SNOW · Full Research · 2026.10.21
AI 拉来的消费，有多少是 Snowflake 自己的
照着 TEAM 和 DDOG 的研究路径走一遍 Snowflake：公司基础、thesis、逐条证据、产品、反向思考、框架打分、估值。原文在前，分析在后。
76份材料
9场管理层发言
965条逐字原文
$210–240买点 (现价 $320)
总体判断
这份是什么：不是 primer，也不是 initial review。它照着我们做 TEAM 和 DDOG 的整条路走一遍 Snowflake：先把公司讲清楚，再写出市场在定价的 thesis 和我自己的假设，用 9 场管理层发言和 Knowledge Base 里的专家、渠道、卖方材料逐条验证，再去理解产品实际怎么用、怎么收钱，最后反过来想怎么会错，落到估值和买点。每条引文都用脚本对原文逐字核对过，补成完整段落，带文件名和行号。
Executive Summary
一句话：Snowflake 的加速是真的，而且比 DDOG 更猛 (产品收入增速一年里从 26% 到 37%，全年指引从 +27% 一路上调到 +36%)；但拆开看，加速里大约一半是 Snowflake 自己的高毛利核心 credit，另一半里混着低毛利的模型 token 转售和"把未来的消费提前"的一次性成分。同时，管理层亲手把平台做得更开放 (Iceberg、Polaris 让外部引擎读写)，CEO 自己说"迁入变快，迁出也一样变快"。现价 $320 按"不扣 SBC 的每股 FCF 5 年复合约 27%"定价，SBC 约占收入 30%，扣掉后 FCF 是负的。不买，放观察名单，买点 $210–240。
加速是真的，也是广泛的。产品收入增速 26% → 32% → 29% → 30% → 34% → 37%，Q3 指引 +37–38%，按惯常 3% 的 beat 是 40% 以上。NRR 从 124% 回到 126%，Q1 新增客户 616 家 (同比 +38%)，管理层说加速"来自很广的客户，不集中"，AI 原生公司只占收入的一小部分。渠道商看到的方向一致但幅度偏低：北欧 SI 看到 29% → 30% → 31%，一位专家的客户样本 22–23% → 26% → 29–30%，北美渠道只有 25–27%，一家大型药企客户只有 10–15%。
"一半来自 AI"要打折。CFO 说的"一半"有两个版本：业绩会上是"约一半加速来自 AI 产品"，在 Piper 会议上是"一半 beat 来自 AI 产品"。AI 收入按 SKU 划分，口径很宽 (CoCo、CoWork、AI functions、AI Gateway 都算)，其中包括 Snowflake 买 Anthropic/OpenAI 的 token 再转卖给客户的部分。一位北欧 SI 专家说 token 收入约 85–90% 付给了模型厂商。另一半里，管理层自己列了 Notebooks、Streamlit 应用和"迁移提速"。一位专家给的按产品拆分是：核心数仓占消费 60–65%、增速 20–25%；Snowpark 占 20–25%、增速约 40%；Cortex/AI 占 10–15%、增速 70% 以上。按这个拆法，37% 里核心数仓贡献约 15 个点，Snowpark 约 9 个点，AI 约 7 个点以上 (我的估算)。
有一部分是"提前"。迁移周期从约 2 年缩到 1–2 个季度 (一家网络设备厂商的 Teradata 迁移不到 3 个季度)；新客户把合同容量用到 80% 的速度在最新一批里明显变快；FY26 账单增速高于收入，因为客户提前用完容量、加签或提前续约。这些对需求是好信号，但同一笔迁移收入被压缩到更短的时间里确认，加速的一部分会在 2027–28 年变成高基数。领先指标已经在放缓：部署的 use case 同比增速从 114% 降到 89%，每个销售赢得的 use case 增速从 86% 降到 43%，RPO 增速从 38% 降到 30%。
AI 层的收入质量低，而且在通缩。产品毛利率指引从 75% 降到 74%，管理层明说原因是"低贡献利润率的 AI 负载占比上升"；Q1 能守住 75% 是靠 AWS 新合同省下的带宽成本。Funda 的渠道估 CoCo 毛利率约 52–55%，另一位专家说 AI 负载毛利只有 34%。Token 价格自 5 月以来下降一半，客户在用 1/20 价格的开源模型。CEO 说单纯转售模型容量"是假装有业务"，但也说如果 CoCo 普及压低毛利，他愿意接受。
锁定在变弱，而且是 Snowflake 主动让的。Polaris 目录现在支持 9 个外部引擎读取，并开始支持写入；Snowflake 在上海开发者活动上明确说客户可以借开放架构"彻底离开"。Databricks 增速约 80%，是 Snowflake 的两倍多；一位前 Databricks 销售说去年约 200 家客户从 Snowflake 迁到 Databricks，其中包括 Snowflake 前 6 大客户的部分负载；一位 Thoughtworks 的负责人说新平台预算约 60:40 偏向 Databricks。但整体迁走仍然很少：PB 级客户迁移要花数百万美元，正面交锋 Snowflake 仍赢约 60%。
估值已经走在前面。市值约 $1,215 亿，EV/FY27 收入约 19 倍。过去 12 个月 FCF $11.95 亿 (收入的 22%)，SBC $16.42 亿 (30%)，扣掉 SBC 是 -$4.5 亿。把 SBC 当成本 (你的默认口径)，基本情形 (5 年收入复合 22%) 只值约 $134；按每股 FCF、把 SBC 当稀释看，基本情形约 $262，乐观情形约 $435。现价比后一种口径的基本情形还高 22%。
逐条结论
| Thesis 环节 | 结论 | 证据强度 | 一句话 |
|---|---|---|---|
| E01 加速来自哪里 | 真实且广泛，但混着前置 | 强 | 公司数字连续 4 季加速、客户分散；核心数仓只增 20–25%，迁移提速和容量提前用完是一次性成分；领先指标在放缓 |
| E02 AI 让核心 credit 变多 | 方向成立，没有量化 | 中 | 个案里 AI 让账户消费 +40% 到 5–8 倍；但 CoCo 的前十大技能之一是帮客户省钱，管理层拒绝披露 AI 采用者的提升倍数 |
| E03 AI 收入质量 | 偏低 | 强 | 毛利率指引 75% → 74%；token 转售 85–90% 付给模型厂商；token 半年降价一半 |
| E04 单价和竞争 | 锁定在变弱 | 中 | CEO："迁出也一样变快"；Polaris 开放写入；Databricks 拿走更多新预算；但整体迁走仍然很少 |
| E05 新产品变现 | CoCo 有量，其余还早 | 中 | CoCo 9,100+ 账户、约占收入 2–2.5%；CoWork 增速只有 +11% 环比；Observe 进展慢；CoCo 本身不收费 |
| E06 预期 | 很高 | 强 | Q2 只超买方预期约 1%；Q3 指引已隐含 40%+；空头回补基本完成 |
| E07 股东经济学 | 差，但在改善 | 强 | SBC 约收入 30%，FCF − SBC 为负；GAAP 盈利目标 FY28 Q4；say-on-pay 被否；稀释股数指引 376M → 382M |
对照你的两条标准
- 好公司：是，而且这一年变得更好了。用量计费、NRR 126%、828 家年付 $100 万以上的客户、管理层执行快 (Ramaswamy 上任两年，CoCo、CoWork、Openflow、Postgres、Adaptive Compute 都在他任内推出)，AI 不威胁它按人头收费的部分 (它几乎没有人头收入)。它比 DDOG 弱的地方：毛利率低 (产品毛利 74–75%，GAAP 约 67%)，SBC 高得多 (30% vs 21%)，AI 层有一大块是转售别人的 token。
- 两年一倍：不满足。翻倍是 $639，算上稀释对应市值约 $2,580 亿；即使 FY28、FY29 每年增长 30%、27%，也需要约 24 倍 FY29 EV/Sales。现在是 19 倍 FY27、约 15 倍 FY28。
- 按你的规则：放观察名单等价格。买点 $210–240 (按每股 FCF 口径的基本情形，对应 12–15% 年化回报)。这比 mapping pitch 里给的 $185–215 高一些，因为这次的基本情形用了更高的增速 (FY28 +25%，而不是 pitch 里的减速更快的路径)。
Follow-up
| 时间 | 盯什么 | 什么结果说明什么 |
|---|---|---|
| 2026-12 初 Q3 FY27 业绩 | 产品收入 vs 指引 $1.588–1.593B；产品毛利率；use case 和 RPO | 按惯常 3% beat 约 $1.64B (+41%)。≥$1.64B 且毛利率 ≥74% 说明加速在延续、AI 没有进一步稀释毛利；低于 $1.62B 说明"提前"的成分开始退潮。RPO 增速如果继续低于收入增速，说明合同承诺没跟上消费，Q4 续约要看清楚 |
| 同上 | AI 收入口径和毛利率拆分 | 管理层如果第一次给出 AI 收入或 AI 毛利率，就能直接回答"一半来自 AI"里有多少是 token 转售。继续不给，就按低毛利处理 |
| 2027-02 末 Q4 FY27 业绩 | FY28 首次指引；SBC 占比；续约季 | 四个财年的首次指引是 +40/+22/+24/+27%，实际都高 5–10 个点。FY28 首次指引 ≥28% 说明管理层自己看到的 run-rate 足够硬；≤25% 说明 FY27 有前置。多数合同在 Q4 续约，看 RPO 和 cRPO |
| 持续 | Databricks 数字 (可能的 IPO 披露) | 如果 Databricks 数仓业务 (DBSQL) 的增速继续是 Snowflake 的 2 倍以上，"新预算 60:40 偏 Databricks"就在变成存量份额的流失 |
| 可以约的专家 | 年付 Snowflake $500 万以上客户的数据平台负责人 | 最缺的一手数据：2026 年 credit 消耗涨了多少、按负载拆开 (核心 ETL/BI、Snowpark、AI) 各涨多少、续约时每 credit 价格降了多少、有没有把负载放到 Iceberg + 别的引擎上 |
证据限制：管理层材料占证据的大头，关键数字 (AI 收入、AI 毛利率、AI 采用者的消费提升) 管理层都拒绝给；专家多是 SI 和渠道合作伙伴，看到的是自己的客户，样本偏中大型企业；Databricks 一方的说法 (约 200 家迁移客户) 有立场；Funda 和 TMTB 是二手转述；Merit 音频转写有错字，MeritForum 整理稿和音频偶有不一致 (以音频为准)。Q3 FY26 和 Q4 FY26 两场业绩会只取到 IR 转写 PDF 里按主题抽出的单句，不是全文。
I assume just given the size of the core, it's the core growing faster is probably the bigger factor. But I was wondering if there's any way for us to sort of distill down what these newer products are having maybe on their own account. Thanks. I would roughly call it even, our AI products which is a pretty broad swath at this point. Absolutely, it's CoCo and CoWork, but it's also things like AI functions that make data operations proceed at an impressive scale, or even newer products like the AI Gateway. They contributed approximately half of the acceleration that we are seeing. But there are a lot of other products that are also demonstrating robust growth. And Brian touched on some of them. Whether it's Notebooks or applications written in Streamlit or React that are deployed into Snowflake and, of course, migrations themselves going faster.
This includes approximately 1 percentage point of growth from Observe consistent with our previous outlook. In Q3, we expect product revenue between $1.588 billion and $1.593 billion, representing 37% to 38% year-over-year growth. Turning to margins. For FY 2027, we now expect 74% non-GAAP product gross margin. This revised outlook includes a higher revenue mix from fast growing AI workloads, which carry a lower contribution margin today. We're delivering continued operating margin expansion as we offset growing cloud costs with slowing head count expense. We are increasing our FY 2027 non-GAAP operating margin guidance from 13.5% to 14.5%. For Q3, we expect non-GAAP operating margin of 15.5%.
Want to ask the switching cost question both ways, which is, we've talked already about speed of migrations accelerating to Snowflake. We've also talked in the last year about things like standardization of data tables and data graph is becoming flat, grave (00:30:00), for lack of a better word. How do you think about the longer-term implications from switching costs potentially going down in the data infrastructure world? I mean, just bluntly, I told my teams, I'll admit this to you, if migration into Snowflake can be made a whole lot faster, migration out of Snowflake can be made a whole lot faster. That's the world we live in. And it applies to data platforms. It applies potentially to consumer software. It applies everywhere.
在增长方面，传统 Snowflake 的增长大概在 20% 到 25% 左右。Snowpark 和机器学习的同比增长约为 40%。Cortex 和 agents 的同比增长超过 70%，而其他领域我估计大约为 30%。核心的数据仓库仍然是一块大饼，但增量增长已经不再主要来自这里，因为 Snowpark 正在把 Snowflake 从 SQL/BI 扩展到以 Python 驱动的数据工程和机器学习工作负载。Cortex 在结构化数据处理上增加了推理和 agentic 工作负载，Cortex AI 本身的定价现在主要基于消费，包括基于 token 的计费。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。
公司基础
它是做什么的
一句话：Snowflake 是企业放"全部业务数据、拿来分析"的云上大仓库。一家公司的订单、库存、客户、财务、App 日志散在几十个系统里 (SAP、Salesforce、自家数据库)，要回答"上季度哪个区域哪个产品卖得最好、为什么"，就得先把数据集中到一个地方，清洗好，再让分析师写 SQL 或用 BI 看板去查。Snowflake 就是这个地方。
它的技术核心是"存储和计算分开"。数据放在 AWS、Azure 或 GCP 的对象存储里 (很便宜)；要查的时候，Snowflake 临时起一组计算机器 (叫"虚拟仓库")，查完就关。不同团队可以各起各的仓库、互不抢资源。2014 年这是革命性的设计 (老一代的 Teradata、Oracle 必须买一整台机器)，也是它 2020 年成为史上最大软件 IPO 的原因。
它在一家公司里干的几件事：
- 数据接入和加工 (ETL/ELT)：每天夜里把各系统的数据导进来、清洗、合并成干净的表。这部分是最大、最稳定的计算消耗，跑在固定时间，量随数据量增长。
- BI 和报表：Tableau、Power BI 这类看板背后的查询都打到 Snowflake 上。
- 数据工程和机器学习 (Snowpark)：用 Python/Spark 风格的代码直接在 Snowflake 里处理数据、训练模型，以前这部分要搬到 Databricks 或 Spark 集群上做。
- AI (Cortex)：在 Snowflake 里直接调用大模型做总结、分类、抽取 (AI functions)；CoCo 是帮数据工程师写 SQL/管道的编程 agent；CoWork (原 Snowflake Intelligence) 是给业务人员用自然语言问数据的 agent。
- 数据共享和 Marketplace：一家公司可以把表"共享"给供应商或客户，不用复制文件；第三方数据商也在上面卖数据。
谁在用：约 14,500 家客户，828 家过去 12 个月付了 $100 万以上，福布斯 2000 强里约 830 家。前 25 大客户平均每年花 $3,400 万 (两年前 $2,200 万)，而 G2000 客户平均只有约 $240 万，所以管理层说渗透还很浅。大部分收入来自"六到八年前就上来的老客户"。
怎么赚钱：约 95% 是用量制。客户买 credits (计算额度，标价按版本和云区域约 $2–4/credit)，查询按秒扣；存储按 TB/月 另收 (约占收入 11%)。大客户签 1–3 年的容量合同预付，用完了可以加签或提前续约，没用完的到期确认。所以 Snowflake 的收入等于"客户实际跑了多少计算 × 每 credit 价格"，和人头无关。这让它不受"agent 替代人"的冲击，但也让它直接暴露在"客户优化"和"计算单价下降"之下 (2023 年客户集体优化，增速一年从 70% 掉到 38%)。底层机器是向云厂商租的，所以毛利率比纯软件低。
怎么收钱：计费单位表
从 snowflake.com 定价页和官方文档的 credit 消耗表整理 (2026-10)。最后一列是我对"AI 用量增长时这个计费单位会不会跟着涨"的判断。大客户实际成交价远低于标价。
| 产品 | 计费单位 | 标价 | 随 AI 增长? | 为什么 |
|---|---|---|---|---|
| 虚拟仓库计算 (Platform Credit) | credit/小时，按秒计费 (60 秒起)；XS=1、S=2、M=4… 每档翻倍 | AWS 美东按需 $2.00 (Standard) / $3.00 (Enterprise) / $4.00 (Business Critical) / $6.00 (VPS) 每 credit；容量合同按 ACV 分档折扣。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | agent、CoCo 生成和跑的 SQL 仍在客户仓库上计 Platform Credit，这是 Snowflake 自有高毛利层 |
| Gen2 标准仓库 | credit/小时 | AWS 上 XS 1.35、M 5.4、XL 21.6 credit/小时 (Gen1 为 1/4/16)，即单价贵 35%，换更快速度。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 部分 | 单价高 35% 但速度更快，跑完同样任务总 credit 可能下降，就是管理层说的性能改进收入逆风 |
| Adaptive Compute / Adaptive Warehouse | Platform Credit，只在查询运行时计费，大小自动 | 没有固定的每小时费率，按用量和设置浮动。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 部分 | 自动调大小、消除闲置，对客户省钱、对 Snowflake 是单价通缩；Investor Day 称相对 Gen1 分析快 1.6x |
| 存储 | $/TB/月 (压缩后平均) | 标准 $23/TB/月；Hybrid Tables $0.34/GB/月；Postgres 存储 $117.76/TB/月 (HA $235.52)；归档 Cool $4、Cold $1/TB/月。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 部分 | 非结构化/AI 数据增加存储，但存储基本按云成本转嫁、低毛利，且 Iceberg 外部存储不收费 |
| AI Credit (所有 Cortex AI 功能的计价单位) | AI Credit | 按需 Global $2.00 / Regional $2.20；容量最高档 (ACV $4,000 万+) $1.88/$2.07；Platform Credit 折扣不适用于 AI Credit。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | AI 单独一种货币、折扣更少，便于单独核算 AI 收入，但底层是模型 token 成本 |
| Cortex AI_COMPLETE (LLM 函数) | AI Credit / 百万 token (输入/输出) | claude-sonnet-4-6 1.80/9.00 (约 $3.60/$18.00)；claude-opus-4-6 3.00/15.00；claude-haiku-4-5 0.60/3.00；gpt-5.4 1.50/9.00；gpt-5.1 0.75/6.00；llama3.3-70b 0.432/0.432。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | 转售第三方模型，按 Anthropic 标价 Sonnet 约 $3/$15 推算 Snowflake 加价约 20%，毛利薄且随模型降价通缩 |
| Cortex Code / CoCo | AI Credit / 百万 token；CLI/Desktop 另有固定月度订阅额度 | claude-sonnet-4-6 输入 1.65 / 输出 8.25 / cache 写 2.07 / cache 读 0.17 AI Credit (约 $3.30/$16.50)。Snowsight 内 2026-04-01 起计费；CLI 独立订阅 (2026-02-24 推出，不需要 Snowflake 计算) 价格未公开。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) ; https://docs.snowflake.com/en/user-guide/cortex-code/cortex-code | 是 | token 加价只有约 10%，几乎是成本价引流；真正的钱在它生成并执行的 SQL/管道所消耗的仓库 credit |
| Cortex Agents / Snowflake CoWork (原 Snowflake Intelligence) | AI Credit / 百万 token + 工具调用 | claude-sonnet-4-6 输入 1.95 / 输出 9.76 / cache 写 2.44 / cache 读 0.20 AI Credit (约 $3.90/$19.52)；Snowflake 托管工具 0，第三方 agent 调用非 Snowflake 端点 8 AI Credit/千次；没有按用户收费，但有每用户配额。另加查询和仓库费用。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) ; https://docs.snowflake.com/en/user-guide/snowflake-cortex/governance-and-availability/ai-cost-management-and-governance | 是 | 面向业务用户，token 加价约 30%，高于 CoCo；每次回答还触发 Cortex Analyst 生成 SQL 在仓库执行 |
| Cortex Analyst (text-to-SQL API) | 每千条消息 | 表中为 '67 Platform Credits per 1,000 messages' (其中 '6' 可能是脚注标号，即 7 credit/千条，需核实)；仅直接调 API 时计费。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | 把自然语言转成 SQL，生成的 SQL 再消耗仓库 credit，是 AI 带动核心消费的直接机制 |
| Cortex Search | AI Credit / GB/月 (索引数据量) | 6.3 AI Credit/GB/月；Batch Search 0.12 AI Credit/GB/小时。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | 非结构化数据检索 (RAG)，按索引数据量收费，随文档入库增长 |
| AI_PARSE_DOCUMENT / AI_EXTRACT / AI_CLASSIFY | AI Credit / 千页或百万 token | Layout 3.66、OCR 0.68、Agentic 8.18 AI Credit/千页；AI_EXTRACT (arctic-extract) 5.55、AI_CLASSIFY 1.62 AI Credit/百万 token。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 是 | 自有/开源模型处理文档，毛利可能比转售 Claude/GPT 高 |
| Openflow (数据接入，NiFi) | Platform Credit / vCPU-小时 (BYOC) 或节点小时 (SPCS) | BYOC 0.0225 credit/vCPU/小时；Snowflake 部署 CPU_X64_S 0.11 credit/小时；Oracle 连接器 $70/月许可 + $40/月维护 (36 个月)。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 部分 | 接入本身单价很低，价值在把更多数据带进 Snowflake 后的下游计算 |
| Snowflake Postgres | Platform Credit / 实例小时 + 存储 | STANDARD_M 0.0356 (AWS) / 0.0376 (Azure) credit/小时，HIGHMEM_L 0.1024；高可用 (HA) 翻倍。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 部分 | 把 OLTP 和 agent 应用的状态库拉进 Snowflake，是新的消费池，但单实例价低 |
| Serverless Tasks / Open Catalog / 数据传输 | 倍数 / 每百万请求 / $ per TB | Serverless Tasks 计算倍数 0.9 (Flex 0.5)；Open Catalog (Polaris) 0.5 credit/百万请求，限时免费；跨云/公网传输 $90/TB，同区 $0。https://www.snowflake.com/legal-files/CreditConsumptionTable.pdf (Effective: October 15, 2026) | 否 | Iceberg 目录基本免费，开放格式下 Snowflake 只能靠计算收费 |
读这张表的要点：Snowflake 的收入有两种性质完全不同的 credit。一种是"自己的计算"：虚拟仓库、Snowpark、Openflow 这些跑在 Snowflake 自己的引擎上，毛利率高 (产品毛利 75% 以上)，护城河是数据已经在这里、权限和治理已经配好。另一种是"转手的模型"：Cortex AI functions、CoWork、CoCo 背后调用的是 Anthropic、OpenAI 或开源模型，Snowflake 按 token 或 AI credits 收钱，再把大部分付给模型厂商。CoCo 本身"不收费"，客户付的是它驱动的消费，一部分是模型 token，一部分是 CoCo 写出来的 SQL 在仓库里跑的 credit。第二种 credit 增长快，但毛利低、单价在跟着模型降价。
AI credit 的加价有多少：2026-10-15 版的 credit 消耗表里，所有 Cortex AI 功能用单独的 AI Credit 计价 ($2.00/个，Platform Credit 的容量折扣不适用)。CoCo 调用 Claude Sonnet 4.6 每百万 token 输入/输出收 1.65/8.25 AI Credit，约 $3.30/$16.50；CoWork/Cortex Agents 约 $3.90/$19.52。Anthropic 对外标价是 $3/$15 (这个数是我按已知信息填的，这次没有重新抓取核实)，所以 Snowflake 在 token 上的加价约 10% (CoCo) 到 30% (CoWork)，再扣掉自己的推理和带宽成本，token 这一层接近成本价转售。Snowflake 在 AI 上真正赚钱的地方，是 CoCo 和 agent 生成的 SQL 在仓库里跑掉的 Platform Credit。
历史、股价和季度数字
历史：2012 年由两位前 Oracle 数据库架构师 Benoit Dageville、Thierry Cruanes 和 Marcin Żukowski 创办，2014 年产品上线。2019 年 Frank Slootman (前 ServiceNow CEO) 接任 CEO，2020-09 以 $120 上市，首日收盘 $253.93，是史上最大软件 IPO，Berkshire 和 Salesforce 各认购 $2.5 亿。2021-11 收盘创新高 $401.89，当时产品收入增长 106%。
$100 亿目标和它的落空：2022 年投资者日提出 FY29 (2028-02 至 2029-01) 产品收入 $100 亿，FY23 实际 $19.4 亿，隐含年复合约 32%。2023 年加息后大客户集体压缩云账单，用量制让节省立刻变成收入下降：FY24 首次指引 +40%，5 月下调到 +34%，股价单日 -16.5%。2024-02-28 是转折点：FY25 首次指引只有 +22% (共识约 +27%)，同时撤回 $100 亿目标、宣布 Slootman 退休，由 Sridhar Ramaswamy (前 Google 广告负责人，Snowflake 2023 年收购了他的搜索公司 Neeva) 接任，股价单日 -18%。管理层当时给了三个理由：性能改进让客户少用 credit (约每年 6% 的收入逆风)、大客户存储降价、Iceberg 让部分存储流走。2024 年 4–6 月约 165 家客户账户因没开 MFA 被盗 (含 AT&T、Ticketmaster)，9 月股价收在上市后最低 $108.56。
新 CEO 的两年：2024-11 宣布与 Anthropic 合作、收购 Datavolo (后来的 Openflow)，股价 +33%。FY25 实际 +30% (首次指引低了 8 个点)，FY26 首次指引 +24%、实际 +29%。2025-09 CFO 从 Mike Scarpelli 换成前 GitLab CFO Brian Robins；2026-03 CRO 从 Mike Gannon 换成内部老将 Jonathan Beaulier；2026-06 联合创始人 Benoit Dageville 退出日常运营。
这一年发生了什么：2025-11 到 2026-04，股价从约 $263 跌到 $121 (-54%)，原因是整个软件板块被"AI agent 取代 SaaS"杀估值 (4 月 Anthropic 发布 Managed Agents 那两天 SNOW 跌了 19%)。其实 2 月的 Q4 FY26 已经是 +30%、RPO 加速到 +42%，CoCo 2 月 5 日刚 GA。之后：5 月底 Q1 FY27 +34%，全年指引 +27% → +31%，同时宣布 5 年 $60 亿的 AWS 采购承诺，股价单日 +36.5% (上市以来最大)；9 月初 Q2 +37%，全年指引一次上调到 +36%，股价 +16.6%。10 月初周收盘到 $369，现在 $320，比 4 月低点涨了约 2.6 倍。
周收盘价，来源 IBKR，2021-10-25 至 2026-10-20 (最后一个点为 10 月 20 日收盘 $319.61)。52 周区间 $118.30–$384.26 是日内价；历史收盘最高是 2021-11-16 的 $401.89 (日线)。
季度数字
来自公司新闻稿 (SEC 8-K 附件和 BusinessWire)。财年截至 1 月，FY27 Q2 = 2026 年 5–7 月。SBC 是现金流量表口径。
| 季度 | 产品收入 | 同比 | 产品毛利率 | FCF (率) | SBC | NRR | RPO | $1M+ 客户 | 当季给的全年指引 |
|---|---|---|---|---|---|---|---|---|---|
| FY25 Q1 | $789.6M | +34% | 77% | $331.5M (40%) | $331.9M | 128% | $5.0B, +46% | 485 | $3 (+24%) |
| FY25 Q2 | $829.3M | +30% | 76% | $58.8M (7%) | $356.0M | 127% | $5.2B, +48% | 510 | $3 (+26%) |
| FY25 Q3 | $900.3M | +29% | 76% | $78.2M (8%) | $363.3M | 127% | $5.7B, +55% | 542 | $3 (+29%) |
| FY25 Q4 | $943.3M | +28% | 76% | $415.4M (42%) | $428.1M | 126% | $6.9B, +33% | 580 | $4 (+24%) |
| FY26 Q1 | $996.8M | +26% | 76% | $183.4M (18%) | $379.5M | 124% | $6.7B, +34% | 606 | $4 (+25%) |
| FY26 Q2 | $1,090.5M | +32% | 76% | $58.2M (5%) | $404.2M | 125% | $6.9B, +33% | 654 | $4 (+27%) |
| FY26 Q3 | $1,158.4M | +29% | 75.9% | $113.6M (9%) | $412.3M | 125% | $7.88B, +37% | 688 | $4 (+28%) |
| FY26 Q4 | $1,226.6M | +30% | 75% | $765.1M (60%) | $403.6M | 125% | $9.77B, +42% | 733 | $5 (+27%) |
| FY27 Q1 | $1,334.3M | +34% | 75.1% | $232.8M (17%) | $402.5M | 126% | $9.21B, +38% | 779 | $5 (+31%) |
| FY27 Q2 | $1,491.9M | +37% | 74.7% | $83.8M (5.4%) | $423.6M | 126% | $9.00B, +30% | 828 | $6 (+36%) |
这张表里最该注意的四件事：(1) 产品收入增速从 26% 一路到 37%，NRR 从 124% 回到 126%，$1M+ 客户从 485 家到 828 家，加速是真的；(2) 产品毛利率从 77% 降到 74.7%，FY27 指引 74%，收入越加速毛利越低，和 DDOG 一样，而且原因更明确 (AI 负载)；(3) SBC 每季约 $4 亿，没有随收入加速而下降，FCF 季度波动很大 (Q4 集中收款)，全年 FCF 率约 23–25%；(4) RPO 增速从 Q4 FY26 的 42% 降到 Q2 FY27 的 30%，在收入加速的同时，合同承诺端在放缓。管理层的解释是客户提前用完容量后转为按需付费、续约集中在 Q4。
竞争格局
竞争格局，按威胁从低到高：
- Teradata、Oracle、Cloudera 等老一代数仓：Snowflake 的主要迁移来源。CoCo 和 Datometry (把 Teradata 层虚拟化的工具) 让迁移从 2 年缩到不到 3 个季度。这一块是 Snowflake 的顺风，但也是会用完的存量。
- Google BigQuery、AWS Redshift、Microsoft Fabric：云厂商自带的数仓，常和云合同打包、有补贴。GuidePoint 的专家说，三家云厂商都在用数百万美元的补贴把客户从 Snowflake/Databricks 迁出，补贴从 $10–20 万涨到"包揽整个迁移"。Snowflake 的防线是跨云中立：大企业不愿把全部数据锁进一家云。
- ClickHouse、DuckDB 等开源/轻量引擎：在实时分析和小规模分析上更便宜。Snowflake 自己承认 500 毫秒以下新鲜度的数据"目前做得不好"。
- Databricks (私有，估值约 $1,900 亿)：最大的对手。从 Spark (大数据处理) 起家，主打"湖仓一体"，2024 年以超过 $10 亿收购 Iceberg 的创始团队 Tabular，开源了 Unity Catalog。年化收入超过 $70 亿、增速约 80%，数仓 (DBSQL) 业务 run-rate 约 $15 亿。它在 AI/ML、流处理和工程师心智上更强；Snowflake 在"开箱即用"、治理和分析团队友好度上更强。专家说正面交锋 Snowflake 赢约 60%，但新平台的预算约 60:40 偏向 Databricks。
- 开放格式本身 (Iceberg + 开放目录)：长期最大的结构性威胁。数据用 Iceberg 存在客户自己的 S3 上以后，同一份数据可以被 Snowflake、Databricks、Trino、DuckDB 读写，客户可以按"每个查询哪个引擎便宜"来分配计算。Snowflake 的应对是把锁定上移到"平台层"：Horizon 治理、Polaris 目录、Snowflake 托管的 Iceberg 存储、CoCo/CoWork 这些 agent。我们 mapping 里对数仓的判断 ("锁定从格式移到平台层，风险是单位计算价格") 在这次的证据里得到了印证。
- 外部 AI agent (Claude、ChatGPT、Cursor) 通过 MCP 直接读数据：如果企业的 AI 层落在外部 agent 上，Snowflake 就退化成"被 agent 查询的存储和计算"。Snowflake 自己的工程师承认这是目前大多数企业的用法；CEO 的偏好排序是"先卖 CoCo/CoWork，其次卖模型容量，最后才是只当 Claude 的数据后端"。
Thesis：市场在定价什么 vs 我的假设
市场现在在定价什么
把卖方报告、TMTB 和 Funda 的买方情绪、管理层的说法放在一起，市场隐含的 thesis 是：
"Snowflake 是企业 AI 的数据控制面。企业的数据已经在这里、治理已经配好，agent 要用数据就得在 Snowflake 上跑，所以 AI 会持续把 credit 拉上去。CoCo 是公司史上最成功的产品，它让迁移更快、让更多人用 Snowflake。加速是结构性的，35–40% 可以维持，毛利率的下降是暂时的，SBC 在快速下降，FY28 实现 GAAP 盈利。"
这个 thesis 的证据看起来很强：Benchmark 在 Investor Day 后把目标价提到 $290，理由是 Snowflake "正在成为企业 agentic AI 的控制面"；Funda 在 Q2 后说 SNOW 是 AI 基础设施里"叙事最清晰"的，比 DDOG (大客户逆风) 和 MDB (AI 收入基数小) 都好；Q2 后 TMTB 把 SNOW 称为"软件里最拥挤的名字"。
我的假设：AI 拉来的消费，有多少是 Snowflake 自己的、能持续多久
做 TEAM 的时候，问题是"AI 让 seat 变多还是变少"；做 DDOG 的时候，问题是"AI 让数据量暴涨，Datadog 收到了多少"。Snowflake 的问题更进一步：它的加速里，有多少是 Snowflake 自己的高毛利 credit，有多少是转手的模型 token，有多少是把未来的消费提前到今天？
我把加速拆成三层：
- 第 1 层，AI 拉动的核心 credit：agent 和 AI 应用要更多干净的数据、更多数据加工、更多查询，跑在 Snowflake 自己的引擎上。这是最好的一层：毛利高、粘性强，和"数据控制面"的叙事一致。
- 第 2 层，低毛利的 token 转售：Cortex AI functions、CoWork、CoCo 背后的模型调用。增长最快，但毛利只有核心的一半左右，单价跟着模型厂商降价。
- 第 3 层，一次性成分：CoCo 让迁移从 2 年缩到几个季度、新客户更快用完合同容量、Observe 收购贡献约 1 个点、AWS 合同带来的 GTM 投入。这些让今年的数字好看，但会变成明后年的高基数。
反方向还有一组"价格"力量：Iceberg 和开放目录让计算可以被分流到别的引擎；Databricks 拿走更多新预算；云厂商补贴迁移；Gen2 和 Adaptive Compute 让每个查询用的 credit 更少；CoCo 自己在帮客户优化账单。
所以我的假设是：只有当第 1 层是加速的主体、而且第 1 层的增长快于"价格"力量的侵蚀，市场的 thesis 才成立。如果加速主要来自第 2 层和第 3 层，Snowflake 仍是一家很好的公司，但 FY28 起会回到 25% 左右，毛利率继续向 72% 漂移，不值 19 倍收入。
把 thesis 拆成可以验证的 7 条预期
| 编号 | 预期 | 要找什么证据 |
|---|---|---|
| E01 | 加速来自哪里 | 产品收入增速、NRR、新客户；按产品拆分 (核心/Snowpark/AI)；迁移提速、容量提前用完、Observe 的贡献；渠道数字是否对得上；领先指标 |
| E02 | AI 让核心 credit 变多 | AI 采用者的核心消费提升；具体客户案例 (消费从多少到多少)；反例 (AI 帮客户省钱、外部 agent 绕开 Snowflake) |
| E03 | AI 收入的质量 | AI 产品毛利率、token 转售占比、模型降价、开源模型替代；管理层对毛利的取舍 |
| E04 | 单价和竞争 | Iceberg/Polaris 的影响、Databricks 份额、云厂商补贴、迁出案例、Gen2/Adaptive Compute 的定价 |
| E05 | 新产品变现 | CoCo、CoWork、Openflow、Postgres、Observe 的采用和收入；收费方式 |
| E06 | 预期和仓位 | 指引方法、beat 幅度、卖方/买方预期、股价反应 |
| E07 | 股东经济学 | SBC、稀释、回购、FCF、经营利润率、GAAP 盈利路径、治理 |
反常识 Check List
按对 thesis 的冲击大小排序。每条先说人话，再放原文。
1. CEO：迁入变快，迁出也一样变快 → 冲击 E04
CoCo 最被市场看好的一点是让迁移变快 (Teradata 迁移从 2–3 年到不到 3 个季度)。9 月在高盛会议上，CEO 自己把另一面说了出来：同一套工具也能让客户更快迁出。第二天 CFO 在花旗会议上重复了这句话。这对 Snowflake 的含义是：转换成本这条护城河在被 AI 工具双向削平，存量消费的粘性比过去低。
Want to ask the switching cost question both ways, which is, we've talked already about speed of migrations accelerating to Snowflake. We've also talked in the last year about things like standardization of data tables and data graph is becoming flat, grave (00:30:00), for lack of a better word. How do you think about the longer-term implications from switching costs potentially going down in the data infrastructure world? I mean, just bluntly, I told my teams, I'll admit this to you, if migration into Snowflake can be made a whole lot faster, migration out of Snowflake can be made a whole lot faster. That's the world we live in. And it applies to data platforms. It applies potentially to consumer software. It applies everywhere.
From a data coming into Snowflake and then other companies trying to be a database or be a database site, the thing about Snowflake that is incredible is we are an open system. We have Iceberg tables where we're encouraging you to input your data from any anywhere anyhow or whatever, it's our -- we have to create great products like CoCo to incentivize you to use us. One of the things around that we talk a lot about internally is it's customers first. If you don't make a great experience for your customer, we talk about it's really easier to migrate into Snowflake with all these tools, it's just as easy to migrate out of Snowflake. And so, let's say the reverse of it, too, right? And so, we have to create those great customer experiences to keep them within Snowflake and to give them the ROI of doing that. And since we don't run -- we are in HR application, we are in CRM, we don't run all these various components like a lot of people want to put all their data in Snowflake.
2. Snowflake 自己在宣传"客户可以彻底离开" → 冲击 E04
在 9 月的上海开发者活动上，Snowflake 的工程师介绍 Iceberg + Polaris 开放架构时说，客户可以借此迁移甚至彻底离开 Snowflake，没有厂商锁定；Polaris 已经支持 9 个外部引擎读取，现在又开始支持外部引擎写入。只读意味着 BI 查询可以走；可写意味着数据加工 (ETL) 这种最大、最稳定的计算消耗也可以走。Snowflake 这样做是为了赢新客户 (不开放就进不了门)，但代价是存量计算的锁定变弱。
Snowflake 的开放架构建立在 Apache Polaris、Iceberg 和 REST Catalog 之上。Apache Polaris 由 Snowflake 于2020年捐献给 Apache 基金会，并于去年成为 Apache 基金会的顶级项目。它能够真正实现双向互操作，不存在 vendor lock-in，并提供开源、开放的能力。任何第3方 Lakehouse Engine 都可以无缝接入。即使未来因技术迭代需要调整，企业客户也可以在必要时选择迁移，甚至彻底离开，从而避免任何 vendor lock-in。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。
2024 年，Polaris 已经捐献给 Apache 社区，目的是为客户提供更多选择。到 2025 年，我们发布了 Polaris Catalog，并将其集成在 Polaris 和 Iceberg 的 RAF 上。随着时间推进，我们支持了 9 个外部引擎读取 Polaris Catalog，现在又支持外部引擎写入 Polaris Catalog。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。
3. 毛利率指引下调，原因是 AI 负载 → 冲击 E03
Q2 业绩会上 CFO 把全年产品毛利率指引从 75% 降到 74%，原因明确是"低贡献利润率的 AI 负载收入占比上升"。Q1 时 75% 能守住，是靠 AWS 新合同省下的带宽成本。这说明第 2 层 (token 转售) 已经大到能把全公司毛利拉低 100bp。CEO 在高盛会议上说，如果 CoCo 大规模普及压低毛利，他愿意接受。
Yeah. Karl, thanks for the question. You're absolutely right. Our AI products have a lower gross margin than our core platform. The one thing that we want to do with our AI products when we launch a new product like CoCo is make sure that we develop a great product, that we get massive adoption, and we've seen really good adoption with CoCo. We're up to roughly about – a little over 7,000 accounts have adopted CoCo. With that said, we're offsetting that and keeping the same product gross margin at 75% for the full year, and lower bandwidth costs, i.e., I talked about the AWS contract, and so we're offsetting it there. So that's how we're able to do that. We're committed to find efficiency to be able to maintain that 75% gross margin. Okay.
I just want the data platform, and it can be a back end to Claude, we will do that as well, somewhat more reluctantly. It's important that you have your Maslow's hierarchy of where you are creating value and acting according to that. And Brian and I are super aligned on what are the business outcomes that we want to drive. Let's say, if CoCo adoption were to go up massively and it has an impact on our gross margins, I'm happy to come and explain that to you all day along. That's not an issue because it will drive a meaningful acceleration in our overall business. And I can also tell you, as we grow and scale, as we do these things, we get better at optimizing. We get better at running open source models, which will have much better margins. We obviously buy a bigger quantity from the suppliers like we do with AWS and the $6 billion contract, which gives us better economics.
4. CoCo 的前十大技能之一是帮客户省钱 → 冲击 E02
CoCo 被当作"AI 拉动消费"的核心证据，但它同时是一个很好的成本优化工具：让它找出最慢的查询、最闲置的仓库，写出更高效的 SQL。管理层在业绩会、高盛、花旗、Piper 四个场合都说"成本优化是 CoCo 前十大技能之一"，理由是不想让客户的账单变成"定时炸弹"。一位 Snowflake 合作伙伴的专家说，同样的提示下 CoCo 写的 SQL 消耗更少 credit；Snowflake 自己的工程师说，借语义层可以让 agent 的 token 用量降 20–50%。
First, I think we see the acceleration come from a very broad swath of customers. It is not concentrated. For example, with, let's say, AI-native companies, they continue to be a small part of our overall revenue stream. And I think the thing that's also materially different this time around with folks that are investing is that the products like CoCo make optimization far, far easier than before. You can point CoCo at a query that's taking too long to run, or you can basically have it debug the top 10 longest running queries are the most idle warehouses, things like that are a lot easier to do. And in fact, our cost management skill in CoCo is a top 10 skill. And it's also the case that as a company, we have learned the lessons of the pandemic. And one thing that we stress with each and every one of our customers is the need to drive spend in an efficient way. And this is also a mantra that our sales team itself adopts pretty aggressively, because they know that every such case, where they go to a customer and point out things that they could be doing better, is a trust building exercise that is going to more than pay for itself in new projects that customers will implement on Snowflake.
相比之下，Cloud Code 虽然底层也能访问基础模型，但它并没有针对为 Snowflake 或 Cortex 编写 SQL 查询进行优化。如果将相同的提示词分别输入给 Cortex Code 和 Cloud Code，Cortex Code 生成的 SQL 查询在效率、速度和准确性上都远超 Cloud Code 生成的查询。这能为客户带来更低的积分消耗和更高的性能。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。
在生产环境以及部分客户的数据中，Token使用量实际上降低了20%到50%，因此不用担心突然收到天价账单。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。
5. "一半来自 AI"，AI 的口径很宽，而且管理层不给数 → 冲击 E01/E02
CFO 在花旗会议上解释，AI 收入按 SKU 划分：凡是 AI 相关的 SKU 都算 AI，其他 (数仓、存储、分析、摄取) 都是核心，不做归因。所以"一半来自 AI"里包括模型 token 的转售。分析师在业绩会上问 AI 采用者和非采用者的消费提升倍数，管理层说"现在还不准备分享"；唯一的数字 (采用者核心消费多 11%) 来自 CEO 的一条 LinkedIn 帖子。
But can you just sort of talk about those two dynamics and how do you kind of see the growth of the core business trending just given these new use cases that are coming on top of it? Yeah, absolutely. When we put our revenue together, there's no -- and I know a number of companies do attribution to saying, AI revenue is this or maybe they're selling a bundle, and they have to -- for accounting purposes, they have to break it out. Ours is down to the SKU level, and so we have a series of products that are AI related. And so, when we talk about AI and the performance, it's based on SKUs. And then the core is -- the core business, data warehouse, and storage, analytics, ingestion, what you know is Snowflake as a whole. And so, one of the things that we're seeing that is extremely encouraging and really goes to the variability of the business is we're seeing that when people adopt our AI products primarily CoCo, because we have the most data around CoCo, they're actually consuming more of the core. And so, it may be more personas. And so, you're reaching more people within a business.
How has that developed across the earliest cohorts and what evidence are you seeing or maybe what is giving you the confidence that all of this reflects higher lifetime consumption rather than projects just being pulled forward? Thank you. Yeah, I'll take a first cut and Brian will add on. At this time, we aren't ready to share the exact uplift numbers, but we do measure cohort behavior and as CoCo adoption gets deeper, more users within an account adopting and more accounts and more customers themselves adopting. The effect is pretty noticeable for all the different cohorts that we have worked with. And what gives us confidence that this is not merely projects being pulled forward is both the breadth and depth of use cases that are coming our way in terms of what people are doing with CoCo and CoWork. It is allowing people to do fairly sophisticated actions that previously would have required things like applications.
And so that's what we're in the process of doing. But it's really a – the 5% comes down to the core business is doing well and is accelerating. We're also seeing benefit in the core because of AI. And so Sridhar did a post on LinkedIn a little while ago, and I think it was something to the effect that when someone adopts CoCo or AI, we're seeing 11% uplift to their core business that they would do otherwise. And so both of those are sort of contributing to the beat. And maybe you can drill down on that a little bit more because CoCo is creating a great unlock just in terms of migrations, moving data over. Yeah.
6. 领先指标在放缓 → 冲击 E01/E06
收入在加速，但领先指标在掉：部署的 use case 同比增速从 Q1 的 114% 降到 Q2 的 89%，每个销售赢得的 use case 增速从 86% 降到 43%；RPO 增速从 38% 降到 30%，低于收入增速 37%。管理层的解释是客户提前用完容量后按需付费、续约集中在 Q4。这个解释说得通，但也正好说明一部分加速是"把消费提前"。
We have deployed CoCo and CoWork across the sales team to analyze pipelines, prepare for customer conversations, and accelerate the onboarding of new reps. Our teams are using these products every day, learning firsthand what they can do and taking those insights directly to our customers. We're seeing the results in how quickly customers are putting Snowflake to work. The number of use cases, individual customer projects deployed on Snowflake increased 89% year-over-year as customers moved more workloads into production. At the same time, use cases won per account executive increased 43% year-over-year, demonstrating both growing customer demand and strong sales productivity. And we're pairing this investment in growth with continued operational discipline. We remain on track for GAAP profitability in Q4 fiscal 2028, and the operating leverage we build along the way strengthens the durability of that outcome. Taken together, our rapid pace of innovation, tighter go-to-market execution and operational discipline positions us well to capture the huge opportunity ahead.
This expansion is underpinned by growth in both migrations and AI use cases. In Q2, 48 net new customers surpassed $1 million in trailing 12-month spend. We now have 828 customers spending above the $1 million threshold. Remaining performance obligations grew 30% year-over-year, totaling $9 billion. As a reminder, we continue to see customers favor Q4 renewals. As a result, we expect bookings to be increasingly weighted towards the fourth quarter. Of the $9 billion RPO, we expect approximately 54% to be recognized as revenue in the next 12 months. This represents approximately 42% year-over-year growth compared to our estimate in the same quarter last year. Our Q2 results reinforce our commitment to delivering both growth and margin expansion. In Q2, non-GAAP operating margin expanded over 400 basis points year-over-year to reach 15%. Our outperformance was driven by strong revenue growth and disciplined head count management.
That's the magic of these coding agents. And in a funny kind of way, CoCo also makes it far easier to create agents and get value from the data itself. And this is the combination that makes Snowflake so attractive and it's not just acquiring customers. We track this metric called like time to 80% of purchased consumption for new logos that we acquire and we measure it cohort by cohort. Basically, of the customers that you acquired, let's say, in January, what fraction of them are consuming more than 80% of their purchased capacity, call it three months after their purchase month. And this metric has very, very visibly improved for the newest cohorts of customers that we are acquiring. That's the power of AI. It's faster to get projects done. It's faster to get value from data.
7. 迁移从 2 年缩到几个季度 → 冲击 E01
迁移变快对需求是好消息，但同一笔迁移收入被压缩到更短的时间里确认。Investor Day 上管理层说迁移数量 FY25 到 FY26 增长 1.9 倍、新客户达到消费的时间从 10 个月缩到 7 个月；Piper 会议上 CFO 说迁移时间从 10–11 个月降到约 6 个月，"受客户因素限制无法降到零"。这是一次性的台阶：下降到 6 个月以后不会再降一半。
Maybe it's the first agent that you're kind of going to launch and you're not stopping there. Maybe there's a number of other ones that are coming, so maybe just help us think about how that changes the potential spend profile of a customer over time. Among the biggest, like, impact that products like CoCo have with our customers is simply one that of expectation. I talked, again, in my remarks about how we did a two-year Teradata migration. We are engaged in more migrations, but the timelines for doing those now run between a quarter and two quarters. Why? Both my team and the customer expects and demands that, and we have the ability to deliver against that. I think the – both the impatience and hunger for what people can do with data, along with the expectation for how quickly we can get them done, I would say that's like – that's a huge sea change.
When we – when I went back and looked at average days sales cycle for this last quarter, it was the lowest in the last four quarters. You would expect with more choices and more evaluations out there that the sales cycles actually expand and actually take longer. They are actually doing the opposite of that so the sales cycles are accelerating. We constantly talk about how we are using our AI tools to actually get new customers to consume faster. We've taken that from ten months down to seven months and we're continuing to see how we can decrease that. And then for all customers, migrations are accelerating. And so, we've shown a 40% improvement. And so, AI across the business is having a dramatic improvement on our time to ramp. All right. Let's talk about a couple customer examples.
And people, depending upon what organization they're in, use it for a number of different things. And so, for instance, in our group that does our migrations, they use it to actually help migrations. And so they'll – we've got a whole bunch of skills and agents that can actually take the migration time down. We're around 10 or 11 months. Now we're around six months. There's some customer dependency there. So, we can't get that down to zero, although we'd like it. But we're currently working as much as possible to bring down the migration timeframe. For myself, I use it for, call it, like, analytics and business intelligence.
8. 渠道和客户层面看到的增速低于公司 → 冲击 E01
北美渠道专家说整体消费增速约 25–27%、过去 3 个季度稳定、没看到加速；一家大型药企客户的 Snowflake 支出只增 10–15%；北欧保险客户的 AI 负载涨了 5–8 倍，但按狭义口径只占其 Snowflake 消耗的 15–20%，其余 80–85% 靠迁移和整合每年增长 20–30%。这几条都是单一视角，但它们指向同一个方向：核心数仓的增速在 20–30%，不是 37%。
回 答 者：整体增速约为25%至27%。过去3个季度的增长较为稳定，均保持两位数增长。今年的增长趋势维持在25%至27%左右，与去年的增速相近。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。
回 答 者：按照 year over year 的口径，同比增长应该为 10%到15%。由于体量较大，有时会高一些，有时会低一些，但每年的支出增幅基本处于这一范围。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。
其余80%至85%主要是数据仓库等工作负载，与直接的AI用例相对独立。这部分也会增长，因为该公司仍在将更多工作负载迁移至Snowflake，这是长期IT供应商整合的一部分，但增幅不会达到5倍至10倍，预计每年增长约30%，即每年20%至30%。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。
9. Databricks 一方说去年约 200 家客户从 Snowflake 迁走 → 冲击 E04
一位前 Databricks 销售说，Databricks 增速约 80%，Snowflake 约 35%；去年近 200 家客户从 Snowflake 迁往 Databricks，Snowflake 前 6 大客户也在迁，主要是 AI/ML 和流处理负载。这是对手视角，数字可能夸大；但 Thoughtworks 的负责人 (中立的 SI) 也说新平台预算约 60:40 偏向 Databricks。
回 答 者：目前，客户迁移主要是从 Azure 或 Microsoft 平台转向 Databricks；第2类常见迁移是从 Snowflake 转向 Databricks，现场工程团队正在大量支持此类项目。这类迁移主要涉及 AI/ML 和流式处理用例。去年，我们看到接近 200 家客户进行此类迁移。事实上，即使是 Snowflake 的前 6 大客户，也在向 Databricks 迁移。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。
BJ: I think it's been pretty split. Honestly, I would say net new is 60-40, Databricks net new versus 55-45, widening slowly, but not where me and you would automatically default to. And it's not because Databricks is winning head-to- head against Snowflake or so. It's that the net new work skews towards engineering heavy agent builds and streaming, which is where Databricks started from the beginning, right? And where the work is more governed analytics and business facing insights and agentic associated with the Snowflake still takes it comfortably. And then I think vendors are calling it out as well, in terms of their growth rate accordingly, right? And then, respectfully, frankly, the architecture is now equal between both of these vendors. So it's just how one got in an account and then how it stayed there and then who the buyer persona is. That's fundamentally where it's going for. But if you think about, again, agent builds, if it's purely engineering, analytics, custom app, Databricks. If it's more BI and business-facing
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。
Evidence：逐条验证
按 7 条预期组织。每节先给判断，再列关键原文，最后可以展开该节全部原文。绿色边框是支持 thesis 的原文，红色是反对，灰色是口径不齐或中性。
E01 加速来自哪里
判断：加速是真的、广泛的，但主体仍是核心数仓 (迁移 + 整合 + 更多人用)，AI 产品是增长最快的一小块；其中有相当一部分是"提前"：迁移提速、容量更快用完。领先指标已经在放缓。
数字：产品收入增速 26% (FY26 Q1) → 32% → 29% → 30% → 34% → 37% (FY27 Q2)，Q2 环比新增 $1.58 亿创纪录。NRR 124% → 126%。Q1 新增客户 616 家 (+38%)、新增 13 家 G2000 (去年同期 4 家)。Top 25 客户平均每年花 $3,400 万 (两年前 $2,200 万)。CEO 说加速"来自很广的客户，不集中"，AI 原生公司只占收入的一小部分。这和 DDOG 不一样：Snowflake 的加速不依赖一两个 AI 实验室。
管理层怎么拆：CEO 在 Q1 把加速拆成三部分：AI 拉动核心数据平台、AI 产品本身、CoCo 带来的二阶核心消费 (含迁移)，没有给比例。Q2 说"约一半加速来自 AI 产品"，另一半是 Notebooks、Streamlit/React 应用和迁移提速。Piper 会议上 CFO 换了说法："一半 beat 来自 AI 产品，一半来自核心"，这是相对共识的超预期拆分，不是增速加速的拆分。
专家怎么拆：一位专家 (MeritForum 8/12) 给的按产品拆分是：核心传统数仓占消费 60–65%、增速 20–25%；Snowpark/ML 占 20–25%、增速约 40%；Cortex/AI agent 占 10–15%、增速 70% 以上。我按这个比例算，37% 的增速里核心约贡献 15 个点、Snowpark 约 9 个点、AI 约 7 个点以上。北欧保险客户的例子更直观：AI 负载 5–8 倍，但只占消耗 15–20%，其余 80–85% 靠迁移和供应商整合每年增长 20–30%。
一次性成分：(1) 迁移提速：从约 2 年到 1–2 个季度，迁移数量 FY26 增长 1.9 倍；(2) 新客户更快用完容量：新客户在购买后约 3 个月消费到合同容量 80% 的比例"明显改善"，FY26 账单增速高于收入，因为客户提前用完、加签或提前续约；(3) Observe 收购全年贡献约 1 个点。
外部印证：方向一致但幅度偏低。北欧 SI：Q4 约 29%、Q1 约 30%、Q2 约 31%；一位专家的客户样本 22–23% → 26% → 29–30%；北美渠道 25–27% 且没看到加速；一家大型药企 10–15%。
对"前置"判断不利的证据：管理层在 Q3 FY26 说传统本地数仓的迁移只完成了 15–20%，跑道还很长；如果这是真的，迁移提速更像是把一条长队排得更快，而不是把有限的存量一次吃完。我仍然认为提速本身是一次性台阶 (6 个月以后很难再减半)，但迁移这块收入的总量未必会很快见顶。值得注意的是 2 月：一位专家当时预计 Q4 FY26 只有 27–29% (低于 Q3 的 32%)，另一位合作伙伴则看到 Q4 约 38%、Q1 可能 40%。加速出现之前，专家意见是分裂的。
So, really sort of a question on the quality of the acceleration you're seeing. This is Sridhar. Let me take a first cut at this, other folks can add on, since it's a pretty broad question. First, I think we see the acceleration come from a very broad swath of customers. It is not concentrated. For example, with, let's say, AI-native companies, they continue to be a small part of our overall revenue stream. And I think the thing that's also materially different this time around with folks that are investing is that the products like CoCo make optimization far, far easier than before. You can point CoCo at a query that's taking too long to run, or you can basically have it debug the top 10 longest running queries are the most idle warehouses, things like that are a lot easier to do. And in fact, our cost management skill in CoCo is a top 10 skill.
Can you talk about where specifically you're seeing the inflection? Thank you very much. Absolutely. No, I would break this up into three parts. First, AI is accelerating the value that people can get from the data that they have put into Snowflake or that they can put into Snowflake. So we saw a – like a healthy secular tailwind for our core data platform. And part two is really that agentic products, the control plane products like Snowflake Intelligence and Cortex Code/CoCo came into their own in Q1. Recall that CoCo went into GA on Feb 5, so just as we were opening up the quarter. And we've seen very strong traction with both the products. And the really interesting thing with Cortex Code is that it, in turn, drives more consumption on the core data platform simply because it's much easier to get projects done, whether it's a pipeline or creating a new agent, or setting up a new dynamic table or even, honestly, a migration. So it's driving the second order effect as well. But it's really – this is the one, two, three, and that's why I like to think of this as AI compounding Snowflake's strengths in data. And I'll hand it off to Brian for the mechanics of how these came together in our forecast for the quarter and the year.
And that doesn't include all the analytics and business intel, but does include all the other stuff around it that you can do with CoCo and CoWork and so forth. And so that's what's super exciting for us at Snowflake. And that's one of the reasons that was what helped our print. We talked about 50% of the beat was related to our AI products and 50% was related to the core. And there are these secular tailwinds sort of in industry now sort of pushing towards getting your data in a database such as Snowflake so you can actually get value out of it. When I talked about – and don't mean to hop on this too much, but when I go talk to CFOs and CEOs, and then show them what we can do, they're like, I want that now. And so, when I joined Snowflake over a year ago, I wasn't having any of those conversations.
A:按产品线划分，核心的传统 Snowflake——仓库计算、SQL、存储、BI 和 ELT——大约占消费份额的 60% 到 65%。Snowpark 或机器学习占 20% 到 25%。然后 Cortex AI、生成式 AI 和 agents 占 10% 到 15%。像流式应用和协作工具等其他部分大约占 5%。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。
So a common problem I see is that you build your pipeline uh based on a POC and then when you scale it, you don't rebuild your pipeline uh uh sort of architecture. And then you have an uh a sort of over proportional consumption increase. Uh, but but but but maybe giving you a flavor, we're still talking about a kind of sort of massive consumption uh compute consumption increase. So, so for for the insurance company, I think they have increased their snowflake compute by, you know, five to eight times, uh, you know, going from pilot to to multi-line production. Uh, and this is driven by uh query volume from claims agents, etc. So it's not so much data volume growth. Um, and and this is something that IT budget, you know, it's it's hard to really budget for this. Um, and and uh I I'm not private to exactly how they did the budgeting, but but it's uh, um, I I don't think uh they they sort of calculated with this, basically. And and and for the electronics manufacturer with MongoDB, it's uh Atlas consumption, it's a similar range, um, uh, basically 5 to 10.
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。
You can see last year it was 86%. Two years before, that was 80% but really remains consistent. There is some noise between billings growth and revenue though. And so, you can see on the far left-hand side, in FY 2024, revenue grew 36% and we had 29% billings growth. But in FY 2026, it is actually reversed. When our customers actually come to the end of their capacity or use the capacity that they purchase, they really have two choices and we allow them to do either the following. They can actually do an addendum and actually extend the current contract to when the contract renewal comes up or they can actually do an early renewal. And so, this causes some noise between billings and revenue. But over a long period of time, this actually normalizes out. So if you look at the three-year, they actually are normalized.
I have talked pretty much in every single earnings call over the past six quarters about migrations, and that is an area where we continue to get faster and faster. And some of the recent advances, both in models and harnesses, are letting us run long-duration tasks of a scale and complexity that we haven't been able to do before. And the rate at which workloads are coming onto Snowflake is also an important factor. And one anecdotal example that a big network equipment manufacturer is doing a Teradata migration in less than three quarters this year, and this is something that would have taken probably two to three years in any previous time. So, these are some of the things that are contributing to our acceleration and beat. Thanks so much, Sridhar. Thank you.
回 答 者：如果没有记错，Q4的增速为29%，Q1为30%。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。
回 答 者：去年Q4约为22%至23%，今年Q1约为26%，今年Q2约为29%至30%。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。
A:我们对第四季度的估计是27%到29%，所以比32%略低。这就是对第四季度的估计。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。
…aswamy: the complete data offering will end up playing a larger and larger role Brent Thill (Jefferies): And when do you expect this batch of go-lives to go up Sridhar Ramaswamy: We gave guidance for Q4. It's a pretty hefty beat and raise. Sridhar Ramaswamy: accelerating go-lives will continue to be a priority Dan [ph] for Brad Zelnick (Deutsche Bank): help frame the impact that migrations had product revenue this quarter Sridhar Ramaswamy: We are super-early with migrations. Sridhar Ramaswamy: we are 15% to 20% of the way through kind of on-prem legacy migrations. Raimo Lenschow (Barclays): Zero Copy comes up a lot in the conversation. Sridhar Ramaswamy: Zero Copy generally comes out in the context of SaaS vendors, Sridhar Ramaswamy: Zero Copy agreements just make data flow more smoothly Sridhar Ramaswamy: We see these as a win-win. Ardy [ph] for Mark Murphy (JPMorgan): explicitly stated that Snowflake budget is now tied to their AI budget, and they're kind of broadening Sridhar Ramaswamy: The strongest pattern that we have had to work hard and earned th…
Sanjit K. Singh (Morgan Stanley): congrats on reasserting 30% product revenue growth in Q4. Brian G. Robins: it's really based on the high stable growth that we see in our core business. Brian G. Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type. Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake. Brian G. Robins: we guide based on the Observe customer behavior up until really the point of earnings. Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
展开本节全部 147 条原文 (支持 64 / 反对 15 / 其他 68，按时间)
- 支持 产品收入 +29%、NRR 125% 稳定、RPO 加速到 +37%
Sridhar Ramaswamy: Product revenue in Q3 was $1.16 billion, up 29% year-over-year. Sridhar Ramaswamy: Remaining performance obligations totaled $7.88 billion, with year-over-year growth accelerating to 37%. Sridhar Ramaswamy: Our net revenue retention remained stable at a very healthy 125%, and we added a record 615 new customers this quarter. Sridhar Ramaswamy: $100 million in AI revenue run rate, achieved one quarter earlier than anticipated, Sridhar Ramaswamy: Because we operate as a consumption-based business, this number reflects real-world enterprise usage.
- 中性 NRR 稳定在 125%，新增客户创纪录 615 家——加速更多来自新客户而非老客户扩张
Sridhar Ramaswamy: Product revenue in Q3 was $1.16 billion, up 29% year-over-year. Sridhar Ramaswamy: Remaining performance obligations totaled $7.88 billion, with year-over-year growth accelerating to 37%. Sridhar Ramaswamy: Our net revenue retention remained stable at a very healthy 125%, and we added a record 615 new customers this quarter. Sridhar Ramaswamy: $100 million in AI revenue run rate, achieved one quarter earlier than anticipated, Sridhar Ramaswamy: Because we operate as a consumption-based business, this number reflects real-world enterprise usage. Sridhar Ramaswamy: It's a direct signal of how customers are using our AI capabilities in production to create value today.
- 支持 传统本地数仓迁移只完成 15-20%，迁移跑道长
It's a pretty hefty beat and raise. Sridhar Ramaswamy: accelerating go-lives will continue to be a priority Dan [ph] for Brad Zelnick (Deutsche Bank): help frame the impact that migrations had product revenue this quarter Sridhar Ramaswamy: We are super-early with migrations. Sridhar Ramaswamy: we are 15% to 20% of the way through kind of on-prem legacy migrations. Raimo Lenschow (Barclays): Zero Copy comes up a lot in the conversation. Sridhar Ramaswamy: Zero Copy generally comes out in the context of SaaS vendors, Sridhar Ramaswamy: Zero Copy agreements just make data flow more smoothly Sridhar Ramaswamy: We see these as a win-win.
- 口径不齐 大型迁移不均匀，管理层用来解释 beat 缩小——迁移是一次性/波动来源
Sridhar Ramaswamy: We clearly invested pretty heavily in our sales and marketing teams in the first two quarters Sridhar Ramaswamy: I don't think of this as an either/or. Mike Cikos (Needham): Is it fair to think that 3Q here was just a strong execution quarter Sridhar Ramaswamy: We've consistently told all of you that we view a 3% beat, as a very good beat Sridhar Ramaswamy: I view the Q3 beat as actually still a very solid beat at some 2.5%. Sridhar Ramaswamy: large migrations are lumpy and not all that easy Sridhar Ramaswamy: there is also a natural variability in a consumption business Brian G. Robins: the quarterly variability is not the right way to evaluate the consumption model. Brian G.
- 支持 签了 4 个九位数 (≥$1 亿) 大单，单季纪录
Brian G. Robins: Our product revenue grew 29% Brian G. Robins: Together, these factors contributed to a stable net retention rate of 125%. Brian G. Robins: Financial services and technology verticals led growth in Q3. Brian G. Robins: Our Global 2000 customers now total 776 Brian G. Robins: each of these accounts spending, on average, $2.3 million on a trailing 12-month basis. Brian G. Robins: We achieved strong booking results, signing four nine-figure deals. Brian G. Robins: This represents a record number of large deals signed in a single quarter. Brian G. Robins: Q3 non-GAAP product gross margins was 75.9%. Brian G. Robins: Non-GAAP operating margin expanded more than 450 basis points year-over-year to 11% Brian G. Robins: Non-GAAP adjusted free cash flow margin was 11%. Brian G. Robins: In Q3, we used $233 million to repurchase 1 million shares at a weighted average price per share of $223.35. Brian G. Robins: We still have $1.3 billion remaining on our original authorization for $4.5 billion through March of 2027. Brian G. Robins: We ended the quarter with $4.4 billion in cash, cash equivalents Brian G. Robins: For Q4, we expect product revenue between $1.195 billion and $1.2 billion, representing a 27% year-over-year growth. Brian G. Robins: We expect non-GAAP operating margin of 7%. Brian G. Robins: We are raising our FY 2026 product revenue guidance. Brian G. Robins: We now expect product revenue of approximately $4.446 billion, representing 28% year-over-year growth. Brian G. Robins: We expect non-GAAP product gross margin of 75%, non-GAAP operating margin of 9%
- 中性 金融服务和科技行业领涨
Brian G. Robins: Our product revenue grew 29% Brian G. Robins: Together, these factors contributed to a stable net retention rate of 125%. Brian G. Robins: Financial services and technology verticals led growth in Q3. Brian G. Robins: Our Global 2000 customers now total 776 Brian G. Robins: each of these accounts spending, on average, $2.3 million on a trailing 12-month basis. Brian G. Robins: We achieved strong booking results, signing four nine-figure deals. Brian G. Robins: This represents a record number of large deals signed in a single quarter. Brian G. Robins: Q3 non-GAAP product gross margins was 75.9%. Brian G. Robins: Non-GAAP operating margin expanded more than 450 basis points year-over-year to 11% Brian G. Robins: Non-GAAP adjusted free cash flow margin was 11%. Brian G. Robins: In Q3, we used $233 million to repurchase 1 million shares at a weighted average price per share of $223.35. Brian G. Robins: We still have $1.3 billion remaining on our original authorization for $4.5 billion through March of 2027. Brian G. Robins: We ended the quarter with $4.4 billion in cash, cash equivalents Brian G. Robins: For Q4, we expect product revenue between $1.195 billion and $1.2 billion, representing a 27% year-over-year growth. Brian G. Robins: We expect non-GAAP operating margin of 7%. Brian G. Robins: We are raising our FY 2026 product revenue guidance. Brian G. Robins: We now expect product revenue of approximately $4.446 billion, representing 28% year-over-year growth. Brian G. Robins: We expect non-GAAP product gross margin of 75%, non-GAAP operating margin of 9%
- 支持 OpenAI 原来在 BigQuery 上做 GPT-4 日志和 AB 测试，后出于性能考虑转到 Snowflake (AI Lab 客户)
对，比如说可以在不同维度上做各种分析，比如比较日度、周度、月度的数据，选取不同用户画像的数据，分析他们对某个话题的反馈等。BigQuery 可以嵌入任意大模型，然后通过 API 直接生成text的结果。OpenAI 以前确用 BigQuery做类似Anthropic的任务，像GPT-4 相关的日志和 AB test 结果都在 BigQuery 上，因为OpenAI当时用的三方实验平台也是基于 BigQuery 做的。后来他们转到 Snowflake，主要还是出于性能考虑。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。 - 反对 专家样本 Q4 FY26 消费增速 27-29%，低于 Q3 的 32%——加速前专家看到的是放缓
A:我们对第四季度的估计是27%到29%，所以比32%略低。这就是对第四季度的估计。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 放缓归因于年末预算重置和假期，消费低但未履约义务在增长
所以，像圣诞节这样的假期是原因之一，但并不是唯一原因。Snowflake的企业客户通常在此期间设定年度预算。决策会放缓，大型合同较不可能签署，这导致第四季度较为疲软。然而，未结义务（outstanding obligations）在增长。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 合作伙伴样本 Q4 消费约 38% (处于此前预测 37-38% 高端)，Q1 估约 40%
A:我们在看到第一季度有加速。这些是估算，但我们相信第一季度可以达到 40%。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 中性 Q1 甚至可能超过 45%，但对比 2025 年末高基数不确定性大
A:完全正确。我们对第一季度的同比估计是 40%，并且我们看到的加速趋势贯穿了去年。如果我要估计，可能会超过 45%，尽管随着季度推进并与 2025 年末强劲增长进行比较，变得更难更快地增长。因此超过 45% 是可能的，但存在相当大的不确定性。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 中性 Snowpark 约占消费 20%，Q1 估 22-23%，年底约 25%；增长是演进而非突变
A:采用在增加，但不是突发性的激增。能力确实在增长，我们与 Snowflake 客户的大部分工作涉及提供数据工程资源，包括使用 Snowpark。总体上这是演进而非革命，并没有导致消费出现突变式增长。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 Snowflake NRR 从峰值 171% 回落到约 125%，Databricks >140%
NRR衡量的是现有客户的扩张意愿。Databricks >140%意味着老客户每年多花40%以上；Snowflake 125%看起来仍然不错，但趋势线令人担忧——它从IPO时的158%一路攀升至峰值171%（SaaStr数据）、再回落至135%、127%、如今约125%。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 支持 ETR 2026 年 1 月数据显示 Snowflake 核心数仓支出意向大幅反弹，Net Score 全数据集上市公司最高
* **易用性**：SQL分析师仍然偏爱Snowflake的使用体验。 * **装机量**：688个$1M+客户、766个Forbes Global 2000客户——这是一笔稳固的经常性收入来源。 * **现金储备**：截至2025年10月（Q3 FY2026），现金、等价物及投资合计约$37亿，有能力通过收购切入邻接领域。 * **数据重力正在回馈Snowflake**：ETR 2026年1月数据显示，Snowflake的核心数据库/数据仓库支出意向大幅反弹，推动其成为ETR全数据集中Net Score最高的上市公司。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 支持 加速由核心业务稳定增长 + AI 工作负载贡献上台阶驱动
Robins: Product revenue grew 30% year-over-year. Brian G. Robins: Our results were driven by stable growth in our core business and a step-up in growth contribution from AI workloads. Brian G. Robins: We saw no decline in our net revenue retention rate, which remains at 125%.
- 支持 RPO $97.7 亿，同比加速到 +42%
Sridhar Ramaswamy: The winners will be the platforms that combine trusted enterprise data, governed business metrics, Sridhar Ramaswamy: With Snowflake Intelligence, we brought enterprise-grade agentic capabilities directly to business teams. Sridhar Ramaswamy: With the general availability of Cortex Code, we extended that to builders, accelerating the entire data lifecycle Sridhar Ramaswamy: Snowflake Intelligence and Cortex Code are meaningful steps in Snowflake's evolution Sridhar Ramaswamy: Turning to our results, product revenue in Q4 grew 30% year-over-year to reach $1.23 billion. Sridhar Ramaswamy: Remaining performance obligations totaled $9.77 billion with year-over-year growth accelerating to 42%. Sridhar Ramaswamy: Our net revenue retention was at a healthy 125%. Sridhar Ramaswamy: We added 2,332 net new customers this year, and we are seeing more and more businesses move over to Snowflake.
- 口径不齐 $4 亿+ 史上最大单来自老客户，已在 run-rate 内——RPO 加速不等于新增消费
Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake.
- 口径不齐 Observe 并购贡献 FY27 约 1pt 产品收入增长
Robins: For FY 2027, we expect product revenue of approximately $5.66 billion, representing 27% year-over-year growth. Brian G. Robins: We expect Observe to contribute approximately 1 percentage point of product revenue growth in FY 2027. Brian G. Robins: We expect FY 2027 non-GAAP product gross margin of 75%.
- 支持 $1M+ 客户 733 (+27%)，$10M+ 客户 56 (+56%)
Robins: We now have 733 customers spending more than $1 million on a trailing 12-month basis, growing 27% year-over-year. Brian G. Robins: bringing a total of 56 customers above this $10 million threshold, growing 56% year-over-year. Brian G. Robins: FY 2026 non-GAAP product gross margin was 75.8%.
- 中性 FY27 收入指引 +27%，其中有机 +26%，约 1pp 来自并购 (Observe 等)
* 2026年收入指引同比27%，其中Organic收入同比26%，参考2025年的Beat幅度隐含31%+
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 中性 Q4 beat 小，公司解释为圣诞节增长慢，也可能与半天 outage 有关
可能也与SNOW在4Q25有一次为期半天的outage有关
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 支持 渠道商 Q1 需求高于计划，产能而非需求受限，业务增速预期从 high 20s 上调到 high 30s
BJ: Absolutely. And again, it's a pleasure. Our Q4, as you recall, was sold out and a lot of it from my vantage point was, I believe, year-end budget exhaustion on the client side, in addition to some programs starting to scale into production. Now Q1 for us is above plan from a demand perspective, and I continue to operate in that manner for Q2 as well. We're effectively capacity constrained rather than demand constrained at this stage. And a lot of it, I've had to expand the team multiple times during the quarter, and we're adding another 30 to 35 people, senior executives, to go after the space and also deliver against it. My guidance initially was high 20s year-over-year growth. I believe directionally I'm at the point where I can confidently project high 30s as long as I'm delivering with quality.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 口径不齐 high 30s 是 Thoughtworks 自身咨询业务增速，而非 Snowflake 消费增速
BJ: Absolutely. And again, it's a pleasure. Our Q4, as you recall, was sold out and a lot of it from my vantage point was, I believe, year-end budget exhaustion on the client side, in addition to some programs starting to scale into production. Now Q1 for us is above plan from a demand perspective, and I continue to operate in that manner for Q2 as well. We're effectively capacity constrained rather than demand constrained at this stage. And a lot of it, I've had to expand the team multiple times during the quarter, and we're adding another 30 to 35 people, senior executives, to go after the space and also deliver against it. My guidance initially was high 20s year-over-year growth. I believe directionally I'm at the point where I can confidently project high 30s as long as I'm delivering with quality.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 口径不齐 渠道商 Snowflake 业务 Q1 增速 24%，低于上季 30%，全年目标 25%
FC: Sure. Yeah, let's start off with the quarterly numbers since, yes, we had a great quarter last quarter, which was at that 30% mark, which set that high watermark for us. For this most recent quarter, we fell in at 24% growth, which is still good. It's still very good. But again, with that high watermark of that 30%, we were sort of expecting a few more Snowflake projects to roll through. It was slightly below last quarter. However, we're still on track or ahead on the year since we're over the 25% mark that we're setting for ourselves.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 中性 瓶颈是工程师产能而非需求：招了 5 人，实际需要 10-15 人
We need a coding agent for data and analytics, and that's exactly what Cortex Code brings to the table. And so we've really, in order to meet that gap that I was talking about last quarter, we've had to find these efficiencies using Cortex Code. Since, yes, we hired 5 more engineers, but of those 5, we needed probably 10 to 15. So we had to supplement using Cortex Code to become more efficient and really do the job of that plus 15.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 Cortex Code 在上个项目上带来约 40% 效率提升，进场首日 '100% 效率'
FC: 5 new engineers. I wouldn't say that it has given me a 200% efficiency, because that would be very boastful. But I would say from a project perspective, we are gaining... I mean, on that past project, we made ourselves about 40% more efficient for these projects. Somewhere around there where it allows you to hit the ground running and really make it even though you might not know these environments, because we're coming in cold, right? We come into a client and we don't understand what they've set up already. And so having that efficiency of quickly having Cortex Code scan everything, understand all the schemas, understand a lot of the code bases where we literally will just plop in the code base that they have today, tell me what this is doing and what the purpose of it is, which pipelines is it supporting, and then that allows us... so we've generated this sort of hit-the-ground-running code base where we literally have these already pre-generated and then as soon as we land on day 1, it's hitting the beach running where we kick off all these scripts and we collect all the information. It populates into our work areas and now it's like we have a workbook that immediately, even if the client doesn't have documentation, we generate our own from scratch
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 中性 项目块状：约 8 个 pipeline 项目，单个约 3000 万美元、18-36 个月
FC: That's right. We definitely need to close on those project wins. But again, these are $30 million projects, right? So these are big projects for a lot of these enterprises that we're talking about. So yes, it is going to be a little chunky to your point where we close on a couple of those. A lot of project work, and this is why also my hiring and bringing new Snowflake engineers on board is also very chunky, right? Because suddenly we need you BB: Yeah. FC: know, 5 to 10 more engineers to help with this project. And it's a potential 18 to 36 month project, so yes, as you can imagine, it will be chunky. And it's also how quickly we can do these comparisons and bake-offs with their competitors.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 北美伙伴 A 的 consumption 增速从 3Q25 的 20-25% 升到 2Q26E 的 30-35%，但预计 2H26 回落到 25-30%
* 3Q25: +20–25% YoY * 4Q25: +25–30% YoY * 1Q26: +25–30% YoY * 2Q26E: +30–35% YoY * 2H26E: +25–30% YoY
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 反对 伙伴 A 认为 2Q26 签约客户过多，2H26 会出现季节性疲软 (Funda 认为专家混淆了咨询收入与消费)
* The expert believes too many customers were signed in 2Q26, leading to some seasonal softness in 2H26. (Note: We believe the expert is conflating Consulting Revenue with Consumption — this is not a negative read-through for 2H26 Consumption.)
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 反对 伙伴 C 尽管 1Q26 强劲，仍下调 2Q-4Q26 预测，原因是一家大型金融客户在优化，抵消了 CoCo/Intelligence 的上行
* “We had previously revised our expectations. Despite a strong 1Q26, we have modestly cut our 2Q26–4Q26 forecasts, and we now expect 2Q26 to land at the lowered target.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 伙伴 D 1Q26 consumption +28-29% 高于 24-26% 预期，全年需上调约 2%；驱动包括 Intelligence/CoCo、数据现代化、从 Redshift/Fabric 迁入
* Multi-cloud migration trend: customers migrating off Redshift, or shifting away from heavy reliance on Microsoft Fabric toward third-party solutions like Snowflake
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 CoCo 1Q26 贡献小是因为大量免费试用 credit，3-4 月才开始计费；客户学会使用后发现很贵、部分不满
+ Customers received large amounts of free trial credits and subscriptions, started experimenting in January–February, and began incurring charges in March–April. + “Snowflake has been giving the product away for free to drive learning. What we’re finding is that as customers learn, they’re also discovering how expensive it is. Some customers are unhappy.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 伙伴 D 称 AI 相关消费占比从 3Q25 约 20% 升到 1Q26 近 30%，年底 agentic AI 消费可能超过数仓 (口径宽，含 Snowpark)
* By year-end, agentic AI consumption may exceed data warehousing (since DWH consumption stabilizes once mature) * Agentic AI includes: the Cortex AI suite (AI functions, AI search, Cortex analytics), Snowflake Intelligence, Snowpark, Cortex Code, Cortex Code SDKs, etc. * AI-related Consumption mix:
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 CoCo 让遗留迁移快 45-50%、管道开发快 30-35%、分析看板快 20-25%，迁移类收益最大 (指向一次性迁移加速)
- Legacy migration: 45–50% faster (high end, since requirements are already defined and the work is mostly development and testing) - Analytics and dashboarding: 20–25% faster (more user dialogue, data preparation, ingestion, and formatting)
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 Funda 认为 OpenAI 到 1Q26 已成为 Snowflake 最大客户之一，AI Lab 的 mid-training/RL 消费是增量
**4. In this quarter’s Preview, we believe that as OpenAI’s demand continued to grow in 4Q25, by 1Q26 OpenAI has become one of Snowflake’s largest customers**
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 Q1 FY27 产品收入 $1.334B，同比增速从 Q4 的 30% 加速到 34%，创历史最大环比美元增量
Snowflake Intelligence gives business users a natural language interface to enterprise data, context, and actions, while CoCo gives builders a natural language way to create applications, pipelines, agents, and workflows directly on Snowflake. Snowflake is uniquely positioned to help customers become agentic enterprises, as evidenced by our Q1 results. Product revenue came in at $1.334 billion, with growth accelerating to 34% year-over-year, up from 30% last quarter and 26% a year ago, marking our strongest sequential dollar growth in company history. Our net revenue retention rate increased to 126%. And with our continued focus on executing with discipline and operational rigor, our Q1 non-GAAP operating margin expanded over 300 basis points year-over-year to 12%.
- 支持 NRR 升至 126%，老客户扩张是加速来源之一
Snowflake is uniquely positioned to help customers become agentic enterprises, as evidenced by our Q1 results. Product revenue came in at $1.334 billion, with growth accelerating to 34% year-over-year, up from 30% last quarter and 26% a year ago, marking our strongest sequential dollar growth in company history. Our net revenue retention rate increased to 126%. And with our continued focus on executing with discipline and operational rigor, our Q1 non-GAAP operating margin expanded over 300 basis points year-over-year to 12%. I want to take a moment to touch on our outlook.
- 中性 CEO 把加速拆成三部分：AI 拉动核心数据平台、AI 产品本身、CoCo 带来的二阶核心消费 (含迁移)，但未给量化拆分
Thank you very much. Absolutely. No, I would break this up into three parts. First, AI is accelerating the value that people can get from the data that they have put into Snowflake or that they can put into Snowflake. So we saw a – like a healthy secular tailwind for our core data platform. And part two is really that agentic products, the control plane products like Snowflake Intelligence and Cortex Code/CoCo came into their own in Q1. Recall that CoCo went into GA on Feb 5, so just as we were opening up the quarter. And we've seen very strong traction with both the products. And the really interesting thing with Cortex Code is that it, in turn, drives more consumption on the core data platform simply because it's much easier to get projects done, whether it's a pipeline or creating a new agent, or setting up a new dynamic table or even, honestly, a migration. So it's driving the second order effect as well. But it's really – this is the one, two, three, and that's why I like to think of this as AI compounding Snowflake's strengths in data.
- 支持 迁移时间从约 2 年缩到 1–2 个季度——这是 thesis 第 (3) 层'一次性迁移加速'的直接证据，意味着迁移收入被提前
Maybe there's a number of other ones that are coming, so maybe just help us think about how that changes the potential spend profile of a customer over time. Among the biggest, like, impact that products like CoCo have with our customers is simply one that of expectation. I talked, again, in my remarks about how we did a two-year Teradata migration. We are engaged in more migrations, but the timelines for doing those now run between a quarter and two quarters. Why? Both my team and the customer expects and demands that, and we have the ability to deliver against that. I think the – both the impatience and hunger for what people can do with data, along with the expectation for how quickly we can get them done, I would say that's like – that's a huge sea change. Christian? No, align with what you're saying, Sridhar. There's a massive backlog...
- 支持 新客户净增 616 家 (+38% YoY)，新增 13 家 G2000 (去年同期 4 家)，'公司史上最多净新增'
JB brings more than a decade of experience with Snowflake, deep knowledge of our customers and platform, and a strong operational focus as we continue to evolve our go-to-market motion. That strong execution is translating into continued customer momentum and broader product adoption of the platform. In the quarter, we added 616 net new customers, up 38% year-over-year. We're also seeing customers deploy and scale workloads at a faster pace. The number of use cases, individual projects managed on Snowflake deployed in the quarter increased 114% year-over-year as customers moved more workloads into production on the platform.
- 支持 部署的 use case 数同比 +114%，每个 AE 赢得的 use case +86%
In the quarter, we added 616 net new customers, up 38% year-over-year. We're also seeing customers deploy and scale workloads at a faster pace. The number of use cases, individual projects managed on Snowflake deployed in the quarter increased 114% year-over-year as customers moved more workloads into production on the platform. At the same time, the number of use cases won per account executive increased 86% year-over-year, underscoring both growing customer demand and improved sales execution across the organization. We are also continuing to strengthen our ecosystem as we deepen our strategic partner relationships and extend the reach of our AI Data Cloud. Just today, we announced an expanded collaboration with AWS through a new $6 billion multiyear agreement to accelerate enterprise AI adoption globally, leveraging Graviton compute and AI services.
- 中性 Observe 在 Q1 贡献 <1pt 产品收入增长，全年约 1pt
In Q2, we expect product revenue between $1.415 billion and $1.42 billion, representing 30% year-over-year growth. Our Observe acquisition is progressing well, consistent with our initial expectations. Observe contributed less than 1 percentage point of product revenue growth in Q1, and we continue to expect the acquisition to add approximately 1 percentage point of revenue growth – product revenue growth for the full year. Turning to margins. We expect 75% non-GAAP product gross margin for FY 2027.
- 口径不齐 $1M+ 客户数转录为 79 (Q1 新增 46)，与 Q2 披露 828 不衔接，应为约 779/780，疑转录错误
Snowflake's AI workload is now a significant revenue engine in its own right. AI products like Cortex Code are expanding our opportunity with existing customers as CoCo encourages faster, more consumption of the data platform. We now have 79 customers spending more than $1 million on a trailing 12-month basis. 46 customers crossed the $1 million threshold in Q1 compared to 26 a year ago period. Remaining performance obligations grew 38% year-over-year, compared to 34% in Q1 of last year. We continue to see customers favor Q4 renewals.
- 支持 FY27Q1 产品收入 +34%，超预期 5.3%，剔除 Observe 并购仍超 4%
* 1Q26 product revenue grew 34% YoY, beating by 5.3%—the largest beat in recent history. Even excluding the Observe acquisition, the beat was still 4%. * 2Q26 revenue guidance implies 30% YoY growth. Applying the typical 3% beat, this implies 34% YoY growth—meaning growth is still not decelerating. * FY26 revenue guidance was raised from 27% to 31%, one of the largest full-year revenue raises in Snowflake’s history. Based on this guide, Snowflake still believes growth will not decelerate for the full year. * 2Q26 OPM guidance of 12.5% also came in 1% above expectations. Margins were not impacted by Cortex Code.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 Top 25 客户年均消费 $34M (两年前 $22M)，G2000 客户 FY26 平均仅 $2.4M，暗示老客户扩张空间
And over the five-year period, our market has grown roughly 30%. We really have conviction in this when we look at our large customers. And so, the top 25 large customers, they spend on average $34 million a year with us. That's grown over the last two years from $22 million a year. When we look at our Fortune 2000 customers, G 2000 customers, in FY 2026, they on average have only spent $2.4 million with us. We feel that we have the right to actually increase those customers up to our large customer spending. We'll jump into the core growth drivers of the business primarily in the core data platform and our AI workload.
- 支持 迁移数量 FY25→FY26 增长 1.9 倍，use case 增长 1.7 倍
And so, with the land motion then they expand over time, with the high gross retention rate, this is really a powerful revenue engine for the company. We do that through a number of different ways but I want to touch on migrations and use cases. Migrations from FY 2025 to FY 2026 grew 1.9 times. Use cases grew 1.7 times. Very meaningful increase. We actually are able to get this wallet share from a number of different sources and where the real benefit comes in for us and our customers is when they're consolidating all this into a single platform as Snowflake.
- 口径不齐 新客户达到消费的时间从 10 个月缩到 7 个月、迁移快 40%：这是时间前移 (一次性台阶)，不一定是终态消费更大
You would expect with more choices and more evaluations out there that the sales cycles actually expand and actually take longer. They are actually doing the opposite of that so the sales cycles are accelerating. We constantly talk about how we are using our AI tools to actually get new customers to consume faster. We've taken that from ten months down to seven months and we're continuing to see how we can decrease that. And then for all customers, migrations are accelerating. And so, we've shown a 40% improvement. And so, AI across the business is having a dramatic improvement on our time to ramp. All right.
- 支持 上季基础业务 (base) 加速，归因于 CoCo 和行业顺风，未量化
Before I dive into the specific examples, I'll talk about some broad trends that we're seeing. And so, one of the things that we talked about on the last earnings call was their secular tailwinds that we actually benefit from in the overall industry. And then second is CoCo is actually expanding our personas that we're selling into and allow more people to consume. And then thirdly, in our base business, we saw an acceleration in the base and last quarter as it relates to CoCo and the secular tailwinds. And so, this particular customer was a large customer that was an equipment company and they adopted Snowflake CoWork. And they adopted CoWork because they had over 1,600 locations and the reps were having a tough time getting the information out to the customers and answering in a unifying way.
- 中性 FY26 账单增速高于收入：客户提前用完容量后选择 addendum 或提前续约，带来 billings 噪音
Two years before, that was 80% but really remains consistent. There is some noise between billings growth and revenue though. And so, you can see on the far left-hand side, in FY 2024, revenue grew 36% and we had 29% billings growth. But in FY 2026, it is actually reversed. When our customers actually come to the end of their capacity or use the capacity that they purchase, they really have two choices and we allow them to do either the following. They can actually do an addendum and actually extend the current contract to when the contract renewal comes up or they can actually do an early renewal. And so, this causes some noise between billings and revenue. But over a long period of time, this actually normalizes out.
- 支持 新客户数和新客户 ACV 同比都显著增长 (未给数字)
We get enormous leverage. And we have talked to you many times about things like new logos. And this was a remarkable quarter for us because both the number of new logos that we won and the ACV, the total contract value that we got out of these new logos, both went up significantly year-on-year. That's because there are a set of people who are obsessed about this motion, who are obsessed about getting these customers onboarded, getting these customers live. And that's the efficiency that I push for, that we push for.
- 口径不齐 CoCo 收入不单独披露，各客户用量差异大，消耗要随时间才看得清；收入贡献靠带动计算/存储/token (没有量化)
每个客户对 Cortex Code 的使用量各不相同，具体的消耗情况需要随着时间的推移才能明朗。但使用 Cortex Code 会带动其他附加服务的消耗，例如计算、存储和 Token 用量的增加，这将为 Snowflake 贡献相当可观的收入。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 口径不齐 其组合/公司在 Snowflake 上的消费 Q1 同比约 +30%，Q2 跟踪略高于 +35% (主体是谁不清楚)
回 答 者：去年至今，我们的 Snowflake 消费量持续增长，主要得益于数据平台采用率的扩大、AI/ML 工作负载的增加以及业务驱动型分析消费的提升。第一季度，消费量同比增长约 30%；进入第二季度，随着更多生产工作负载迁移至我们的平台，该增速追踪至略高于 35%。增长的主要驱动力来自数据工程、自助式分析以及新兴的 AI 使用场景，特别是在 AI 和基于代理（Agent-based）的应用中，Snowflake 被用作数据治理层。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 Q2 加速的主要驱动不是传统数仓，而是 AI 产品组合带来的增量工作负载
回 答 者：第二季度的增长确实强于第一季度，其主要驱动力并非来自传统的数据仓库业务，而是源于 AI 产品组合，特别是 Cortex 相关的增量工作负载，这包括 Cortex AI、Cortex Agent Search 和 Cortex Code。此外，Snowpark 的更广泛使用、Apache Iceberg 的相关项目以及生产环境中代理工作负载带来的更高计算消耗，共同推动了第二季度的加速增长。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 贡献排序：Cortex AI Agents/Search 最高，CoCo 次之，Snowpark 第三，核心数仓和分析只是中等，Iceberg 适中但在增长
如果按贡献度排序，我认为贡献最高的是 Cortex AI Agents 和 Search；其次是 Cortex Code，贡献度非常高；再次是 Snowpark，贡献度高；核心数据仓库和分析业务的贡献为中等；最后是 Apache Iceberg，其贡献度适中但正在增长。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 口径不齐 按他的组合，CoCo 约占增长贡献 15-25%，AI agents 和 search 约 50%；年底 CoCo 升到 25-35% (是增量占比，不是消费占比)
回 答 者：从我的产品组合来看，Cortex Code 的增长贡献率在 15% 到 25% 之间，而 AI 代理和搜索的贡献率约为 50%。到今年年底，我预计 Cortex Code 的贡献率将达到 25% 至 35%。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 Q2 里 Cortex AI 生态是最大贡献者，核心平台扩张对增长的贡献在下降；Snowpark 约占增量的 10-20%，是 AI 的下游效应
可以说，Snowflake 以前主要用于存储数据以供查询，但现在模式已转变为：存储数据，然后构建并运行应用程序，并在其上运行 AI 工作负载。这为每个客户创造了更多的消耗。从第二季度的数据来看，Cortex AI 生态系统是最大的增长贡献者，核心平台的扩张对增长的贡献有所下降。Snowpark 是一个重要但次要的贡献者，约占增量增长的 10% 到 20%。但这部分增长很大程度上应被视为 AI 采纳的下游效应，而非独立的 Snowpark 需求。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 口径不齐 摘要稿把 CoCo 写成占 Q2 消耗的 15-25% (转写原话是增长贡献，而且专家说直接消耗很小)，这个说法被放大进了标题
A:：Cortex Code 在第二季度大约贡献了 15% 至 25% 的消耗，而 AI 代理和搜索约占 50%。到年底，预计 Cortex Code 的贡献将约为 25% 至 35%。
来源背景：同专家B。本稿标题给出身份“海外头部咨询公司离职专家”，作者 Lantian，没有给职位和年份。视角局限同专家B。 - 口径不齐 他的 Snowflake practice 环比增长约 10%，没有同比数字 (practice 收入 ≠ Snowflake 消费)
回 答 者：我们正在增长，业务表现良好，但我手头没有确切的同比增速数据。不过，我们的环比增长率大约在 10% 左右。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 Q2 的驱动是 CoCo (提升开发者生产力) 和平台规模化后自然增长的消费；Snowflake Intelligence 带来一些早期收入
回 答 者：Cortex Code 是驱动因素之一，它提升了我们开发人员的生产力，进而推动了整体消费。此外，Snowflake 业务的一个典型特征是，一旦开始规模化并构建数据架构，消费量通常会随之持续增加。除了 Cortex Code，Snowflake Intelligence 也是一个驱动因素，它激发了客户的兴趣并带来了一些早期收入。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 Snowflake 的 AI 相关收入增长很快，今年可能翻倍甚至更多
回 答 者：我仍然认为 Databricks 在人工智能、机器学习和基础模型方面处于领先地位，而 Snowflake 的优势在于其源自数据仓库背景的、以 SQL 为主导的数据环境。不过，Snowpark 确实在快速增长，Snowflake 的人工智能相关收入增长也非常迅速，预计今年可能会翻倍甚至超过一倍。从这个意义上说，他们正在追赶 Databricks。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 Benchmark 称 AI 产品在加速迁移和提升消费，time-to-value 从月缩短到周
Benchmark raised its PT on SNOW to $290 from $270 and reiterated Buy following Snowflake Summit and Investor Day, arguing the company is increasingly positioning itself as the agentic AI control plane for the enterprise. The firm highlighted strong traction for CoCo and CoWork, noting AI products are accelerating migrations, increasing consumption, and helping customers move from experimentation into production faster. Benchmark also pointed to improving operating metrics, including faster sales cycles, expanding use cases, stronger customer acquisition, and a reduction in time-to-value from months to weeks. The analyst believes Snowflake is successfully shifting from a data platform to a broader AI workflow and orchestration platform, supporting sustained growth and a path to GAAP profitability by FY28.
来源背景：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 - 支持 本地客户把数据从 Cloudera 接入 Snowflake/Databricks 用于 AI 与现代分析，高级用例流向两家
**AS** [00:37:21]: As far as the existing workloads are concerned, they are not very successful. What I've seen is more customers basically create a pipeline from Cloudera into Databricks or Snowflake so that they can leverage the AI modelling tools, the modern analysis tools that they run. I have not heard a lot of success stories in terms of Snowflake and Databricks being able to take away on-prem workloads from Cloudera.
来源背景：Ashish Shah，前 Cloudera 高级产品经理 (2016 年起在 Cloudera 9 年多，先后任销售工程师、ISV 合作伙伴管理 2018-21、产品管理，最后三年负责 SDX 治理组件，明示)；现任 Dataiku Technical Account Manager 。视角局限：Cloudera 本地/混合视角；对 Snowflake 的认识来自竞争丢单和朋友转述 ('I don't have any direct evidence')；离开 Cloudera 时间约 2025-2026。 - 反对 消费模式普遍存在承诺额用不完的问题，marketplace 是消化承诺的手段
**TL** [00:05:30]: Yes. With consumption, as we've seen with all of the players, even Anthropic recently, they announced their own marketplace, a company only creates a marketplace if they're struggling to get their customers to realise their commits and realise their obligations. Snowflake call it an RPO, remaining performance obligation. "Hey, I've committed USD 2m, but I've only computed USD 1m. How do we make up that shortfall?" The answer is often a marketplace. I think there's only so many times where that makes sense as a model. The biggest friction point is you can win the contracts, and you can win large contracts in this space, but then actually getting people to utilise their full commitment is a big challenge. It's a challenge that every single vendor sees regardless of size. That's exactly why AWS, Microsoft, Google, they all have marketplaces for this specific reason. I don't see how that trickles down to the smaller vendors where it's easy for them to have a long-lasting consumption model that allows them to recognise that revenue as computed or used commit and the trends that we're seeing across the industry, but also, what you would typically see is somebody commits, but they don't utilise that commit, that then creates friction. That friction then creates a need to find a way to use that commit.
来源背景：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 - 支持 Databricks 存量客户新增收入数据和 AI 各占一半；agent 上线后 ETL 频率从每周变每日，数据收入翻倍 (AI 拉动核心数据消费的机制)
在过去12个月，Databricks的机器学习与GenAI相关产品的年度同比增长率为184%。从存量客户来看，新增收入由数据和AI两大产品线大致平分。AI产品增量最高的来自AgentBricks，该产品帮助客户构建智能体，其收费模式包括调用大模型的Token费用以及构建Agent本身消耗的计算资源（Model Serving）。以太古可口可乐（Swire Coca-Cola）为例，作为AgentBricks用量最大的客户之一，该公司已使用Databricks的数据产品（如ETL、Data Warehouse）三年。其主要应用场景是构建了两个智能体：一个面向内部员工，作为知识库助手（Knowledge Assistant），约有两三百名活跃用户；另一个规模更大，供销售部门使用，通过嵌入在移动应用中的Agent，支持拍照上传、查询库存、销售额及促销活动等功能。仅后一个Agent的Model Serving费用每月就达到约2万美元。为支持这两个智能体，该公司的数据ETL频率从按周批次处理提升至每日一次，使其数据方面的收入也实现翻倍。对于新增客户，其收入的80%至90%来自AI产品本身。这类客户，尤其是数字原生企业（Digital Native Business, DNB），希望快速利用Databricks平台调用各类大模型。一个典型的新增客户是Minimax，作为Databricks全球第三大AI客户，其主要使用AI Gateway和Smart
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 中性 从首次接触到 AI 应用上生产通常约 8 个月，所以 2026 年才显著放量
AI应用场景的丰富化是加速客户数据上传意愿的主要驱动力。例如，远洋运输公司OOCL需要导入其从不删除的五年历史数据，金融服务业的客户如渣打银行、泰国银行、AXA及FWD保险等，则有大量非结构化的单据、图片、PDF等数据需要处理。数据量越大，可支持的机器学习场景就越丰富。尽管AI技术已存在数年，但相关业务收入的显著增长发生在今年（2026年），这与客户的采纳周期有关。一个客户从首次接触（First Pitch）到最终在生产环境上线AI应用，通常需要大约8个月的时间。这个过程包括初步接洽、小型POC验证、商业流程（询价与竞品分析）以及最终的生产环境部署。其中，POC阶段最为复杂且耗时，通常需要约两个月的高频接触（每周两到三次），以明确业务用例并设计解决方案。因此，两年前开始接洽的客户，其需求在近期才逐步转化为生产环境的实际用量和收入。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 口径不齐 专家给出的'Q1 +26%、NRR 124%、收入 $997M'实为 FY26 Q1 公司披露数，不是本年 (FY27) 数据
回 答 者：今年消费同比增长约26%，NRR为124，净续约率为124%。第1季度收入约为997 million美元。第2季度的指引约为31%，但给出的增长指引约为32%。与去年同期相比，增速有所加快，收入约为1.909 billion美元，增幅为32%，NRR为125%。总体而言，尽管此前指引约为26%，但实际增速更接近32%。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 上季度新增客户中超过 50% 由 AI 能力驱动
回 答 者：业绩加速增长的最主要原因，是许多客户开始采用AI。Snowflake上季度新增客户中，超过50%主要受其AI能力驱动。此外，上季度部署的新用例占比超过26%。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 口径不齐 专家把加速归因于 AI：Q2 指引约 26%、实际 32%，但这是 FY26 Q2 的旧数
特别是当企业拥有数100GB或数TB数据时，每次查询都可能消耗大量Token和Snowflake计算能力。因此，Agent是Snowflake应用及消耗增长的最大驱动力。Snowflake最初对Q2收入增长的指引约为26%，但由于AI带动Snowflake被大规模采用，增长率提升至32%。这也推动Snowflake股价大幅上涨，AI是推动Snowflake整体股价上涨的因素。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 SnowConvert 等 AI 工具把 50 人 6 个月的迁移缩短到 2 个月 (30% 以内)——迁移加速是一次性拉动的来源
目前，我们能够提供智能体能力、各类大语言模型以及大量自动化工具，帮助客户完成工作。以前，如果客户拥有一支50人的Snowflake团队，手动迁移需要6个月；现在借助Snow Convert等Snowflake AI工具，同样的迁移可以在2个月内完成，或者将时间缩短至原来的30%以内。客户由此能够大幅节省迁移投入和迁移时间。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 渠道口径 Snowflake 支出增速处于 20% 高位到 30% 低位，Q2 约 31%，略高于预期
回 答 者：支出增速基本处于20%高位至30%低位。第2季度同比增长约31%，处于我们预期的较高水平。我们并未进行非常详尽的预测，但该增速高于预期。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 Q4 约 29%、Q1 约 30%，逐季小幅加速
回 答 者：如果没有记错，Q4的增速为29%，Q1为30%。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 SI 侧 Snowflake 工作量增加且加速，支持 Q3 33-34%，Q4 持平或略高；但看不到签约
回 答 者：我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购，因此我们通常不了解具体的签约情况。不过，从系统集成商的角度来看，Snowflake 相关工作量正在增加，而且增速有所加快。然而，系统集成商能否承接并转化这些增量需求，仍存在疑问。总体而言，我观察到的情况支持 Snowflake 在 Q3 实现约33%-34%的季度增长，Q4 的增速可能与此相同或略高。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 口径不齐 摘要标题写明'Snowflake 2026 财年第二季度收入同比增长 32%'，即 FY26 Q2，不是本年
title: "Snowflake离职专家|Snowflake 2026 财年第二季度收入同比增长 32% 至 19 亿美元，远超 26% 指引，由于 AI 采用推动，超过 50% 的新客户与 AI 相关，AI 账户增至 6,200 个以上；超过 50% 的 AI 用户采用 Cortex Code，开源权重模型加速计算消耗，Agent 使用将用户基数扩大十倍，剩余履约义务达 67 亿美元，近期增长指引维持在 30–35% 区间"
来源背景：Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。 - 反对 该大客户 Snowflake 支出同比 10-15%，远低于公司整体 30%+
回 答 者：按照 year over year 的口径，同比增长应该为 10%到15%。由于体量较大，有时会高一些，有时会低一些，但每年的支出增幅基本处于这一范围。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 环比平均至少 3-5%，有时持平，按项目里程碑付款
如果取平均值，quarter over quarter 并不是以 10%的速度增长，但至少会有 3%到5%的增长，整体呈 positive、upward trending。有时某个季度也可能是 flat，但今年总体规划的投资规模是确定的，因此不同季度可能有高有低，总量仍在增长。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 很少提前作 commitment，项目真正进入开发使用才签单 (对 RPO 的含义：保守客户不提前锁量)
回 答 者：我不了解其他公司的操作方式，但就我们而言，很少会提前作出 commitment 或签订意向订单。根据我个人在目前公司及此前一些大公司的工作经验，这种实践比较少见。我们一般不会提前作出承诺并签单。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 反对 渠道口径整体消费增速约 25-27%，过去 3 季度稳定，未见加速
回 答 者：整体增速约为25%至27%。过去3个季度的增长较为稳定，均保持两位数增长。今年的增长趋势维持在25%至27%左右，与去年的增速相近。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 反对 企业采用 AI 受治理、合规、成本制约，今年预算基本用完
回 答 者：从AI的整体情况来看，企业组织采用AI并不容易，其中涉及大量治理和合规工作，同时还要考虑成本。今年客户的预算基本已经全部用完，因此很难继续投入与AI相关的项目。AI对客户而言尤其具有挑战性，因为与各合作伙伴提供的能力相比，目前任何企业组织实际使用的AI都非常有限。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 新客户贡献约 29-30%；老客户合同规模和期限增幅超过 20%，出现超过 1 年的合同
回 答 者：净新增客户的增长趋势约为29%，最高达到30%，这对我们而言是一个非常积极的信号。从新客户采用Snowflake的情况来看，他们从一开始就在Snowflake上投入了较多支出。 现有客户方面，无论是合同规模还是合同期限，增幅都确定超过20%。我们还看到合同期限有所延长，出现了超过1年的合同。因此，我们的收入实现了大幅增长，原因在于客户持续使用Snowflake。此外，部分客户正在将AI投入生产环境，虽然数量不多，但确实已有一些客户开始这样做。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 FDE 把迁移到 Snowflake 的周期 (原 6 个月-1 年) 缩短约 50%，降低迁移成本——迁移拉动的来源
FDE能够有效节省时间。就迁移至Snowflake所需的周期而言，FDE可以将时间缩短50%。从成本来看，FDE也大幅降低了迁移成本。此外，FDE有助于提升整体使用量，因为大量相关人员会接受培训，了解如何使用Snowflake、如何迁移至Snowflake以及相关流程。FDE还有助于从技术层面建立客户与供应商之间的关系，因此是一项很有吸引力的举措。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 部分客户已大量投资但硬件尚未交付，环境未就绪——明年才完全可用
回 答 者：增长将持续到今年年底。到明年，许多客户的环境将发生变化。例如，目前已有客户对大型环境进行了大量投资，但我们尚未向其中许多客户交付硬件，因此这些环境还没有完全准备好。到明年，无论是在IT支出、环境、能力还是整体准备程度方面，客户都将完全具备使用条件。客户使用得越多，后续需求也会越大。因此，我认为AI工作负载将持续增长。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 专家口径 Q4 约 30%、Q1 约 32%，Q2 预计低 30% 区间
A:我看到 Snowflake 的消费有一个令人愉快的加速。第四季度，我会把消费增长描述为约 30% 的同比增长。第一季度，我认为超过 30%，大约在 32% 左右，这与上一季度相当一致。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 反对 预计 Q2 较 Q1 温和回落 2-5 个百分点 (实际 Q2 FY27 报 37%，专家判断方向错误)
A:第二季度仍在进行中。轨迹上，我可能会说差不多，但我们最终可能会比第一季度略低，因为第一季度增长相当快且较高。我们在第一季度看到特别强劲的加速。对于第二季度的消费增长，我预计相比第一季度会温和下调大约 2 到 5 个百分点。但我仍然预计它会徘徊在大约 30% 左右。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 按产品拆分消费份额：核心传统 Snowflake 60-65%、Snowpark/ML 20-25%、Cortex AI/agents 10-15%、其他约 5%
A:按产品线划分，核心的传统 Snowflake——仓库计算、SQL、存储、BI 和 ELT——大约占消费份额的 60% 到 65%。Snowpark 或机器学习占 20% 到 25%。然后 Cortex AI、生成式 AI 和 agents 占 10% 到 15%。像流式应用和协作工具等其他部分大约占 5%。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 增速拆分：传统 20-25%、Snowpark/ML 约 40%、Cortex/agents 70%+；增量已不主要来自核心数仓
在增长方面，传统 Snowflake 的增长大概在 20% 到 25% 左右。Snowpark 和机器学习的同比增长约为 40%。Cortex 和 agents 的同比增长超过 70%，而其他领域我估计大约为 30%。核心的数据仓库仍然是一块大饼，但增量增长已经不再主要来自这里，因为 Snowpark 正在把 Snowflake 从 SQL/BI 扩展到以 Python 驱动的数据工程和机器学习工作负载。Cortex 在结构化数据处理上增加了推理和 agentic 工作负载，Cortex AI 本身的定价现在主要基于消费，包括基于 token 的计费。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 获得实质 FDE 参与的账户 12 个月内约多 25% 消费；影响在 2026 下半年到 2027 年初显现
通过将FDE工程师直接放在你的战略客户身边，你可以缩短部署周期，更重要的是，在该AI应用之下创造额外的核心Snowflake消费。对于获得实质性FDE参与的账户，我可以看到在12个月内相比过去大约25%的额外消费，因为由FDE驱动的Cort ex应用不仅消费Cortex本身——它还推动仓库计算、数据工程、存储、检索和重复的代理查询。这是我的观点。实际上有三个商业影响：第一，更快地将PoC转入生产；第二，扩展到那些可能本来会流向Databricks、AWS或Azure的工作负载； 第三，在现有Snowflake账户内带来更高消费。这就是我认为FDE更像是一个扩展和消费引擎，而不是传统的许可销售引擎的原因。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 AI 采用到收入有滞后：试点进入生产、用户增加、agent 重复调用后才产生有意义消费
我估计如果第一季度是 34%，第二季度的底层消费可能在上下浮动约 4% 的范围内，因为在第一季度仍处于试点或早期生产的 AI 和 agent 工作负载，我预计会转入商业化生产。这就是为什么我说大约 30% 更为保守——这是我们对该季度的预期。它不应被解释为底层消费的运行速率，因为 AI 采用不会立即转化为收入。客户可以在第一季度启用 Cortex 或构建 agent，但有意义的消费通常在工作负载进入生产、用户量增加、推理频率上升以及 agent 开始重复进行自主数据调用时才会开始。这会产生消费的滞后效应，因此预期会出现商业化收益。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 只有 5-10% 客户把 agent 投入生产，且仅 1-3 个——AI 消费仍极早期
从这个角度看，我们确实才刚刚起步。在我负责的客户群体中，可能只有5%至最多10%的客户已经将智能体投入生产，而且应用仍然非常有限，可能只有1个、2个或3个。除此之外，他们还在进行大量测试，但尚未进入生产阶段。在这些客户中，我认为95%的相关工作确实是在Databricks上完成的。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 中性 Databricks 非 AI 业务靠数据重力有机增长，非 AI 其实靠 AI 带动，没有新产品
非AI其实还是靠AI带动。Databricks希望客户把数据平台留在Databricks这上面，因为数据本身有gravity，只要数据平台在Databricks这上面，就会有organic growth，这就是Databricks的目标。非AI方面没有新的产品。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 支持 专家客户样本中 Snowflake 消费同比增速逐季加速：Q4 FY26 22-23%、Q1 26%、Q2 29-30%
回 答 者：去年Q4约为22%至23%，今年Q1约为26%，今年Q2约为29%至30%。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 按项目储备，全年消费增速约 32-34%，下半年快于上半年
回 答 者：随着智能体应用的普及，客户对数据架构进行现代化改造，并将多个数据仓库整合至数据湖，Snowflake的消费需求持续增长。此外，Snowflake围绕Cortex AI、Cortex Code、Snowflake Intelligence及Snowflake Co-work进行了一系列功能增强，这些产品在Q2的采用率有所提升。根据目前的项目储备和发展趋势，预计今年整体增速约为32%至34%，下半年增速将较上半年进一步加快。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 新业务中 40-45% 与 Snowflake 的 AI/智能体相关
There are other vendors also out there. Um, to answer the last part of the question, I would say, you know, at least, um, I would say, you know, most of the conversations that we have with customers these days are agentic related, right? So, 40 to 45% of the new, new business that we are getting is all AI and agentic related with snowflake. Uh, Does that make sense? Does that make sense? Yeah, yeah, yeah. Yeah, makes sense. 对，然后想问一下就是有多少的就是专家的客户里面有多少把AI的话从POC阶段的话开始逐步的走向某些的场景的应用阶段，然后有哪些，然后有多少比例的客户的话，他们把单个场景的一个AI的应用的话，开始扩展到全公司的一个应用了，对，然后这里面的应用，这里面的不同阶段的公司数量的话，它的一个，它是否有有有比较明显的一个增长，对。 Hm, among your customers, how many are moving from POC to the scenario-based application? And also, how many has um changed from a single scenario application to the like whole enterprise applications? And for the company numbers in the different stages, do you see a significant increase? an obvious increase. Okay, multiple questions there.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 加速的第一层来源是遗留数仓 (Informatica/Oracle/Exadata) 迁移和数据整合，这是 Snowflake 自有的核心 credit
回 答 者：确实如此。Snowflake正在多个领域持续扩展。首先，使用Informatica、Oracle、Exadata等传统数据仓库的客户，正在向数据仓库、数据湖等环境迁移，我们已经看到这类数据现代化趋势。数据完成现代化后，客户会进一步集成智能体解决方案，包括自定义智能体或其构建的第三方智能体。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 口径不齐 消耗结构：底层数据仓库 55-60%、Snowpark 15-20%，两者属于工作负载自然扩展
回 答 者：核心的Snowflake数据仓库仍占据显著份额，底层数据仓库约占总消耗的55%至60%。Snowpark的消耗正在增加，越来越多客户开始采用Snowpark，目前约占15%至20%。数据仓库和Snowpark的增长主要来自工作负载的自然扩展，以及更多应用的接入。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 渠道商 H1 同比 40%+ 超目标，80%+ 利用率，新增 50+ 高级人员，需求超出产能
BJ: For sure. Our H1 did really well. We were above target, approximately over 40% year after year. Again, we're rebooting, resetting the data in our practice, so the percentages look good. But overall, I am absolutely sold out, which is a meaningful number to highlight here because we run 80% plus utilization and have been from Q1 to Q2. I've onboarded another 50 plus senior hires, not juniors, seniors. I guess the original plan where I had maybe 30 in plan and I've gone to over 50 or so.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 支持 增长来自为 agent 工作负载重建数据基础，是多年工程而非季度性
That is a very strong signal from where I'm sitting and more so than the growth rate that obviously looks really well as well. And I would like to also flag that majority of my growth wasn't because I was underpriced or underscoped. If anything, we were on more of the premium side from a pricing standpoint of view and scoping was more deliberate bottoms up as the market moves away from T&M and that's where the driver of a lot of what we were focused on was actually not just old school re- platforming where clients were just doing data migrations and modernization; it's actually more of rebuilding the data foundation to fund the agentic workloads, and that's a multi-year revision and an execution accordingly, not a quarter by quarter thing, which is great for me but it also makes some of the current workforce planning a little bit more challenging as well because you end up leaning more towards more senior MDE type profiles versus more of where the consulting margin was built back in the days. What I would say, my signal though, is for the rest of the year, it is to pay very close attention to agent governance work, especially because that's the leading indicator of everything else underneath it from a data foundation to modeling to applications. It is fundamentally the work stream clients add when they decide to go to production rather than another pilot.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 传统 lift-and-shift 迁移需求减少，对数据工程业务转为保守
BJ: I think we're now starting and leaning with AI first and then fundamentally as we map the agent workflow, we start looking at the data ecosystem and data state and the readiness of it and that's where we end up actually deploying our data engineering talent and modernizing components as needed. I am not seeing that clean-cut demand curve that I was benefiting from last year, or even in early Q1 fundamentally, where a lot of it was lift and shift that I'm not seeing much of. So if anything, I'm bullish on the AI side; I'm not, I'm more conservative, rather, on the data engineering, and if anything, both of them at the end of the day work out, but it does have an impact on how you do workforce planning and the profile and the person persona that you ultimately have. Some of the personas that we're going after can mix in from data engineering to engineering with a bit of software engineering on top of it. Those are harder profiles to train up and maintain and then also acquire easily, right? At the same time, more commodity... more standardized skill set in data engineering is actually on a downward trajectory, not just in the industry itself. There are a lot more of that available, and at the same time, the nature of what they were solving for is no longer in demand. So you know, it's a different flavor of growth.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 口径不齐 按狭义 (智能体自动化) 口径，AI 负载只占该客户 Snowflake 消耗 15-20%；Snowflake 自己的定义下 AI 占比会高得多——这正是'AI 贡献一半加速'口径要拆的地方
Good. Um, and again, you know, they they contract directly with uh with um uh, you know, with snowflake, so I I don't know exactly. Uh, but I would say that AI related workloads are around maybe you know, 15 to 20%. Um, uh and that that depends how you can there I'm using a sort of a narrow definition. So, you know, workloads related to agentic automation. I think uh Snowflake has their own definition, which is, you know, points to AI related consumption being a lot higher. Uh but but you know, in in pilot production, I would say it was 15 to 20%. And then as they scale, right? Um, then I'm I'm just doing some some some notes here as well. Uh they then as they scale, you know, that the 15 to 20% part is the part that will increase massively. Uh, you know, 5x to even 10x when they are fully fully scaled it. So they're not quite there yet, but you know, once they have it.
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 支持 其余 80-85% 是仓库等非直接 AI 负载，靠负载迁入和 IT 供应商整合每年增长 20-30%——即核心 credit 增长主体仍是迁移整合
其余80%至85%主要是数据仓库等工作负载，与直接的AI用例相对独立。这部分也会增长，因为该公司仍在将更多工作负载迁移至Snowflake，这是长期IT供应商整合的一部分，但增幅不会达到5倍至10倍，预计每年增长约30%，即每年20%至30%。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 支持 专家接触的 Snowflake 客户中 75-80% 在增加容量，原因有 AI 也有常规数据准备/整合
Uh I want to understand like the proportion of companies that have uh tried to increase the capacity or promised to increase the usage. Yeah, exactly. I would say about it's it's almost, you know, 75, 80% of the of the snowflake customers I have are increasing their capacity. Because again, it's like they either do it for AI adoption and or for uh general data preparation and and data consolidation. So, uh, so it's a vast majority. Say say 75, 80% of our customers.
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 口径不齐 2-3 倍只是 AI 负载，账户整体消耗只增加约 40%
A:所述特性，包括数据工程、特征准备和管道，都是与 AI 工作负载相关的。该增长并不适用于 Snowflake 的整个资产。在整体账户层面，我认为增幅更为温和，约为 40% 的增量消耗，因为现有的分析和数据工作负载本来就已经很可观。那两到三倍的增长专指支持或驱动 AI 的数据处理、RAG、embedding、检索和相关工作负载。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 支持 新客户只占收入约 5%，约 95% 来自存量客户；增长主力是存量大客户扩张
Snowflake 的主要增长动力来自现有大客户的扩展，而不是新客户获取。对 MongoDB 来说，增长来自扩展和多产品采用的组合，属于典型的先落地后扩展（land-and-expand）模式。AI 为 Snowflake 带来更多的计算、数据工程和 Snowflake Cortex 工作负载。对 MongoDB 则带来更多的运营型 AI 应用、向量检索工作负载和上下文数据。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 中性 大客户消费增速约 30%，长尾约 25%
长尾客户可能占客户总数的约 95% 到 96%，MongoDB 的情况也类似。他们在消费或收入中的份额对 Snowflake 约为 30%，对 MongoDB 约为 40%。大客户的消费增长很可能高于长尾：Snowflake 约为 30%，MongoDB 约为 25%。长尾的消费增长约为 Snowflake 25%，MongoDB 20%。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 中性 Snowflake Tier 1 客户不足 10% 却贡献 60-70% 收入，少数大客户投产即可拉动 3-5 个点；但价格变化使金额不一定数倍增长
对于拥有存量客户的软件公司而言，只要有部分大客户进入生产环境，消耗量基本上就会实现数倍增长。由于价格存在变化，相应金额不一定也会实现数倍增长。即便如此，只要有几个大客户进入生产环境，就可能直接带动这些公司在未来几个季度的收入提升3到5个点，业绩也很容易超出预期。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 中性 Tier 1.5 以下客户的新签与扩容案例不多——增长靠存量大客户
许多参与渠道调研的IT咨询公司观察到，Tier 1.5以下客户，尤其是在新签约和客户扩容方面，新增案例并不多。我们本次也详细询问了是否存在较大的案例，并进行了具体了解，这些情况与财报数据能够相互印证。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 同案例传统数据存取消费同步 +20-30%——AI 拉动核心数仓 (第 1 层)
回 答 者：客户在数据存取方面的各类consumption也同步增长了20%到30%。AI会加大对数据库的查询和调用量。目前，逐步进入生产阶段的客户比例为30%到40%，高于25年年底。整个整改时间可能接近半年到一年，因此预计今年下半年至明年初会出现较大变化。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 中性 POC 到生产需半年到 8-9 个月，平稳加速而非爆发
同时，企业AI从POC走向生产阶段不会出现爆发式增长。大部分客户在部署阶段都需要半年，甚至8、9个月。因此，我们更倾向于认为这是平稳加速的过程，并将持续拉动消费。同时，去年Snowflake也发布了一些新产品。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 Partner C 2Q26 consumption +32% (预期 27-28%)，3Q/4Q 预期从 30% 上调到 34-35%
1. 3Q25 grew 24% YoY, exceeding the 22% expectation 2. 4Q25 grew 28% YoY, at the high end of the expected 26-28% range 3. 1Q26 grew 28-29% YoY, exceeding the expected 24-26% 4. 2Q26 grew 32% YoY, exceeding the expected 27-28%, making it the strongest quarter over the past several quarters 5. 3Q26 and 4Q26 YoY growth expectations have been raised from 30% to 34% to 35%.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 在净新增 consumption 中，CoCo 约占 15%，Snowpark 约 10-12%
1. Snowpark and Cortex Code contribute the most. Cortex Code has a lower base but is growing the fastest. 2. Next are Snowflake CoWork, Cortex AI, Openflow, PostgreSQL, and other newer products such as Observe.AI. 3. Based on contributions to net-new Consumption, Cortex Code accounts for approximately 15%, while Snowpark accounts for approximately 10-12%.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 Partner D (实施收入口径) 认为 2Q 加速主因仍是核心数仓及其迁移；零售/制造迁移活动回升
1. The primary driver remains the core data warehouse business, including data warehouse migrations and the resulting customization and integration work. Within my customer base, retail and consumer products customers had more migration activity in 2Q than in the prior 2-3 quarters. Manufacturing migration and implementation activity also recovered, alongside ERP, supply chain projects, and upgrades to existing ERP systems.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 Partner D 估计 2H26 增长中约 60% 来自数仓/数据湖现代化/数据工程，约 25% 来自 Snowpark/Unistore/动态表/Iceberg，约 15% 来自 CoCo 等 Cortex 产品
1. Data warehousing remains the largest contributor and will account for at least half of 2H26 growth. Data warehousing, data-lake modernization, data engineering, and ETL pipelines could collectively account for 50% to 60%. For simplicity, call it roughly 60%. 2. After data warehousing, Snowpark, Unistore, Dynamic Tables, and Iceberg Tables could contribute approximately 25%. Under my classification, this would bring the cumulative share to approximately 85%. 3. The remaining 15% will ultimately come from Cortex Code, with Cortex Search, Cortex Analyst, and other Cortex products also contributing. Cortex Code should have greater visibility by year-end.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 迁移恢复部分是关税/地缘不确定性推迟后的补做，存在一次性成分
1. The core data warehouse contribution has increased versus last year and may remain elevated through year-end. Earlier tariff and geopolitical uncertainty caused some enterprises to postpone planned migrations, including some ERP modernization projects. 2. These projects began recovering around the end of 1Q this year as enterprises recognized the need to keep their systems current. The same dynamic applies to data warehouses. Customers that have been using Snowflake for 3-4 years are also actively evaluating and expanding functionality and capabilities. 3. How Agentic Workloads Benefit the Data Warehouse
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 CoCo 占增量 consumption 从 1Q26 的 5-7% 升到 2Q26 的 12-15%，占总消费接近两位数
1. Cortex Code contributed approximately 12-15% of incremental Consumption in 2Q26. 2. On the same basis, it was approximately 5-7% in 1Q26. 3. Cortex Code now represents a nearly double-digit share of total Consumption.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 CoCo 最大用例是数据迁移和转换
1. CoCo understands the customer’s full Snowflake technology stack and environment, while tools such as Claude Code require more context to work effectively. Think of it as a Snowflake-native coding agent rather than a general-purpose coding agent. Most Snowflake customers are consolidating storage and compute within the ecosystem, making this context valuable, although many, perhaps most, engineering teams will also use Claude Code for development outside Snowflake. 2. I have not yet seen customers use CoCo for general-purpose coding. In projects heavily dependent on Snowflake, teams may use CoCo for a small amount of development outside Snowflake because it already has the relevant context and avoids switching between agents. But this depends on the specific use case. So far, I have not seen an enterprise choose CoCo as its primary coding agent across the organization. Among the enterprise customers I have spoken with, the vast majority, if not all, are using two or three coding agents simultaneously. The market is still somewhere between experimentation and consolidation. 3. CoCo’s unique selling point outside the Snowflake ecosystem would weaken, but even if it does not become the enterprise default coding agent, it can still be a strong growth driver. Enterprise customers continue moving more compute and storage into Snowflake, so even if the coding agent operates only within this expanding environment, it can generate substantial consumption. 4. Data migration and transformation are the largest use cases, while data pipeline development and event-driven workflows are also important.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 专家没有看到核心平台增长中 agentic 工作负载占比的数据，也难以计算
1. The impact is not significant yet, but the trend is real. Enterprise IT environments are highly complex, and architecture teams and procurement departments have been working together to reduce the number of systems they use. Overlapping capabilities are driving platform consolidation, and Snowflake is one of the platforms procurement and architecture teams consider as part of that consolidation. 2. MDB could face some pressure. MongoDB is one of the vendors operating independently of large multiproduct ecosystems, and procurement feedback identifies competition from these ecosystems as one of several competitive factors it faces. This remains a long-standing debate within procurement teams: choosing the best-of-breed point solution versus the best suite. 3. They support core platform growth both indirectly and directly. Indirectly, enterprises recognize that consolidating more workloads and compute on Snowflake makes agentic and AI work easier, encouraging further migration and platform consolidation. Directly, agentic workloads consume compute resources on the core platform and bring more structured data into Snowflake. I have not seen data on what percentage of core platform growth comes from agentic workloads, and that share would be difficult to calculate. The core platform is growing more slowly than some other modules, but given its scale, growth remains strong.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 Funda 承认 2Q25 有一次性迁移收益，2Q26 核心数仓继续加速有难度
1. Our base case is 35% growth, a 4% beat versus guidance, modestly above a typical beat, and 1 percentage point faster than 1Q26. 2. Given a one-off migration benefit in 2Q25, 2Q26 faces a difficult comparison. We believe continued acceleration in the core data warehouse business is challenging in 2Q26, but incremental contributions from Cortex Code, Snowpark, and other products can still drive acceleration in total revenue growth. 3. 3Q26 will no longer face a similar high-base headwind, so we expect further acceleration to 36%.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 Q2 产品收入 $1.49B，增速加速到 37%，两个季度累计加速 7pt，连续第二个季度创纪录环比美元增量
By putting intelligence work at scale, our customers are building faster, executing more efficiently, and reimagining their businesses in ways that weren't possible before. Put simply, the agentic enterprise runs on Snowflake, and the traction is translating into strong business performance, as evidenced by our Q2 results. Product revenue came in at $1.49 billion, with growth accelerating to 37% year-over-year, marking our second consecutive quarter of record sequential dollar growth. After exiting Q4 of last fiscal year at 30% year-over-year growth, we have now added 7 points of acceleration in just two quarters. And with our continued focus on executing with discipline and operational rigor, our Q2 non-GAAP operating margin expanded by more than 400 basis points year-over-year to 15%. Thank you to all of our Snowflakes for the hard work and dedication that made this performance possible.
- 口径不齐 '约一半加速'来自 AI 产品，但 AI 产品口径很宽 (CoCo、CoWork、AI functions、AI Gateway)；另一半包括 Notebooks、Streamlit/React 应用和迁移提速——不是严格的'核心数仓'
But I was wondering if there's any way for us to sort of distill down what these newer products are having maybe on their own account. Thanks. I would roughly call it even, our AI products which is a pretty broad swath at this point. Absolutely, it's CoCo and CoWork, but it's also things like AI functions that make data operations proceed at an impressive scale, or even newer products like the AI Gateway. They contributed approximately half of the acceleration that we are seeing. But there are a lot of other products that are also demonstrating robust growth. And Brian touched on some of them.
- 支持 另一半加速包括迁移本身提速；某网络设备厂商 Teradata 迁移不到 3 个季度完成，以前要 2–3 年
And some of the recent advances, both in models and harnesses, are letting us run long-duration tasks of a scale and complexity that we haven't been able to do before. And the rate at which workloads are coming onto Snowflake is also an important factor. And one anecdotal example that a big network equipment manufacturer is doing a Teradata migration in less than three quarters this year, and this is something that would have taken probably two to three years in any previous time. So, these are some of the things that are contributing to our acceleration and beat. Thanks so much, Sridhar.
- 支持 加速来自广泛客户，不集中；AI-native 公司仍只占收入一小部分
This is Sridhar. Let me take a first cut at this, other folks can add on, since it's a pretty broad question. First, I think we see the acceleration come from a very broad swath of customers. It is not concentrated. For example, with, let's say, AI-native companies, they continue to be a small part of our overall revenue stream. And I think the thing that's also materially different this time around with folks that are investing is that the products like CoCo make optimization far, far easier than before. You can point CoCo at a query that's taking too long to run, or you can basically have it debug the top 10 longest running queries are the most idle warehouses, things like that are a lot easier to do.
- 中性 NRR 持平 126%；管理层称扩张由迁移和 AI use case 共同支撑；GRR 近几季'相对平稳'
We added 14 net new Global 2000 customers, bringing our total to 829. Our AI Data Cloud now supports over 41% of the Global 2000. Within our existing base, customer expansion is healthy, as evidenced by our net revenue retention rate of 126%. This expansion is underpinned by growth in both migrations and AI use cases. In Q2, 48 net new customers surpassed $1 million in trailing 12-month spend. We now have 828 customers spending above the $1 million threshold.
- 口径不齐 新客户在购买后约 3 个月消费到 80% 合同容量的比例在最新 cohort 中明显改善——支持需求强，也可能意味着容量提前耗尽、续约前置
And in a funny kind of way, CoCo also makes it far easier to create agents and get value from the data itself. And this is the combination that makes Snowflake so attractive and it's not just acquiring customers. We track this metric called like time to 80% of purchased consumption for new logos that we acquire and we measure it cohort by cohort. Basically, of the customers that you acquired, let's say, in January, what fraction of them are consuming more than 80% of their purchased capacity, call it three months after their purchase month. And this metric has very, very visibly improved for the newest cohorts of customers that we are acquiring. That's the power of AI. It's faster to get projects done.
- 反对 部署的 use case 同比增速从 Q1 的 114% 放缓到 89%，每个 AE 赢得的 use case 增速从 86% 降到 43%
Our teams are using these products every day, learning firsthand what they can do and taking those insights directly to our customers. We're seeing the results in how quickly customers are putting Snowflake to work. The number of use cases, individual customer projects deployed on Snowflake increased 89% year-over-year as customers moved more workloads into production. At the same time, use cases won per account executive increased 43% year-over-year, demonstrating both growing customer demand and strong sales productivity. And we're pairing this investment in growth with continued operational discipline. We remain on track for GAAP profitability in Q4 fiscal 2028, and the operating leverage we build along the way strengthens the durability of that outcome.
- 支持 核心需求仍强，AI 收入显著提升 (TMTB 定性，未量化)
Product revenue grew 37%, ahead of both Street and buyside expectations, while the Q3 guide calls for further acceleration and FY27 growth was raised five points to 36%. With this q3 guide, you now very likely get 40%+ growth in q3 a quarter earlier than even bulls expected. This is an unusually clean print: core demand remains strong, AI revenue stepped up meaningfully, and management passed substantially more than the quarterly upside into the annual outlook. Bulls in full control here. This explaisn the CoCO + core magnitude of the beat well (from our [weekly](https://substack.com/redirect/1aa17942-4a71-4deb-beb9-3d8d65ad436f?j=eyJ1IjoiNDR3MndrIn0.NBlbpuapVvjXEQ-22Fcu951Q9Y3oMbVG8TDoGPfg2x8) on Sunday):
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 - 口径不齐 管理层称 AI 产品组合贡献约一半加速，另一半来自核心平台、迁移、Notebooks、Streamlit/React 应用；未给计算方法
* Management also said the AI product portfolio contributed about half of the quarter’s growth acceleration, with the other half coming from the core platform, migrations, Notebooks, and Streamlit and React applications.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 口径不齐 新客户用完 80% 购买容量的速度显著变快，公司未披露具体数字 (提前消费风险)
* The flywheel is not only about new customer acquisition. It also helps newer customers consume purchased capacity faster. The company did not disclose a specific uplift figure, but said the speed at which new customers reach 80% capacity consumption has improved materially.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 Funda 估计核心数仓在高基数下仍加速 1%，剔除 2-3% 基数影响，底层加速 3-4%
We believe the biggest highlight this quarter is that Core Data Warehouse still accelerated by 1% despite tough comps. Adjusting for the 2% to 3% headwind from tough comps, underlying growth accelerated by 3% to 4%. This should drive a more meaningful acceleration in FY27Q3.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 大型制造客户计划不到三个季度完成 Teradata 迁移，2024 年不可想象
I think AI is having a pretty profound impact on how quickly you can get those done and expectations not just from me, for my team. I've talked before about how I want migrations to be mostly automated, but even customers are expecting it. We have a very large manufacturing client, for example, do a Teradata migration. Planning aggressively to finish it in less than three quarters is not something you would have heard of in 2024. That's part one of a lock mode is possible. Let's go get it done.
- 支持 客户爬坡到消费 run rate 的曲线越来越陡 (未量化)
And so whether it's our partner network or what we're doing internally, what Sridhar talked about is outcome-based pricing, has really built the credibility with our customers. If someone comes to me and says, I can guarantee you X for this set price, and I knew historically that took a lot of time and a lot of materials and so forth, like I'm all in, and we're seeing that from a customer perspective. We do track internally how long it takes for them to get up to their consumption run rate. And we're seeing those curves gets steeper and steeper and steeper. And so, customers are deploying quicker. They're consuming quicker. They're using – partners, us and themselves are using our agents to actually do that. And so, it's really exciting to see.
- 中性 分析师担心 CoCo 是试验性使用、可能褪色；称公司此前说过毛利留存稳定
So, look, it's a new product cycle, and you're going to have customers that are experimenting. You're going to have questions around gross retention durability. I think you've already said gross retention has held stable even as CoCo is scaled. My question for you as an analyst, we try to model CoCo, and we also try to model the impact that migrations and the speed of migrations is having on your business. What advice would you give us as we try to think about some of the blue-sky scenarios over the next 18 months? And how do you de-risk the forecast from customers getting really excited about CoCo, but ultimately it's still very competitive and you've got experimentation? And could that usage pattern actually fade over time? Absolutely. It's something that we struggle with internally as well when you launch a new product.
- 口径不齐 同一案例给了两个口径：该用例使消费增 25-30%，全部功能合计增近 40%——与 4 万→7 万 (+75%) 不一致
该客户最近将该用例投入生产，现在每天处理约10,000份申请。该用例使消费增加了约25%至30%。包括 OCR、Document AI、Cortex LLM、生成式 AI、SQL 查询和 ML 模型推理在内的整个 Snowflake 功能和操作使消费几乎增加了40%。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 水泥/制造客户上线 2-3 个用例后消费增长 25-36%，同比约 35%，若扩展可能翻倍
A:对于该特定客户，消费增长约为 25% 至 36%，目前仅实施了两到三个用例。已识别出更多用例。如果最初的两到三个用例成功，客户预计会扩展到更多用例，消费可能会翻倍。目前的同比增长约为 35%。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 增长主要来自存量扩张 (NRR 126%)，新客户需 5-6 个月才稳定
在完成入驻后，客户通常需要五到六个月才能让 Snowflake 平台稳定下来，因为必须完成设置、开发和入职工作。大多数增长来自已经建立了 Snowflake 平台和数据基础的现有客户。Snowflake 最近的年报显示净收入保留率为126%，意味着现有客户在增加采购、扩展他们的 Snowflake 部署并将更多用例引入平台。这是收入增长的主要来源。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 中性 迁移完成后消费会平台化，裁员或数据量不增长会导致持平甚至下降
PM: Yeah, what plateaus is: Once the migration is done, the other part I didn't forget to mention is some cases like if it's warehousing, then the number of users also drives the growth and the consumption, obviously. So it's the data size and the number of users. So what plateaus is if the number of use cases within that use case plateaus, not more people are going to access the environment for whatever reason. Reduction of workforce or other reasons... that would be 1 reason that it goes flat or maybe goes down, the data size doesn't grow, or there is no additional investment in that particular use case. Those would be examples of why it would plateau for that particular case. And then the way you overcome it from a go-to-market perspective is we have industry experts; they look at that use case, they see the value.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 口径不齐 Snowflake 可能把 AI 应用中产生的数据用量单独归为 AI 收入，Databricks 则全部计入数据库用量，两家 AI 收入口径不同
Snowflake的Cortex产品刚刚推出两个月，其对收入的贡献比例需要考虑其发布时间。关于其收入贡献的具体数值，需要明确统计周期是月度、季度还是年度。Snowflake与Databricks在业务统计口径上存在差异。Databricks将所有数据访问，包括AI Agent或访问上下文产生的数据查询，都统一归入数据库用量。而Snowflake则可能将业务分为两套体系：传统的Data Cloud负责数据业务，而参与AI的部分则通过Hybrid Table和Postgres Page等另一套产品体系来支持。因此，Snowflake可能将AI应用中产生的数据用量单独归类为AI业务收入，而不是计入传统的数据查询业务。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 反对 Snowflake AI 场景与数据场景割裂，AI 新功能对存量和新客户影响同质，存量客户使用 AI 的起点与新客户无本质区别 (回应 NRR 持平问题)
Snowflake的AI场景与数据场景之间存在较为明显的割裂。客户即便在Snowflake平台上使用AI功能，其对已有数据的依赖程度也不及Databricks。因此，Snowflake推出的AI新功能对于新客户和存量客户的影响基本是同质的。存量客户的优势主要在于他们已有的合作关系，能够更快地了解到新产品，市场认知度更高。但在搭建AI应用方面，即使是使用了大量数据仓库（Data Warehouse）的存量客户，当他们开始采用AI时，其起点与新客户并无本质区别。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 口径不齐 同一案例的使用量增幅先后说 25-30% 和近 40%，与 $40K→$70K (+75%) 不一致
近期，该客户已进入生产阶段。该用例使 Snowflake 的使用量增加了约25%至30%，因为该客户每天处理约10,000份申请。OCR、Document AI、Cortex LLM、生成式 AI、SQL 查询、机器学习模型推理等整套 Snowflake 功能和操作，使整体使用量增加了近40%。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 增长主要来自现有客户扩展而非净新增；新客户需 5-6 个月才稳定，NRR 126%
回 答 者：建模需要先构建基础，而这需要一定时间。因此，我们的增长主要来自现有客户的业务扩展，而非净新增客户。净新增客户确实是销售管线的驱动因素，但增长主要来自现有客户的业务扩展。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 中性 客户使用 Snowflake 后需 5-6 个月才进入稳定阶段——新客户贡献滞后
即使在上个季度，也有48家新客户的支出突破1百万美元，这是衡量客户增长的一项指标。通常，客户使用Snowflake后，需要5至6个月才能在平台上进入稳定阶段，期间涉及设置、开发和上线等工作。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 加速来源：更多 persona、迁移更快、更多数据进来
Now anybody within the business can get access and we don't charge like CoCo is not -- we don't charge for it. You just get it being a Snowflake customer, and then the charge is the consumption that you actually draw in using the product. And so, we see more personas, and then we're seeing migrations go faster. And so, we're seeing more data come in, relative migrations. So, there's a whole bunch of different things that AI is doing that's driving strength and acceleration in the quarter. Yeah, got it.
- 口径不齐 被问是否会重演 5 年前的优化潮，CFO 只说已过试验阶段，没给数据
I got my colleague Katherine here. And so, anything I get in trouble with, she'll help me out. But from the -- when I look at the data and sort of understand the past and sort of what's going forward, we're well past experimental phase with CoCo and AI, and really seeing the power of what this does within our customer base gives me the confidence in forecasting it with the guidance. When you have sort of a new tool and a lot of people adopt it and use it, that's great. But what you really have to ensure is that you're driving positive business outcomes and there's ROI on that.
- 支持 采用 CoCo 或 AI 的客户，核心业务比原本多 11% (转述 CEO LinkedIn)
But it's really a – the 5% comes down to the core business is doing well and is accelerating. We're also seeing benefit in the core because of AI. And so Sridhar did a post on LinkedIn a little while ago, and I think it was something to the effect that when someone adopts CoCo or AI, we're seeing 11% uplift to their core business that they would do otherwise. And so both of those are sort of contributing to the beat. And maybe you can drill down on that a little bit more because CoCo is creating a great unlock just in terms of migrations, moving data over.
- 口径不齐 Q2 超预期一半来自 AI 产品、一半来自核心 (是 beat 拆分，不是增速加速拆分)
And so that's what's super exciting for us at Snowflake. And that's one of the reasons that was what helped our print. We talked about 50% of the beat was related to our AI products and 50% was related to the core. And there are these secular tailwinds sort of in industry now sort of pushing towards getting your data in a database such as Snowflake so you can actually get value out of it. When I talked about – and don't mean to hop on this too much, but when I go talk to CFOs and CEOs, and then show them what we can do, they're like, I want that now.
- 支持 5% 超预期来自核心业务加速，且 AI 也在拉动核心
And what we're trying to focus on now is get our sales team to go in and talk about the use cases to drive deeper penetration in those accounts so we can get more users. And so that's what we're in the process of doing. But it's really a – the 5% comes down to the core business is doing well and is accelerating. We're also seeing benefit in the core because of AI. And so Sridhar did a post on LinkedIn a little while ago, and I think it was something to the effect that when someone adopts CoCo or AI, we're seeing 11% uplift to their core business that they would do otherwise. And so both of those are sort of contributing to the beat.
- 支持 迁移时间从 10-11 个月降到约 6 个月，受客户因素限制无法降到零
And so, for instance, in our group that does our migrations, they use it to actually help migrations. And so they'll – we've got a whole bunch of skills and agents that can actually take the migration time down. We're around 10 or 11 months. Now we're around six months. There's some customer dependency there. So, we can't get that down to zero, although we'd like it. But we're currently working as much as possible to bring down the migration timeframe.
- 中性 核心业务主要来自六到八年前的老客户，高留存
Like, we're spending as much as we want to spend in the business. We go in to these large enterprises, and our gross retention rate is really high. The core of our business really comes from customer six, seven, eight years ago because they got on Snowflake, they move more and more into Snowflake. Now with stuff like CoCo, more people within the companies are actually using it. And we're seeing – so you can imagine as a customer builds, we're such – we have a great net retention rate, one of the best in the software industry.
- 支持 Warehouse 计算一直是 Snowflake 最主要收入来源，存储和 Cloud Service 占比小
Service。从技术角度来说，这其实就是Warehouse。Snowflake的Warehouse和传统的Warehouse有些不同。对于数据仓库或者数据库，大家通常认为Warehouse就是一个仓库，把数据或文件存进去就可以了。但在Snowflake中，Warehouse其实是一个Computing Node，是资源节点。当用户需要用SQL查询数据时，Snowflake就是通过Warehouse来完成查询的，所以Warehouse本质上是Computing Node。从Snowflake成立到现在，Warehouse的查询一直是主要收入来源，也是最核心的收入部分。客户在付费时，Warehouse这部分占了大头。具体比例我没有查到，但可以确定，Warehouse是收入的主要来源。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。
E02 AI 让核心 credit 变多了吗
判断：方向成立：AI 确实让一些账户的 credit 涨了很多，具体案例都有。但净效果没有量化，管理层拒绝给；而且 AI 同时在帮客户省钱、在让外部 agent 绕开 Snowflake。
成立的案例：一家 BFSI 客户把 Cortex Agents/CoWork 开放给 200 多个分行人员后，每月 Snowflake 消费从 $4 万涨到 $7 万；一个 AI 用例从 POC 到生产，月消耗从 $5,000 到 $3 万；北欧保险客户从试点到多业务线生产，计算消耗涨了 5–8 倍，主要来自理赔人员的查询量；一家零售客户的 AI 负载涨 2–3 倍，但整个账户只涨了约 40%。管理层说使用 CoCo 的账户在核心上也消费更多 (飞轮)，Snowflake 自己取消了所有 BI 和分析软件合同，用 CoCo + Streamlit 替代。
抵消的力量：(1) CoCo 帮客户省钱 (见 s4 第 4 条)；(2) 一个 Spark 遗留客户用 CoCo 把代码转成 Snowpark，快 5 倍、成本更低：迁入变容易了，但单个负载的 credit 更少；(3) 一位专家说 AI 带来的消耗增长有很大一部分来自架构低效 (agent 和向量库之间没有缓存、编排差)，"5–10 倍的资源消耗增长不可能一直被接受"；(4) Snowflake 的工程师承认，大多数企业用外部编码 agent 经 VPC 连 Snowflake，"只把 Snowflake 当存储和数据管道，智能在外部"；(5) 一位专家说客户直接让 AI 检索 SharePoint、ServiceNow、Confluence，不经过 Snowflake/Databricks 这一层。
结论：AI 对核心 credit 的拉动今天是正的，而且可能是加速里最好的一部分。但它的大小没人知道：管理层唯一的数字 (采用者核心消费多 11%) 来自 CEO 的一条 LinkedIn 帖子，而且没说对比的基准。考虑到 CoCo 同时在优化客户的账单，以及 2023 年"客户优化"曾让增速一年腰斩，我把第 1 层的持续性打一个折扣。
回 答 者：在 Cortex Agents、Coco 及这些 Co-work 功能推出之前，客户主要进行传统机器学习建模，例如梯度提升预测、回归分析或 ARIMA 预测等。使用传统机器学习模型时，每月消耗约为 40,000 美元，主要由中央 IT 团队驱动，其他分支机构用户无法使用 Snowflake。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。
最初，概念验证阶段的消耗约为每月 5,000 美元。随着项目逐步进入生产阶段，按照客户希望处理的交易数量和应用数量估算，目前消耗已增至 30,000 美元。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。
A:所述特性，包括数据工程、特征准备和管道，都是与 AI 工作负载相关的。该增长并不适用于 Snowflake 的整个资产。在整体账户层面，我认为增幅更为温和，约为 40% 的增量消耗，因为现有的分析和数据工作负载本来就已经很可观。那两到三倍的增长专指支持或驱动 AI 的数据处理、RAG、embedding、检索和相关工作负载。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。
So, we've seen couple quarters of this behavior. Our sales team is doing a great job with proving the business value of the use cases. And we're continuing to see great new logo additions. CoCo, when we look at CoCo, the accounts that are using CoCo are consuming more of the core as well. And so, there's this flywheel effect that we talk about. We had 9,100 CoCo accounts this quarter. That's up significantly from last quarter.
So, we can't get that down to zero, although we'd like it. But we're currently working as much as possible to bring down the migration timeframe. For myself, I use it for, call it, like, analytics and business intelligence. And so, we've actually canceled all of our BI and analytics contracts. And so, I use CoCo to Streamlit and can spin up dashboards in a heartbeat with whatever I want around the business, whatever time period. They aren't static. They're interactive. Because of the MCP stuff, I can then turn that into an e-mail, send that straight out to CoCo.
I mentioned this morning there's a customer that wanted some legacy Spark API and it's paying to support that legacy Spark API. We've been working on it and we know that at the end the day it's going to be compatible but not super fast. And we're busy doing that. And in the meantime, the customer said, oh, we tried CoCo. We converted to Snowpark and we're done, we're good and it's five times faster and cheaper by implication. So, we're actually – we have a renewed push on Snowpark. We mentioned it to some of you that we are seeing increased momentum of Spark migrations, use of Snowpark just because, hey, the pain that represent that converting code is no longer as painful as it is which is easier to do. And then the last piece that I'll mention in the first column here is the productization of an acquisition we made a company called Datometry.
回 答 者：消耗量确实在增加。值得注意的是，消耗增长很大程度上源于架构效率不足。例如，如果智能体与向量数据库之间没有缓存层，或者编排设计不合理，就会产生额外消耗。供应商通常会将消耗曲线作为AI价值的证明，Snowflake也会这样展示，但对此需要谨慎看待。部分消耗与业务影响成正比，但如果不重视架构设计，很多消耗实际上来自不必要的架构复杂性，通过优化管线本可避免。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。
不过，如果Snowflake或MongoDB的资源消耗分别增长约5到10倍，这种情况不可能一直被接受。未来企业会更加重视有纪律的资源使用和合理的底层架构，并通过针对性设计抑制资源消耗增长。类似情况也会出现在其他AI领域，例如模型公司的Token消耗等。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。
这里也展示了常见的失败原因。大部分企业目前可能都在使用这种模式：将外部编码agent或大模型通过VPC等狭窄管道连接到Snowflake，仅将Snowflake作为存储和数据管道，而所有智能仍位于外部。这种架构在生产环境中很容易崩溃，因为外部agent不了解企业内部情况，每次都会不断寻找错误的表，进而导致token失控。由于每次query都需要扫描全表数据，成本也会相应增加。但如果有了我们，情况会有所不同。这里可以看到，其实……
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。
A:绝对有。我认为这是一种重要的新兴架构。并非每个 AI 工作负载都需要 Snowflake 或 Databricks 做中间层。对于以知识为中心的用例，客户可以部署一个 AI 检索层，直接查询 SharePoint、OneDrive、ServiceNow、Confluence 或其他存储库，同时保留源端的权限和治理。例如，Microsoft 支持从 SharePoint 和 OneDrive 检索内容，而无需客户将这些内容复制并维护在单独的索引中。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。
Sanjit K. Singh (Morgan Stanley): congrats on reasserting 30% product revenue growth in Q4. Brian G. Robins: it's really based on the high stable growth that we see in our core business. Brian G. Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type. Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake. Brian G. Robins: we guide based on the Observe customer behavior up until really the point of earnings. Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
展开本节全部 133 条原文 (支持 75 / 反对 28 / 其他 30，按时间)
- 支持 管理层称 AI 是核心业务强劲的关键驱动，但未量化
Sridhar Ramaswamy: AI influencing 50% of the bookings signed this quarter. Sridhar Ramaswamy: 28% of all use cases deployed during the quarter incorporated AI. Sridhar Ramaswamy: AI is a key driver of the strength that we see in our core business. Sridhar Ramaswamy: today, we've announced an expanded partnership with Anthropic. Sridhar Ramaswamy: This brings native model availability into Snowflake, Sridhar Ramaswamy: enable our customers to move from legacy data warehouses to Snowflake Sridhar Ramaswamy: We recently acquired the technology behind Datometry's software migration solution Sridhar Ramaswamy: Key capabilities like Snowflake Openflow are making it easier than ever to bring in structured, Sridhar Ramaswamy: bringing our total GA product capabilities to 370 year-to-date, Sridhar Ramaswamy: over 40,000 customers, partners and prospects joined us across 23 events Sridhar Ramaswamy: we announced new partnerships with Workday, Splunk, Palantir, UiPath and more Sridhar Ramaswamy: to make the latest Gemini models available to our 12,600-plus customers Sridhar Ramaswamy: we introduced Cortex AI for financial services Sridhar Ramaswamy: We've also agreed to acquire Select Star to enhance our Horizon Catalog Sridhar Ramaswamy: Snowflake has already surpassed $2 billion in sales through AWS Marketplace in a single calendar year,
- 支持 AI 影响了 50% 的签约，28% 新用例含 AI
Sridhar Ramaswamy: we are seeing the fastest ramp in product adoption in our company history. Sridhar Ramaswamy: Already 1,200 customers are harnessing next-generation agentic AI capabilities to drive real business impact at scale. Sridhar Ramaswamy: AI influencing 50% of the bookings signed this quarter. Sridhar Ramaswamy: 28% of all use cases deployed during the quarter incorporated AI. Sridhar Ramaswamy: AI is a key driver of the strength that we see in our core business.
- 中性 客户无需为 AI 预先承诺——AI 消费直接从现有容量里扣
Ardy [ph] for Mark Murphy (JPMorgan): explicitly stated that Snowflake budget is now tied to their AI budget, and they're kind of broadening Sridhar Ramaswamy: The strongest pattern that we have had to work hard and earned this year is to be that genuine player Sridhar Ramaswamy: no amount of talking can make you that, you need project, products that produce the magic Matthew Martino (Goldman Sachs): what about the Snowflake platform is allowing customers to really accelerate their AI journeys? Sridhar Ramaswamy: with things like Snowflake Intelligence and Streamlit, which we are very heavy users of Sridhar Ramaswamy: investments like Openflow or even Postgres are going to expand the aperture Alex Zukin (Wolfe Research): the momentum that you're describing is showing up in bookings Sridhar Ramaswamy: The virtuous cycle of a Snowflake customer is one in which they sign a deal. Sridhar Ramaswamy: they don't have to make any pre-commits towards AI Sridhar Ramaswamy: bookings are an indicator of how much a customer thinks they're going to spend in the coming years. Sridhar Ramaswamy: we had four, nine-figure deals this quarter. Patrick Colville (Scotiabank): could you just remind us what does that $100 million actually include? Sridhar Ramaswamy: The $100 million is primarily the product suite, but it's the whole stack. Sridhar Ramaswamy: Cortex Search and Cortex Analyst, which are our unstructured and structured data products, respectively. Brian G. Robins: The $200 million is a buy side that we're buying from Anthropic.
- 反对 CoCo 不只写代码，也做性能调优、可靠性检查、识别长查询——会帮客户省 credit
Code可以自动从头构建整个管道并部署管道。它用于创建编排和复杂管道，支持更广泛的ETL/ELT项目。它也用于健康检查，例如Snowflake的可靠性检查、性能调优与优化、安全管理以及各种工作负载测试。它不仅仅是用于编写代码。除了代码创作之外，Cortex Code还有多种用例。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 CoCo 可提示式自动构建并部署摄取管道 (如从 PostgreSQL 到 Snowflake)
例如，如果您想从PostgreSQL将数据摄取到Snowflake，以前我们必须在编排工具中手动配置整个框架以创建部署和运行时执行。现在，通过Cortex Code，您只需提示它创建和修改数据摄取管道。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 受监管行业本可直连 OpenAI/Anthropic，但因治理和审计改走 Snowflake
通过在Snowflake中进行，管理员可以启用这些限制并确保正确的数据可被正确的人访问，从而保持对监管要求的合规性。这就是我们看到许多受监管行业采用Snowflake进行此类用例的一个原因。那些本可以直接使用OpenAI和Anthropic 的客户反而通过Snowflake使用它们，而不是直接使用它们。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 无论 OpenAI 还是 Anthropic 的 agent，通常都指向 Snowflake 数据，驱动更多负载和消费
潜在地，这会成为 Snowflake 的消费驱动因素。无论你使用 OpenAI 还是 Anthropic，这些代理通常会指向 Snowflake 的数据，这会驱动更多的工作负载和每个数据集的消费。从这个意义上说，代理式自动化的增长和对 Snowflake 数据集的 AI 查询增长是 Snowflake 的消费驱动因素。这是积极的。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 最坏情景：OpenAI/Anthropic 取代 Snowflake 智能层，把 Snowflake 降级为数据存储
另一方面，你可以认为如果 OpenAI 和 Anthropic 非常成功且激进，它们可能开始与或甚至取代 Snowflake 的智能层竞争。我并不是说我们看到这种迹象，但那是最坏的情景，届时 Snowflake 会被降级为数据存储，这不是他们想看到的情况。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 中性 当时 AI agent 对 Snowflake 是消费驱动但规模不大
让我们拭目以待。目前为止，我认为这是对 Snowflake 的消费驱动，而不是大规模的驱动，因为 AI 代理还处于早期阶段；它们还没有推动大规模消费，但那一天会到来。我们拭目以待。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 支持 ETR 调研：客户部署 Cortex Search 后，采用 Snowflake 其他组件的配对可能性上升约 350%
一个经常被低估的事实是：**模型可以切换，但数据搬不动**。一个企业的数据一旦沉淀在某个平台上——经过清洗、治理、建模、关联——迁移成本极其高昂。这种现象被称为"数据重力"（Data Gravity）。ETR 2025–2026年调研数据明确验证了这一点：当客户在Snowflake上部署了Cortex Search后，采用Snowflake其他组件的配对可能性飙升了约350%。数据重力不是理论，而是可以用采购行为量化的现实力量。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 支持 使用 AI 的账户单季增量最大，总数 9,100+
Sridhar Ramaswamy: Our net revenue retention was at a healthy 125%. Sridhar Ramaswamy: We added 2,332 net new customers this year, and we are seeing more and more businesses move over to Snowflake. Sridhar Ramaswamy: This quarter, we delivered the largest sequential increase in accounts using AI, Sridhar Ramaswamy: bringing the total to more than 9,100 accounts. Sridhar Ramaswamy: And in just three months, Snowflake Intelligence has scaled from a nascent offering to an essential capability Sridhar Ramaswamy: Snowflake Intelligence has scaled from a nascent offering to an essential capability for over 2,500 accounts, Sridhar Ramaswamy: almost doubling quarter-over-quarter. Sridhar Ramaswamy: It's a truly transformational coding agent that's already helping over 4,400 customers Sridhar Ramaswamy: helping over 4,400 customers build and scale AI-powered applications Sridhar Ramaswamy: This year we launched over 430 product capabilities, underscoring the strength of our product velocity.
- 支持 遥测显示平台活动因易用性提升而增加 (未量化)
Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake.
- 反对 AI 让数据工程'数倍'更高效——效率提升可能压低单任务 credit
Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G.
- 支持 管理层称 CoCo 让建 dbt 管道、数据工程、调试性能提速 10 倍以上
* “但Cortex Code才是真正改变游戏规则的产品，因为它是整个数据生命周期的强大加速器……构建DBT管道、运行数据工程、建立动态表，或调试性能问题，速度又提升了10倍以上。” * "本季度，我们实现了使用AI账户数量的最大环比增幅，总数超过9100个账户。仅仅三个月内，Snowflake Intelligence就从一个初期产品成长为超过2500个账户的核心能力，环比几乎翻倍。”
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 中性 消费型平台成本取决于工作负载设计和成本纪律，而非用户数
At the higher end, Snowflake and Databricks were probably the most expensive platforms but also the most complete for enterprise-scale data management. Snowflake is very strong for governed analytics and warehousing, while Databricks felt stronger for end-to-end Lakehouse ML and AI development environments and whatnot. Both are consumption-based, so costs depend heavily on workload design, compute, storage, and how disciplined you are with cost controls. If I had to rank them from cheapest to most expensive for our expected Untrap workload, I would say self-hosted n8n or open-source orchestration is at the bottom, then vector databases and Azure AWS, and then Snowflake and Databricks at the high end. Relative to the cheapest option, Databricks or Snowflake could easily be 30x more expensive, but that comparison is a bit unfair because they are solving a much broader problem.
来源背景：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 - 中性 Snowflake 未管控的查询会产生'惊吓账单'——消费弹性高但客户会因此设防
EXPERT: I will put GCP as 7 out of 10 because GCP is good, but then you need to have some kind of technical skills and alignment to leverage it fully. For example, everything is Kubernetes-based, so skills are needed. That's why 7 out of 10. Databricks I will give 9 out of 10 because if you are really kind of data-science-centric, you are a techie organization, process a lot of data, it can really be cheaper if you can kind of have skills within the organization. Snowflake also I will give 8 out of 10 because more of a BI-based centered, but very soon you'll see this goes into a slippery path where unmanaged queries or even if you check queries or instances will eat up a lot of money and you will not be able to control and will get shocker bills. Then if I come to Azure, I will give them 6 out of 10 because so much centered towards Microsoft Stack and Microsoft Stack ecosystem. It works very well there, but at the same time, you have compatibility issues and many of most of the environments you will have some kind of hybrid cloud or on-prem and all that integration issues. So saying that, you can always be in Azure and Microsoft is - I don't believe that. Going to AWS. AWS also I will give 8 out of 10 because of their breadth. You need to give some kind of longer commitment to get best value, but still they are really good in terms of features and standard offerings and ecosystem. So a lot of other third-party vendors will provide you a lot of feature value add components which you can directly integrate and use within the environment without doing small and unmanaged developments. So those kind of very tested components in the ecosystem are available in AWS. So that's my two cents.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 - 支持 Snowflake 最大 BI 负载来自 Power BI 和 Excel；AtScale 让这些查询和计算几乎无摩擦地流入 Snowflake
The other would be the fact that AtScale speaks the native language of Power BI and Excel. And while that sounds simple, no one else has had that until Cube and even Cube doesn't really have that. So, and it's the biggest BI workload in Snowflake, it comes from Power BI and it comes from Excel and when AtScale is in the mix all of the queries and all of the compute that comes with it can flow into Snowflake near frictionless. So there are some good reasons for why Snowflake should maybe consider buying AtScale and not just being an investor. It would be a lot of engineering time to actually, it'd be a lot of engineering time for them to actually build what AtScale has built over the past 12 years.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 中性 语义层通过智能聚合降低扫描量，用于性能和降成本——会减少部分计算
when a query comes through, if AtScale hasn't seen a query for the first time, and it hasn't predicted that that query was going to happen, then that query is going to hit all of the rows in the database. So the client's just going to have to accept that. But because then AtScale has seen that query before, and it knows the data that it needs to access, it will then create this intelligent, it'll create an aggregation of that, you know, and something we call an intelligent aggregate and it will do that for performance reasons, but it will also do that for cost optimization reasons. You know, we don't want to be scanning through billions of rows of data when you only need 100 million. And so AtScale will do that. And it will also then tune that over time because it's constantly analyzing user query patterns.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 支持 AI 用例落地会反向拉动数据平台现代化 (目录、库存、订单、CLV 数据)
Very simple, but as we're going into, and the production business case was, let's bring it in and that should be a customer-facing AI assistant. We looked at the workflow, we realized the real need was actually better product catalog data, inventory, freshness, order history, returns data, and CLV, the customer lifetime value attributes that would be sitting in CRM or a MarTech platform, and overall, of course, the original customer service knowledge as well. That resulted in basically more of the AI use case pulling forward the modernization of the data platforms, the technology layer to make that happen. It's still underway. And that is the nature of, fundamentally how, as we think about productionization, the boring topics around data management and then governance and traceability and audibility become real concerns.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 50%+ 的 AI 项目包含降 token/模型路由步骤，粗糙原型通常能降 30-60% 算力
BJ: I mean, the high level is, obviously, the thinking behind from the client side is optimization is spreading very quickly, right? In early AI prototypes, many clients, particularly in POC, sent way too much context to too large of models too often. These prototypes absolutely work, but the economics did not always scale. So in skilled engagements, I would say 50% plus now includes some kind of a step where we have to reduce the tokens and we have to think through model routing. We have to look at prompt optimization, teaching people how to do prompts or pre-filled prompts, caching, retrieval, pruning, guardrails, of course, to my earlier point. I'll say when it's a poorly optimized version, so someone just prototyped it, we commonly see anywhere from 30% to 60% compute reduction easily, sometimes more depending on the workload.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 支持 Cortex Code 理解完整 schema/语义模型/权限，快速建新层和 view，从而增加消费
All of the databases that you have set up within Snowflake, it understands what is where and the schemas that connect them all, right? That is the missing piece because Snowflake owns that layer and Cortex is sitting directly on top of that. It allows you to quickly create new layers directly within your Snowflake environment. And that obviously increases the amount of consumption that's created as you create new views that are sitting on top of it. It creates more consumption from an analytics perspective, etc. So that's how they're raising their own tide on all things Snowflake. Now, it also creates a lot of that efficiency from an engineering perspective, and that helps me as a system integrator, an implementer, that helps us deploy these projects faster and with more efficiency and with lower headcount, which reduces the overall cost.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 更低的实施成本让客户预算更多流向 Snowflake 消费和计算
It creates more consumption from an analytics perspective, etc. So that's how they're raising their own tide on all things Snowflake. Now, it also creates a lot of that efficiency from an engineering perspective, and that helps me as a system integrator, an implementer, that helps us deploy these projects faster and with more efficiency and with lower headcount, which reduces the overall cost. And that's a win as well, because now they put more of that into the revenue going directly to Snowflake through consumption and compute. Now, as far as how does it do that? And I know that's the second part of the question around, hey, how is it? What are those use cases that are happening? And I think the big hit that I'm seeing is that, yes, you have a repository of code, but what you don't have, at least historically, was that understanding of what's happening within the Snowflake layer, and how to ultimately get this deployed out into Snowflake Intelligence in a faster perspective. And so let me just go through some examples of the use cases of how we're doing it. Typically, we're plagued with, hey, for instance, from a data governance perspective, I'm always asked, hey, we need to please the auditors.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 反对 Cortex Code 能找出最贵的 pipeline 并优化，'吸走本该流向 Snowflake 的收入'
So you can see how you can just sequentially make this more and more refined and ultimately... I mean, to be honest. By creating that type of efficiency, now you're sucking away some of the revenue that would be going to Snowflake, but it's in the best interest of having a really high performing and humming RDBMS layer, right? And so that's really important. And then it also integrates in with all of your other tools that you're using, right? So, for instance, we're a big dbt shop for creating... and dbt is its own separate product, but you can quickly create DAGs using dbt, which is a directed acyclic graph, but you can quickly roll these out very quickly using these other tools that Snowflake has partnered with so you can see they're growing their ecosystem outwards to all the other data tools like dbt and other ELT engines like Talend and Informatica that allow you to build these efficiencies. And then you can also set up things like, for instance, for Cortex Analyst, you can ask it just to create a brand new semantic view using Cortex Analyst, and it will build that out for you. So you can see how it's using Snowflake-specific Cortex tools, where instead of me having to build out those semantic views, I can just quickly ask Cortex Code to generate that for me and build those Cortex agents.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 中性 每次迭代都用 Cortex Code 审代码并找降本方法
FC: Our team has really responded well. When I challenged them, you know, initially those grumblings around, hey, we don't want to learn this new Cortex Code layer. But I think people have realized how easy it is, right? And the scary part is really that you can have a business analyst running Cortex Code on a daily basis. The only problem that I have run into there is that analysts can get themselves into deep water very quickly. And I always say to my team, although you can use Cortex Code, if you use it and just blindly implement it and put it into production, you will get burned hard. And you need to understand and review all code and we still have to do code reviews, etc. Although, I will say code reviews... you always do a Cortex Code review first and foremost... have it reflect on where is this code going off the rails. How could it and what are the overall costs and how can we reduce those costs as part of every iteration that we do before implementing any code base.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 口径不齐 MCP 若把 Snowflake 内部层传给 Claude 等外部 agent 就'可行'，护城河依赖客户不愿外传 schema——这一点是推测
FC: Excuse me. If you were able to pass all that through to Claude or Cursor or Windsurf, if you were to pass all the internal Snowflake layers into it, it would be feasible. I will say, you know, one area that Snowflake has really glommed onto and worked well with is the MCP layer, which is the Model Context Protocol, which does pass everything from a BB: Mm-hmm. FC: From a model perspective, so more of the AI layer in a standard fashion so that other tools could use it. I know Anthropic, which has created MCP, that's exactly why they want that information, right? From a Claude perspective, if you could pass all the information through MCP, then it would be usable. But with that said, you know, Snowflake is going to... that's their moat, right? If you think about it, anyone that's a Snowflake customer is not going to want to share their Snowflake setup, their semantic models, and their schemas, et cetera, and broadcast that across into Anthropic, right? That is their own internal workings and what they want to keep proprietary as, "Hey, you're a Snowflake customer, we're going to keep your data safe," as well as, "If you want to run it, just use Cortex Code." It's the most powerful coding agent for a data and analytics layer, especially for Snowflake users. And that's a great selling point for why someone who's on another RDBMS would say, "Hey, I want Snowflake as my enterprise data platform." BB: Yeah. And just last question on Cortex. What have you seen on a quarter by quarter basis? How fast is the ramp then?
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 伙伴 B 称 CoCo 在 2Q26 已占 consumption 近 3%，且周环比快速增长
* “Cortex Code is essentially a pixel-perfect replication of Claude Code, with extensive Snowflake-specific optimizations.” * “Cortex Code is the fastest-growing product in Snowflake’s history. Its contribution in 1Q26 was modest, but it became very visible in 2Q26.” * “We didn’t see a clear acceleration in 1Q26, but I expect it to be apparent in 2Q26.” * “Cortex Code already accounts for nearly 3% of Consumption in 2Q26, and is growing rapidly week over week.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 伙伴 D 认为 CoCo 本身不产生很高消费，主要是间接驱动整体消费
+ May ultimately be integrated with Snowflake Intelligence (customers want all of this automated behind the scenes rather than as standalone products) + Today its contribution is still indirect: “Is Cortex Code generating very high consumption in its own right? No. But it is a driver — it makes feature development on Snowflake easier and faster, which will indirectly drive overall consumption.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 Snowflake Intelligence 每次搜索都会额外查询数仓/Cortex，两层都产生消费
+ Warehouse compute consumption: queries against the data warehouse * “Snowflake Intelligence is like the front-end search layer when we use ChatGPT or Anthropic. When you press enter, it has to retrieve data from the data lake (structured and unstructured), then call the LLM. Doing all that work means it issues additional queries against the data warehouse, Cortex, or Cortex Code — every search is incremental consumption.” * Increased data-platform complexity, broader business-line user adoption, new product integrations, and customer demand for the speed and ease-of-use of Snowflake Intelligence are all tailwinds. * **Cortex AI vs. Snowflake Intelligence:** “Cortex AI is a different product. But Snowflake Intelligence uses Cortex AI capabilities under the hood to help answer customer questions. Cortex AI is generative AI intelligence — it’s what Snowflake hosts internally, exposing AI functions, providing access to a range of LLM models, and integrable within the data environment. It includes search capabilities, vector search, re-ranking, fine-tuning, etc. Cortex AI is essentially LLM-type intelligence, vector databases, vector re-ranking, etc. Snowflake Intelligence then runs on top of Cortex AI.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 分析师指出客户可能对 CoCo/Claude token 支出设限；CEO 承认大规模铺开时成本治理'绝对是问题'，Snowflake 在做账户/agent/用户级 token 上限，并用小模型 (Mistral) 替代 Opus 降本——这会压低单位 token 收入
And to give you concrete examples, a very large bank that we work with has told me that while they spend several hundred million dollars on data systems as a whole, it's a very large bank, the amount of money that they spend on the human capital that powers all of these various pieces of software and links them up is three to four times that much (00:38:25). And anything that makes that part of the labor force 10 times more effective is always incredibly welcome. Having said that, when we want to roll, for example, Snowflake Intelligence out to 10,000 users, cost governance is absolutely an issue, just like it's an issue in Snowflake when I want to roll products out at scale. And so we are doing things like cost limits at an account level, or at a particular agent level, or you want to be able to restrict how much tokens a particular user can be spending. Of course, it quickly comes back to having exceptions for very talented users that are actually worth the tokens that they are using. And that's the kind of infrastructure that we are really good at creating.
- 反对 CEO 主动把廉价小模型路由进产品以提高效率，意味着 AI token 收入单价会下行
Plus there is a lot of innovation that we are driving within these coding agent products themselves. As I said, they handle very complicated tasks, but not everything is complicated. If you want to summarize, for example, I mean I did something a couple of days ago to summarize Slack threads, and perfectly small models from Mistral are enough for that. You don't need the latest and greatest Opus models for summarizing Slack threads. We are building that – those kinds of capabilities natively into Snowflake, so that it can be efficient in what kind of models that it uses. But my short answer to your question is they're creating incredible value, but we are not resting on that. We are creating the controls that one needs in order to keep costs manageable as things continue expanding.
- 支持 采用 CoCo 的客户增长更快 (未量化倍数)
Second, Snowflake Intelligence and CoCo are seeing the fastest adoption of any new products in our history, opening new opportunities for growth as the first major product surfaces of the agentic control plane. And third, adoption of these AI products is increasing core platform consumption as customers move from questions to answers, from prompts to pipelines, and from ideas to production workflows on Snowflake. Customers adopting CoCo are growing even faster, and we expect that momentum to continue as adoption expands. The strength of our Q1 results reflects the powerful flywheel effect of the agentic enterprise. Importantly, this momentum starts with the strength of our core business.
- 支持 管理层称 CoCo 带动核心数据平台更多消费 (管道、agent、动态表、迁移)，属二阶效应，未量化
“What’s really interesting about Cortex Code is that it in turn drives more consumption on the core data platform—simply because it’s much easier to get projects done, whether it’s a pipeline, creating a new agent, setting up a new dynamic table, or even a migration.”
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 口径不齐 Funda 认为 CoCo 一半用例不是数据相关，而是吸收 Claude Code 的溢出需求 (与数据控制面逻辑不同，可能不持久)
“An Infinite Lambda engineer used CoCo to build a Customer 360 application in just 5 hours.” If you look closely at this case, CoCo is indeed absorbing Claude Code’s overflow demand. CoCo handled not only the database portion but also the work outside the database. In our Preview, we made one key call: “The reason CoCo is growing so fast is that half of its use cases are not data-centric. CoCo, by virtue of its speed, is absorbing Claude Code’s overflow demand.”
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 口径不齐 Funda 认为 CoCo 在吸收 Claude Code 溢出需求，因为自托管 Anthropic 模型而更快
* We were very bullish in 1Q26, and we think Cortex Code is taking in the spillover demand from Claude Code. Because Cortex Code hosts the Anthropic model by itself, it is faster than Claude Code. This is the fastest-growing new product in Snowflake’s history. * We were bullish in 4Q25, because the Gen2 optimization was close to finishing, and Snowpark Connect also turned from a negative impact to a positive one. * We were bearish in 3Q25, because we saw that Gen2 optimization and Snowpark Connect were both having a negative impact on Snowflake at the same time, and Frontier Labs was optimizing their Snowflake budget... * We have been bullish on Snowflake since 3Q24, and we believed that the new CEO was making the right and aggressive organizational changes to Snowflake.
来源背景：作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。 - 口径不齐 案例：半导体客户最初用 CoCo 是为了优化查询省钱，之后供应链部门扩展使用；公司称消费增加 (无数字)
And they adopted CoWork because they had over 1,600 locations and the reps were having a tough time getting the information out to the customers and answering in a unifying way. And, so they used AI agents to actually build the responses with CoWork and they are able to answer with greater consistency across all the 1,600 locations, faster in answers and increase the customer satisfaction. If you look at the next example, this is a semiconductor company. They originally bought deployed CoCo to optimize their queries and save money. Their supply chain department then actually picked up on CoCo and started using it. And within the supply chain department, they were having a problem because it was really complex manual calculations on ordering inventory. Through the use of CoCo, they are able to decrease the time, increase the accuracy and reduce the cost. And so, this is another example of how we've seen CoCo play out in our customer base. In both of these, as you can see, CoCo is actually increasing the consumption of those customers. All right. Let's talk about how this is working at Snowflake.
- 口径不齐 CEO 承认 CoCo 让成本优化更容易，但称客户做的事更多，净消费仍增加 (未量化)
So I see a slew of techniques like this show up as there are concerns about costs. And honestly, we also want to put them into products like CoCo and CoWork as we continue to innovate with them. One of the things, ironically, that Christine and I are quite happy about is that things like Snowflake optimization is a lot easier with CoCo. And even though people optimize more with CoCo, we actually get more consumption anyway because they just do a whole lot more. And that's the benefit of a general purpose Swiss Army knife like tool that CoCo is and we have already worked on things like per user limits, per account limits for how much some tool should be used. In fact, I'm having a conversation with a company about deploying CoWork for 3,000 odd sales folks within that company and part of the guarantee that they want is a per user limit.
- 中性 Spark 遗留 API 客户用 CoCo 转成 Snowpark，快 5 倍、成本更低：迁入变容易，但单个工作负载的 credit 更少
We've been working on it and we know that at the end the day it's going to be compatible but not super fast. And we're busy doing that. And in the meantime, the customer said, oh, we tried CoCo. We converted to Snowpark and we're done, we're good and it's five times faster and cheaper by implication. So, we're actually – we have a renewed push on Snowpark. We mentioned it to some of you that we are seeing increased momentum of Spark migrations, use of Snowpark just because, hey, the pain that represent that converting code is no longer as painful as it is which is easier to do.
- 中性 OpenAI 新的数据分析产品是调用 Snowflake MCP SQL 连接器的技能：外部 agent 跑在 Snowflake 上，但 AI 层可能落在 Claude CoWork 而非 Snowflake
You saw that Daniela yesterday. We work effectively together. Yes, there'll be some cases where they would prefer obviously to have cloud CoWork rather than CoCo or Snowflake CoWork. That's fine. I think there's an element of maturity that we have about how we approach this relationship. That's 100% calling into our MCP SQL connector. We gave them a quote on Friday. Right here.
- 中性 7,000 个用 CoCo 的客户里多数只有一个用户在用，还没转到 agent 工作方式
And while we have – clearly, we have a lot of work to do in terms of driving adoption by each and every one of our customers. It's still pretty early. When we talk about the 7,000 customers, it will be more the case that there's like one user that's gone and done. It's not that they have switched over to this agentic way of operating and it's not like we have made migrations easy enough that anyone can migrate from anything into Snowflake so there's work to be done. I would say in the near-term, there's just – that's where there's enormous potential for us but it's sort of playing our game of be a good data platform. CoCo accelerates everything that you can do with the data platform, including creating products like CoWork that people can get even more value from.
- 反对 同一提示下 CoCo 写的 SQL 更高效，客户 credit 消耗更少 (AI 让查询更省)
相比之下，Cloud Code 虽然底层也能访问基础模型，但它并没有针对为 Snowflake 或 Cortex 编写 SQL 查询进行优化。如果将相同的提示词分别输入给 Cortex Code 和 Cloud Code，Cortex Code 生成的 SQL 查询在效率、速度和准确性上都远超 Cloud Code 生成的查询。这能为客户带来更低的积分消耗和更高的性能。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 反对 长期看用 CoCo 写 SQL 能帮客户省很多钱
回 答 者：您也可以直接使用它，可以将 Snowflake 直接连接到 Cloud Code。但两者的区别在于生成代码的质量。例如，当您要求编写一个创建客户表或执行复杂连接的 SQL 查询时，Cloud Code 生成的查询质量与 Snowflake Cortex Code 生成的是不同的。Cortex Code 生成的代码针对 Snowflake 进行了更深度的优化，调整得更好，性能也更高。从长远来看，使用 Cortex Code 编写 SQL 查询可以为客户节省大量资金，因为它遵循了 Snowflake 的最佳实践。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 支持 CoCo 把业务分析师等非技术人员带上平台，用户群扩大会拉动计算和存储消耗 (与专家C“95% 是开发者”的说法矛盾)
此外，Cortex Code 的推出也为 Snowflake 平台带来了更多元的用户。过去主要是开发人员使用 Snowflake，而现在，许多业务分析师等非技术人员也开始通过 Cortex Code 使用平台。用户群的扩大将显著提升 Snowflake 计算和存储资源的消耗，从而增加公司营收。尽管竞品价格低 40%，但这种由新用户带来的增量消费是 Snowflake 的核心优势。未来，随着使用量的增加以及模型微调与优化技术的进步，成本会逐步下降，这 40% 的价格差距预计将显著缩小。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 支持 CoCo 直接消耗还很小；消耗来自用它建出来的东西 (管道、RAG、agent、Snowpark 应用)
回 答 者：Cortex Code 最近已正式发布（GA），我们需要区分其直接贡献和生态系统贡献。目前，Cortex Code 的直接消耗量仍然很小，因为它刚刚发布，尚处于企业推广的早期阶段。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 开发路径从 SQL→Snowpark 变成 CoCo→Cortex AI 服务→Snowpark，最终拉动计算和存储
关于工作负载与 Cortex Code 之间的联系，我将其视为一个开发者接口。从历史上看，开发者的路径是使用 SQL，进入 Snowpark，然后作用于 Snowflake。现在，路径变为开发者访问 Cortex Code，后者再调用 Cortex AI 服务，进而访问 Snowpark，这最终增加了 Snowflake 的计算和存储消耗。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 消耗主要不来自编码助手本身，而来自用它构建的成果
Cortex Code 的作用是让开发者能非常便捷地生成 SQL、构建数据管道、创建 RAG 应用、构建代理以及开发 Snowpark 应用，并最终部署 AI 工作流。因此，消耗量主要并非来自编码助手本身，而是来自使用该助手所构建的成果。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 客户账单可能涨 40%，原因是跑了更多 AI/Cortex/Snowpark 工作负载，不是单价上涨 (举例性质)
最后，客户可能会感觉账单金额增加，但这通常是由于消耗增长所致，例如运行了更多的 AI 工作负载、Cortex 或 Snowpark 任务。客户的账单可能会增加 40%，但这主要是因为他们运行了更多的工作负载，而不是因为 Snowflake 提高了单价。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 agent 可能是 Snowflake/Databricks 长期消耗的最大驱动，潜力可能超过最初的 AI 浪潮
回 答 者：简而言之，是的。Agent 可能是 Snowflake 和 Databricks 长期消耗增长的最大驱动力，其潜力甚至可能超过最初的 AI 浪潮。原因是人类消费数据是间歇性的，例如商业分析师（BA）可能每天运行 10 次查询，或每周生成一次报告，消耗有限。而 Agent 会 7x24 小时不间断地消费数据，包括查询数据、检索文档、搜索知识库、执行工作流、调用 API、触发其他 Agent 以及更新记录。因此，1 个 Agent 产生的平台活动量远超 1 个人类用户。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 客户现在有几千名员工、几十个 agent，几年后可能是几万个 agent
这对 Snowflake 至关重要，因为 Agent 的每一项活动，如数据检索、向量搜索、元数据查询、模型推理、工作流执行和结果存储，都会消耗 Snowflake 的服务。我观察到，目前多数客户拥有数千名全职员工（FTE）和数十个 AI Agent。但几年后，他们可能依然是数千名员工，但会拥有数万个 Agent，这个比例将发生巨大变化。因此，更大的机会在于有多少生产环境的 Agent 在 Snowflake 上运行，这会带来极高的计算消耗。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 大部分 agent 还在试验，生产环境里很少，真正的加速要等进入核心业务流程
Agent 的普及也会增加 Snowpark 的消耗，因为许多 Agent 工作流依赖于 Snowpark、Cortex AI、搜索和 Agent 编排等服务。但需要注意的是，并非所有 Agent 都能创造经济价值。当前阶段，多数应用仍处于实验性质，生产环境中的 Agent 数量还很少。真正的消耗加速，将发生在 Agent 成为核心业务流程（例如处理索赔）的一部分时。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 CoCo 消费没有单列；直接贡献很有限，主要是间接推动平台增量消费
回 答 者：这个问题很难回答，因为这部分消费没有被单独列为一项。Snowflake 已经转向了基于 AI credit 的使用模式。Cortex Code 的影响更多是间接的，它通过赋能来驱动整体消费的增长，其直接贡献非常有限，但它确实在推动平台增量消费的增长。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 agent 以机器频率运行，多步工具调用每一步都消耗计算
I mean, one is direct consumption in uh in uh in in Snowflake and then uh they also help save cost because some of these are recurring and they sort of they can run at human uh sort of machine frequency and not human frequency. So, uh you can have them running recurrently daily or monthly or whatever and the business user can only only uh only need the output. Uh, and then you can you can also have multi-step agents to chain multiple tool calls together and each of them hits compute. So they're basically a way of increasing compute uh, and in different well, ultimately save labor basically on the business side. 嗯。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 Jefferies 渠道称 CoCo 自 4 月可靠性明显改善，在 text-to-SQL 和治理数据用例上优于第三方编码 agent
Jefferies came away from Snowflake Summit and Investor Day with increased confidence in Snowflake’s AI monetization story, highlighting rapid adoption of CoCo and improving GTM execution. The firm’s checks suggest CoCo has become materially more reliable since April and is outperforming third-party coding agents on Snowflake-native workflows, particularly text-to-SQL and governed enterprise data use cases. Jefferies argues context and governance are becoming key differentiators in enterprise AI, positioning Snowflake favorably as more agents move into production. The firm also sees improving sales productivity, faster migrations, and a clearer path to GAAP profitability by FY28, while noting Snowflake still trades at a meaningful valuation discount to Datadog despite growing AI momentum.
来源背景：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 - 支持 Snowflake/Databricks 原生监控基本免费，客户只付 SQL 检查的计算费——功能变现为 credit
**TL** [00:19:04]: You can turn on the monitoring in Snowflake or Databricks basically for free. You're going to pay just the compute cost of running the SQL or the checks that you're looking to run. I think companies start that way. We saw that pattern quite a lot where companies would start with that, and it works up until a point, but once they validate we need this, how do we then elevate that into a place where we can measure it across the organisation and have all of the enterprise bells and whistles around it so that you can organise it by domain, by department, by user, give people certain permissions to run certain checks that might not be data engineers the more left the quality checks go or even right. You see a lot of folks in analytics or even as domain specialists within a business unit, they might not have access to the underlying tables and they might not be the admin there, but they still want to deploy checks and balances on those assets, and it's really difficult to do that directly in one of the warehouse providers, if you're not an engineer, as an example.
来源背景：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 - 支持 客户 token 消耗基本按月翻倍，AI 预算翻倍常见
当前企业客户在AI领域的投资态度非常积极，预算投入不成问题。企业普遍认为，如果因为预算不足而延误Agent的上线，将造成重大的商业损失。因此，预算翻倍的情况十分常见。许多CIO认为，在当前的AI时代，企业必须跟上（catch up）甚至力求领先，这意味着必须进行投资。反映在消耗上，客户的Token消耗量基本以每个月翻一倍的速度增长。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 支持 平台数据量月环比约 6%，年化接近 100%，非结构化数据占新增主导
月度环比增长6%指的是当月数据总量相较于上月总量的增长率。例如，若四月份的数据总量为100单位，则五月份的数据总量将增长至106单位。这意味着年化增长率接近100%。这些新增数据主要来源于企业持续运营中产生的业务数据，例如游戏公司新产生的玩 家行为数据，或制造业设备运行产生的日志文件。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 中性 中型科技公司 RAG 放量后评估云厂商以外的数据库，Snowflake/Databricks 只是候选之一
公司关注过 Snowflake 和 Databricks，但并非只考察这两者。重点在于，公司开始评估当前云服务提供商以外的新选择，以寻找查询效率更高的数据库。检索增强生成是一个复杂系统，不只是存储代表文本语义的向量，还需要采用适当的索引方式来提高查询速度，并要求数据库支持不同类型的查询。相关策略很多，并非每种策略都适合企业的基本应用场景，因此需要具备足够全面的能力，才能尽可能适配具体应用场景。
来源背景：未具名专家，自述曾在泰国一家传统大型企业、一家覆盖英美的中型科技公司 (AI 非唯一业务) 和美国小型初创工作 (明示，)，职位未披露；中型科技公司经历中参与过 RAG 早期建设 。视角局限：样本为个人经历的三家公司；更偏工程/AI 应用，未直接谈 Snowflake 账单。 - 支持 agent/自然语言接口把一家 1,000 人企业中能用 Snowflake 的人从 40-50 人扩到 500 人以上，每个问题都会转成 SQL 消耗 credit
回 答 者：Agent应用是Snowflake消耗增长最快的领域。以一家平均拥有1,000名员工、使用Snowflake的企业为例，过去由于需要具备SQL技能，通常只有约40至50人使用Snowflake。如今，借助Agent、LLM和Snowflake的能力，企业中已有超过500人可以访问Snowflake。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 Cortex Code 生成一个 Streamlit 应用消耗数百万 token，上线后多人使用继续消耗算力
如今，借助 Cortex Code，任何业务用户都可以使用英文向 Snowflake 输入提示词，并在不到 1 小时内获得 Streamlit 应用程序。每个 Streamlit 应用程序在创建过程中都会消耗数百万个 token，同时也会消耗数据和算力。每个 Streamlit 应用程序还会由多个用户使用，因此将消耗越来越多的算力。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 Horizon Catalog 的上下文让同一条 SQL 所需 token 少于直接用 Anthropic/OpenAI——对客户省钱，但也意味着单次查询 token 收入更少
另一个差异在于，当客户在 Snowflake 外部或内部使用 Anthropic 时，由于 Snowflake 在 Horizon Catalog 中具备上下文信息，生成同一条 SQL 查询所需的 Token 数量低于 Anthropic 或 OpenAI。Snowflake了解数据表、查询和数据集等上下文，因此针对同一查询使用的 Token 更少，客户也能大幅节省 Token 使用成本。这正是将 Snowflake 作为 AI 和智能体平台使用的优势。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 AI 让 20-50 人的 Snowflake 团队可缩减到 3-4 人 (客户人力节省，专家认为消费反而扩大，未量化)
其次，借助Snowflake的大语言模型和自动化能力，客户可以自行构建数据管道，并通过AI和大语言模型自动完成数据质量、安全、治理和信息安全等工作。如果客户原本拥有一支50人或20人的Snowflake团队，借助Snowflake大语言模型、Snowflake Cortex Core和Snowflake Cover，可以将团队缩减至3人或4人，从而大幅节省人力成本。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 加速由 AI 驱动：每个落在 Snowflake 上的 AI 工作负载都增加仓库消耗，因为 agent 查询更多更频繁；但 AI 采用慢于预期
回 答 者：我们看到AI带来了一定的加速，这就是主要原因。此前我们也讨论过，AI的采用速度可能比预期更慢，但在Snowflake环境中实施的每一项AI工作负载都会增加仓库消耗，因为代理会进行更多、更频繁的查询。因此，这种加速仍然可以认为是由AI驱动的。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 支持 Cortex Code 等 AI 编码 agent 让开发提速 30-40%，推高平台使用量
不应将 Cortex Code 视为终端产品，它更像是 Snowflake 针对系统集成商生态获取利润空间的一种方式。至于具体比例，很难判断。它能够提高 Snowflake 的开发速度，进而增加使用量。根据我们对 AI 智能体或 AI 编码智能体的观察，开发速度通常会提升 30%-40%。我推测，采用率可能大致呈线性增长，但无法提供更具体的数字。这基本就是其背后的发展逻辑。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 反对 AI 编码工具已统一为 Cursor 全球标准，切换到 Cortex Code 需重新评估审批——外部编码 agent 先占位
回 答 者：我们并不是后来转用 Cursor，而是一开始就使用它，因此具备 first-mover advantage。Cursor 已经成为 global standard，并已通过 global architecture approval 和 global governance，是获得批准的工具。大家只能选择1种开发平台或开发环境。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 支持 商业类 AI 应用 (推荐、摘要、工作流自动化) 至少 80% 部署在 Snowflake 上；R&D 约 20% 在 GCP/AWS 等
一般的商业应用，例如 Next Best Action、推荐、使用生成式 AI 进行文档摘要、内容创建、视觉相关应用，以及后台办公自动化和工作流自动化，通常都部署在 Snowflake 上，可以说至少80%都在 Snowflake 上。研发领域可能有20%部署在 GCP 等环境中，但相对较少，而且还在进一步减少。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 agent 架构是多厂商的：一个编排 Hub 调度 Claude Code、Cortex Code、ServiceNow 等不同来源的 agent
但后端有所不同，核心在于编排。不同的问题和工作流可能会触发不同的 Agent，因此系统中会存在多个 Agent。这些 Agent 的开发环境也可能不同，有些通过 Cloud Code 创建，有些未来可能通过 Snowflake Cortex Code 创建。此外，还有 ServiceNow 等不同厂商提供的开发环境，分别负责不同的业务流程和业务部门，因此会形成不同的 Agent。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 反对 真正执行任务的 agent 仍处早期、不成熟，1-2 年后才见进一步发展
回 答 者：一开始，我们也在做一些包含几个 Agent 的案例，并开始与业务部门沟通。总体而言，这方面仍处于较早阶段，尚不成熟，需要进一步完善和开发，使其更加易用，这是目前的主攻方向。接下来 1、2 年会看到进一步发展。如果成功，将带动大规模投资，因为无论是工作效率、业务简化，还是人事调动与布局，都会受到非常大的影响。因此，这是一项具有战略意义的工作。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 支持 Cortex Code 成本优化/可选开源底座后，客户查询量立即增加
回 答 者：是的。我们已经看到客户的查询量增加，客户的关注度也有所提升，这将推动使用量增长。客户非常看重这种灵活性。从灵活性和成本角度来看，这种方式得到了客户的广泛认可，因为Snowflake非常清楚，这最终会为其带来更多流量。我们已经观察到客户使用量立即增长的迹象，这将进一步提升Snowflake的消耗量。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 支持 AI 有叠加效应：agent 带来额外查询和仓库计算，部分 AI 驱动的消费落在传统类别，AI 贡献可能超过 15%
在我们的企业消费中，我会根据产品类别大致归属这些百分比。但还有一个叠加效应：Cortex 不仅仅带来直接的 Cortex 收费。任何针对客户数据运行的 agent 都会产生额外的查询、带有转换的检索和仓库计算。因此有些由 AI 驱动的消费实际上也落入了传统类别。这就是为什么 AI 的贡献可能超过我之前提到的 15%。这也是推动我们讨论的同比和第二/第三季度增长的原因。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 重度采用 Cortex Code 的客户消费同比约 45%，可比传统客户约 30%
这就是 AI 在扩大工作负载和组织内用户基数方面的原因。我想强调的是，直接的 Cortex 消费在今天仍相对较小。Cortex code 导致的消费在战略上比直接消费更为重要。对于那些重度采用 Cortex Code 的客户，我估计其消费同比增长约为 45%，而对于使用传统 Snowflake 工作负载的可比客户，增长可能大约为 30%。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 积极采用 Cortex Code 的客户 Snowflake 消费比原本高约 10-20%
关键是，为了对核心 Snowflake 的提升产生实质性消费影响，我估计积极采用 Cortex Code 的客户其 Snowflake 消费大约会比原本高 10% 到 20%。但我不会把所有这些都直接归因于 Cortex Code，因为一个 agent 会生成 SQL、构建和优化流水线、探查你的 schema，并与 airflow 和加速开发协同工作，最终会产生更多的仓库查询、Snowpark 处理和数据工作负载。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 CoCo 推广受限之一是成本：密集 agent 使用在计算之上再叠加推理消耗
主要对那些有大量开发、数据工程、SQL或机器学习活动的客户相关。一些Snowflake客户主要是BI或分析用户；他们根本没有足够的开发使用场景。在大型客户中，也存在安全和治理审批问题。他们需要对AI工具进行标准化，并且可能已有对GitHub、Cursor或云平台等工具的既有承诺。所以全面部署需要时间。还有成本考虑：密集的代理使用会在你的计算之上引入推理消耗。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 推理成本降 70% 时客户不一定减预算，而是部署更多 agent——Snowflake 在模型层丢经济、在数据和计算层得消费
大型客户已经在朝这个方向迁移，因为开源模型可以降低推理成本并改善数据控制。对 Snowflake 来说，我看到两个影响。负面是开源模型对通过 Cortex 提供的专有推理的经济性带来压力，客户没有那么多理由为访问 LLM 支付溢价。但 Snowflake 的回应是变得更加模型不可知：Cortex 支持像 Llama 这样的开源模型并与专有模型并行，Snowflake 也对这些由 Llama 驱动的模型进行了优化。积极方面是，更便宜的模型可能意味着更多的 AI 使用。例如，如果推理成本下降 70%，客户不一定会减少他们的整体 AI 预算——他们可以部署更多代理并处理更多工作负载，同时仍需要对这些代理和数据进行治理。这就是为什么 Snowflake 在模型层面可能失去一些经济利益，但可以在数据和计算层面获得显著的消费增长。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 Cortex agent 可调用 Cortex Analyst、ML 执行和自定义工具，Snowflake 对 token 和仓库计算分别计费
不过，另一个方面也始终存在。当代理以分析为主时，Snowflake 也可以产生巨大的放大效应。Cortex 代理现在可以调用 Cortex Analyst 搜索、机器学习执行和自定义工具，而 Snowflake 会对 token 搜索和仓库计算分别计费。因此，Snowflake 更适合对客户数据集进行大规模推理的代理，而 MongoDB 则在需要状态、记忆和事务性应用数据的高频操作型代理方面具有吸引力的定位。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 agent 从试点扩到 1000 名用户时，支出可达 10-20 倍；只给 2 人用则几乎不增
从我们接触的案例来看，有些智能体仅供某个部门的2名员工使用，这类智能体从测试阶段进入生产环境时，支出不会明显增加。有些智能体经过测试后会在整个组织内投入生产，并突然面向1000名用户开放，这时支出就会大幅增加，可能达到10倍、20倍。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 反对 FDE 能让组织度低的数据直接被利用，可能蚕食 Snowflake '帮客户组织数据' 的收入
在数据层面，数据厂商自身的渗透程度会影响其产品使用量的增长。但在许多情况下，智能体应用或FDE也会处理非结构化数据及其他数据源，并通过整合使其发挥作用。从这个角度看，FDE可能会蚕食Snowflake的收入，因为Snowflake的价值主张是帮助客户组织数据，并在这方面提供出色的服务；而FDE也可以通过串联工作流，让组织程度稍低的数据得到有效利用。因此，最终结果会是多种情况并存。
来源背景：微软前沿部署工程师 (FDE) 团队成员 (纪要明示其团队直接服务 14 家 FDE 客户、可观察约 70 家，)，现任，职级与年限未披露。视角局限：微软立场，倾向推荐 Azure 体系；对 Snowflake 的判断是推理而非账单数据。 - 支持 实质采用 CoCo 的客户，Snowflake 消耗目前约增长 15-20%；大部分新增支出开始来自 CoCo 和 CoWork
Um so if there are, you know, um 500, 600 named users that are using um Snowflake, probably only 30, 40% of them uh are going to use Coco. So there is a base um license, user-based license for Coco, but the real consumption is the more number of queries they make, and also if the queries are using multiple models, if the queries have um, you know, multi-step uh processes so that uh token or consumption increases, that is where uh customers are spending. So I would say for those customers that have um meaningful, you know, uh investments already, um consumption, at least in my view, is is going up somewhere around, you know, 15 to 20% right now. It's not, it's not still, you know, very big, but it's starting to, you know, um most of the incremental spend um is is coming on um on Coco and Co-work. Mhm, 明白。 OK, 然后想问一下就是Snowflake和其它平台的竞争的一个情况, 尤其是DataDog, 然后专家的话认为在这个季度的话, 两者的竞争有变得更加激烈吗? 对, 包括大家就是DataDog和Snowflake之间的话, 客户的话是否有迁移的一个情况? 对。 Mhm. And next question is about competitions.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 即便客户已有 Claude Code，仍会用 CoCo，因为它在数据环境里建管道、做多步提取
许多客户的关注范围已经远远超出AI编程。编程领域已有多种可用工具，各大软件供应商也提供相应的编程解决方案。Coco是Snowflake提供的编程解决方案。即使客户已经使用Claude Code，仍会使用Coco，因为Coco主要在数据环境中运行，可用于创建数据管道、执行多步骤数据提取等任务。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 反对 并非所有企业 AI 应用都经过 Snowflake：Copilot、SAP Joule、ServiceNow Now Assist 自带数据层
这一问题较难一概而论，因为并非所有客户都需要与Snowflake集成。例如，使用Microsoft Copilot的客户可以在Microsoft生态系统内运行；使用SAP Joule或ServiceNow Now Assist的客户，也不一定需要与Snowflake集成，因为这些产品内部已经具备某种嵌入式数据解决方案。因此，在部分应用中，Snowflake的重要性可能相对有限；但在大量其他应用中，Snowflake仍会与ERP等系统集成，用于支持复杂的企业工作流，并整合来自ERP、CRM、大型主机和内部AI应用等系统的数据。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 agent 需要 agent-ready 数据层，Snowflake 和 Databricks 都会受益
BJ: Great question. And I think we are all moving towards a future where it's multi-model environments, right? And a lot of weight towards the open source models. I think absolutely the Chinese labs have proven that they can do it more efficiently, of course, distilling the outputs from the larger models as well to make that happen. But at the same time, the aggressiveness and how they're sharing it in open source and MIT license accordingly does encourage at least the naysayers to pay attention to the models and not see them as a threat, right? Which has led to even adoption of open source models from Meta and NVIDIA as well. NVIDIA's own Nemotron, and Nemotron is also something in consideration as well. My clients are actively leveraging it for not so complex workflows and it's working effectively and I think we're going to see a lot more of that. I think both Snowflake and Databricks will benefit from that because at the end of the day, the agent needs an agent-ready data layer and I think that's where there's a lot of work to be done. We're not looking at just dashboards anymore; we have to act on that dashboard and there's a lot of work and growth expected.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 中性 所有客户都在担心算力成本，因为难以提前预测
At the same time, I think a lot of that is necessary to really model the governance structure so that we can at least from finance to functional technical aspects, be more reliable in the ability to build business cases and the contracts that are associated with those cases accordingly. And then the movement happens. I think right now, fundamentally, every client of mine is worried about compute costs. And there's DG: Right.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 支持 保险客户从试点到多业务线生产，Snowflake 计算消耗增长 5-8 倍，主要来自理赔人员等用户的查询量而非数据量
And then you have an uh a sort of over proportional consumption increase. Uh, but but but but maybe giving you a flavor, we're still talking about a kind of sort of massive consumption uh compute consumption increase. So, so for for the insurance company, I think they have increased their snowflake compute by, you know, five to eight times, uh, you know, going from pilot to to multi-line production. Uh, and this is driven by uh query volume from claims agents, etc. So it's not so much data volume growth. Um, and and this is something that IT budget, you know, it's it's hard to really budget for this. Um, and and uh I I'm not private to exactly how they did the budgeting, but but it's uh, um, I I don't think uh they they sort of calculated with this, basically.
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 支持 从 PoC 到生产，AI 基础设施消耗上升一个数量级：样本数据→全量数据、非结构化内容进 Cortex、Snowflake Intelligence 查询量上升、管道持续运行
回 答 者：基本增长了1个数量级。以Snowflake为例，主要原因是数据摄取量增加。在POC阶段，通常使用经过筛选的样本数据集；进入全面生产后，数据源会全部接入，因此数据摄取量大幅上升。通常还会有非结构化内容输入Cortex功能，查询量也会随之增加，例如通过Snowflake Intelligence进行查询。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 消耗增长很大一部分来自架构低效 (智能体与向量库之间没有缓存、编排差)；厂商把消耗曲线当 AI 价值证明，Snowflake 也这样展示，需谨慎
回 答 者：消耗量确实在增加。值得注意的是，消耗增长很大程度上源于架构效率不足。例如，如果智能体与向量数据库之间没有缓存层，或者编排设计不合理，就会产生额外消耗。供应商通常会将消耗曲线作为AI价值的证明，Snowflake也会这样展示，但对此需要谨慎看待。部分消耗与业务影响成正比，但如果不重视架构设计，很多消耗实际上来自不必要的架构复杂性，通过优化管线本可避免。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 5-10 倍的资源消耗增长不可能一直被接受，企业会通过有纪律的架构设计抑制消耗
不过，如果Snowflake或MongoDB的资源消耗分别增长约5到10倍，这种情况不可能一直被接受。未来企业会更加重视有纪律的资源使用和合理的底层架构，并通过针对性设计抑制资源消耗增长。类似情况也会出现在其他AI领域，例如模型公司的Token消耗等。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 中性 规模化是多年过程：理赔之后还会扩到承保、财务、HR、采购，底层消耗会持续增加
对于所有企业客户而言，这都将是一个持续多年的过程。企业仍可在流程中的人工环节增加代理，也可以纳入更多产品线。例如，在全面实现理赔和保单管理自动化后，还可以进一步拓展至承保、通用财务与会计、人力资源和采购等领域。因此，尽管这2个案例已经实现规模化，但扩展仍将继续。对任何客户而言，这都是一个持续多年的过程，也意味着底层资源消耗将不断增加。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 建模时 AI 负载倍数不宜用 10 倍上限，5-6 倍或 7 倍更合适；差异主要来自成本控制
A:主要差异不太可能来自特定的行业垂直。底层用例会有所不同，但除此之外各垂直之间没有重大差异。更重要的差异将来自成本控制措施，随着市场成熟这些措施会变得更明显。在建模时，不会更倾向于使用十倍的上限；更合适的假设是五到六倍，或者可能是七倍。除 此之外，预计各垂直之间不会有实质性差别。
来源背景：标题写'北美软件渠道商离职专家'，但与 08-26 Merit 音频为同一场，转写中专家自述负责北欧区域的实施/咨询商；身份局限见转写条 - 支持 进入生产后 6-12 个月，AI 相关 Snowflake 消耗增加约 2-3 倍，来自数据工程、特征准备、RAG 管道、治理和分析
A:在最初的 6 到 12 个月内，进入生产可能导致 Snowflake 的消耗增加约两到三倍，而 MongoDB 的消耗约增加两倍。Snowflake 增长更快，原因在于新增的数据工程、特征准备、RAG 管道、治理和分析工作负载。MongoDB 的增长则更多来自应用层数据、向量检索、上下文存储以及更高的事务量。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 反对 客户直接让 AI 检索 SharePoint/OneDrive/ServiceNow/Confluence，不经 Snowflake/Databricks 中间层，是重要的新兴架构
A:绝对有。我认为这是一种重要的新兴架构。并非每个 AI 工作负载都需要 Snowflake 或 Databricks 做中间层。对于以知识为中心的用例，客户可以部署一个 AI 检索层，直接查询 SharePoint、OneDrive、ServiceNow、Confluence 或其他存储库，同时保留源端的权限和治理。例如，Microsoft 支持从 SharePoint 和 OneDrive 检索内容，而无需客户将这些内容复制并维护在单独的索引中。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 反对 AI 增长不必然一对一转成数据平台消费，平台必须成为治理层
Databricks 也在朝这个模式发展，允许其生态系统中的 agent 连接到 Microsoft 365、Jira 和 Google Drive 等外部源。对 Snowflake 和 Databricks 来说，这一含义很重要：AI 的增长并不必然按一比一转换为数据平台的消费。这些平台必须通过成为治理上下文和数据的那一层来赢得在 AI 架构中的位置。如果它们不是该架构中的治理层，就不能假定每一次 AI 查询都会通过它们。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 支持 跨系统联结结构化数据、历史分析、大规模转换时 Snowflake/Databricks 更有价值，混合架构会占优
然而，AI 绕过数据平台与 AI 使数据平台变得不必要是两个不同的命题。直接检索对文档、政策和知识非常有效。一旦用例需要将结构化数据跨企业系统联结、进行历史分析、大规模转换或构建可复用的数据产品，像 Snowflake 和 Databricks 这样的平臺就变得更有价值。我预计混合架构将占优：AI 会直接在源头访问某些企业知识，而 Snowflake 和 Databricks 则仍然作为复杂分析和跨企业工作负载的受控数据与处理层。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 支持 AI 进入生产后客户不只为推理付费，还产生摄取、转换、治理、语义建模和 RAG 负载，对 Snowflake 有乘数效应
金融服务可能是最清晰的例子。一个大型银行不会只引入一个 AI 助手就停止；它会开始增加欺诈检测、反洗钱（AML）、客户服务 copilots、文档智能、开发者 copilots 和风险分析。对 Snowflake 来说，乘数效应来自数据端。当客户将 AI 推向生产时，他们不仅为推理付费，还会产生额外的摄取、转换、治理、语义建模和 RAG 工作负载。因此 AI 推动了更广泛的工作负载和更大的用户基础，对 Snowflake 产生乘数效应。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 支持 欧洲渠道商估计 Snowflake 超过 40% 的消费可能与 AI/Agent 相关
较为乐观的是1位欧洲渠道商专家。他观察到，随着AI Agent的采用，以及客户开始为企业AI部署做准备，本季度Snowflake及各类数据治理需求均受到关注，并在2季度出现明显增长。他认为，目前Snowflake上超过40%的消费可能与AI和Agent相关。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 北欧养老险投产后 Snowflake 上 AI 相关消费 5-8 倍，AI 原占 15-20%，可带动整体约 1 倍
该公司认为，AI相关Token成本仅占相关人员成本的1/4到1/5左右，因此可以为公司贡献较多利润。他们同时发现，进入生产阶段后，Snowflake上AI相关的Consumption可能增加5到8倍左右。AI相关Consumption原本所占比例约为15%到20%，这意味着AI相关Consumption的增长可能带动Snowflake整体增长1倍左右。同时，他们也提到了Snowflake传统的数据工程。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 北美消费品公司 POC 8-12 个月 + 部署 6-9 个月，Snowflake AI 负载 2-3 倍，整体消费 +40%
北美一家咨询公司的专家提到，一家消费品公司从POC走向生产阶段，POC通常需要8到12个月，生产阶段的准备和部署则需要6到9个月。在此期间，该公司在Snowflake上的AI相关负载增加了2到3倍左右，MongoDB上的负载可能也会增加2倍左右，由此推动Snowflake的consumption增长40%左右。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 AI Agent 对数据的调用量比问答式 AI 高一个数量级
另外1位北美软件渠道商专家认为，其客户主要来自AI采用速度相对较快的领域，下半年AI Agent的采用将逐步加速。尤其是部署在Snowflake上的AI，将持续增加对数据库的调用。AI Agent对数据的调用量比过去问答式AI高1个数量级，因此，他们对于Snowflake今年逐季度加速的趋势也持较为乐观的态度。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 积极采用 Cortex Code 的客户 Snowflake 消费比原本高 10-20%
积极采用 Cortex Code 的客户，其 Snowflake 消费较原本高出10%到20%，约20%左右。Cortex Code 也推动了更多企业使用 AI，效果优于 Cloud 和 Codex。在 Snowflake 的整体架构中，Cortex Code 对数据上下文的理解能力更强，也清楚应从何处获取相关数据，因此能够使用更少的 tokens、更快地响应。对客户而言，其总体成本优于 Cloud、Codex 等闭源大模型产品。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 CoCo 5-7 月消费是 1-4 月近 3 倍，仍月增 15-20%
1. CoCo consumption from May-July nearly tripled versus January-April, and is still growing 15-20% MoM. 2. At the current trend, CoCo growth in August-October could be more than 50% higher than growth in May-July. 3. Cortex Code is fast, and Snowflake can host the models itself, resulting in a strong user experience.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 CoCo 已理解 Snowflake 权限和 RBAC，客户在 Snowflake/Databricks 环境中用 Claude Code 价值较低
1. Cortex Code is still used in the data environment and does not directly compete with Claude Code. If Claude Code needs to pull data from five or six tables, developers must add logic for user permissions and role-based access. Cortex Code already understands Snowflake access and permissions, so it can apply this context without additional work. 2. Therefore, customers see less value in using Claude Code in Snowflake or Databricks environments because they must build additional logic around role-based access control and table permissions. Cowork has broader coverage. When data needs to be pulled from SAP, CRM systems, mainframes, or other sources, it competes more directly with Claude Cowork, OpenAI products, and other solutions.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 一家大型零售客户用 CoCo 后年开发应用数从 15-18 个接近 24 个 (单一样本)
1. Before Cortex Code launched, developers used tools such as Snowpark and container services, with Snowflake providing overall governance. The traditional development lifecycle included design, coding, deployment, and integration with Snowflake services. Cortex Code reduces the design and coding workload and some integration work, enabling capable developers to build more applications or improve applications faster. This will lead to more deployments, customization, and integration, driving incremental usage. 2. One example from my observations is a large retail customer that previously developed approximately 15-18 applications per year to manage e-commerce, supply chain, and related ecosystems. Cortex Code was deployed only recently, and the number of applications is already approaching 24. I estimate that by year-end, its application development capacity could increase by at least approximately 30%, reaching around 24 applications. Note that this is just one observed example and may not represent Snowflake’s broader customer base. Retail customers generate significant data usage through queries, indexes, APIs, and required data pipelines. 5. If application creation in retail grows at this rate, I expect the increase in financial services to be even greater. This demand has not yet fully flowed through to my customer base, but I have seen some activity since last month, and it is continuing this month. I expect this trend to persist through the rest of the year.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 尚未看到企业把 CoCo 作为通用编码 agent，多数企业同时用 2-3 个编码 agent (与 Funda 的 '吸收 Claude Code 溢出' 说法相矛盾)
1. CoCo understands the customer’s full Snowflake technology stack and environment, while tools such as Claude Code require more context to work effectively. Think of it as a Snowflake-native coding agent rather than a general-purpose coding agent. Most Snowflake customers are consolidating storage and compute within the ecosystem, making this context valuable, although many, perhaps most, engineering teams will also use Claude Code for development outside Snowflake. 2. I have not yet seen customers use CoCo for general-purpose coding. In projects heavily dependent on Snowflake, teams may use CoCo for a small amount of development outside Snowflake because it already has the relevant context and avoids switching between agents. But this depends on the specific use case. So far, I have not seen an enterprise choose CoCo as its primary coding agent across the organization. Among the enterprise customers I have spoken with, the vast majority, if not all, are using two or three coding agents simultaneously. The market is still somewhere between experimentation and consolidation. 3. CoCo’s unique selling point outside the Snowflake ecosystem would weaken, but even if it does not become the enterprise default coding agent, it can still be a strong growth driver. Enterprise customers continue moving more compute and storage into Snowflake, so even if the coding agent operates only within this expanding environment, it can generate substantial consumption. 4. Data migration and transformation are the largest use cases, while data pipeline development and event-driven workflows are also important.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 CoCo 让成本优化变得容易：可以让 CoCo 排查最慢查询、最闲置的仓库，成本管理 skill 是 CoCo 前十 skill——AI 也在压缩客户单位消费
It is not concentrated. For example, with, let's say, AI-native companies, they continue to be a small part of our overall revenue stream. And I think the thing that's also materially different this time around with folks that are investing is that the products like CoCo make optimization far, far easier than before. You can point CoCo at a query that's taking too long to run, or you can basically have it debug the top 10 longest running queries are the most idle warehouses, things like that are a lot easier to do. And in fact, our cost management skill in CoCo is a top 10 skill. And it's also the case that as a company, we have learned the lessons of the pandemic. And one thing that we stress with each and every one of our customers is the need to drive spend in an efficient way.
- 支持 使用 CoCo 的账户也在核心上消费更多 (飞轮)，但未量化
Our sales team is doing a great job with proving the business value of the use cases. And we're continuing to see great new logo additions. CoCo, when we look at CoCo, the accounts that are using CoCo are consuming more of the core as well. And so, there's this flywheel effect that we talk about. We had 9,100 CoCo accounts this quarter.
- 口径不齐 分析师要 AI 采用者 vs 非采用者的消费提升倍数，管理层明确不披露
Thank you. Yeah, I'll take a first cut and Brian will add on. At this time, we aren't ready to share the exact uplift numbers, but we do measure cohort behavior and as CoCo adoption gets deeper, more users within an account adopting and more accounts and more customers themselves adopting. The effect is pretty noticeable for all the different cohorts that we have worked with. And what gives us confidence that this is not merely projects being pulled forward is both the breadth and depth of use cases that are coming our way in terms of what people are doing with CoCo and CoWork.
- 反对 CoCo 被用于找出长时间查询和闲置仓库，成本管理是 CoCo top 10 技能之一 (AI 帮客户省 credit)
> “In fact, cost management skills rank among CoCo’s top ten skills.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 管理层称使用 CoCo 的客户也更多使用核心平台 (飞轮)，未给倍数
> “Customers using CoCo also use more of the core platform. That is the flywheel effect we have described.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 反对 成本优化是 CoCo 前 10 大技能之一，账单异常时主动提醒客户
We really, really drive internally this customer-first obsession. And so, people can come in and leave easily. But if you have a customer-first obsession and really focus on the business cases, outcomes and ROI, one of the top 10 skills that we have is around cost optimization. And so, we want people to be fully optimized. We want people to use the product. We want them to get value out of it. And so, we're constantly going back. If we see an anomaly with a customer, we actually go alert them and say, hey, your bill is running higher than it's been before. Were these jobs that you meant to kick off or not? And so, really being obsessed with the customer, I think, really is a key priority for us. Let's stay on the pricing implication here. So I think about architectural enhancements and Snowflake like the Gen2 instances, for instance. You've got this natural pricing deflation in your business like any good technology business. And yet your role is to also abstract value and price one level above the core components of that, so you can capture gross margin.
- 反对 Snowflake 宣称其原生 agent 借语义层可使 token 用量降低 20%-50%，卖点是'不会收到天价账单'——AI 单位消耗被主动压降
在生产环境以及部分客户的数据中，Token使用量实际上降低了20%到50%，因此不用担心突然收到天价账单。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 口径不齐 Snowflake 承认大部分企业目前用外部编码 agent/大模型经 VPC 等狭窄管道连 Snowflake，只把 Snowflake 当存储和数据管道，智能在外部；且外部 agent 会反复找错表、全表扫描、推高成本 (既是外部 agent 拉动计算消费的证据，也是 Snowflake 智能层被绕开的证据)
这里也展示了常见的失败原因。大部分企业目前可能都在使用这种模式：将外部编码agent或大模型通过VPC等狭窄管道连接到Snowflake，仅将Snowflake作为存储和数据管道，而所有智能仍位于外部。这种架构在生产环境中很容易崩溃，因为外部agent不了解企业内部情况，每次都会不断寻找错误的表，进而导致token失控。由于每次query都需要扫描全表数据，成本也会相应增加。但如果有了我们，情况会有所不同。这里可以看到，其实……
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 反对 Snowflake 的 agent 设计目标就是精准定位表、减少无效查询和全表扫描——AI 让单次查询更高效
第3是成本。大家可能会关心，AI如此易用，成本是否也会飙升？实际上不会。因为Agent原生来自Snowflake，所以当你提出“我们的营收是多少”这类问题时，它了解营收的计算方式，会精准找到所需的表，不会像普通方式一样到处询问：“A是数据源吗？这里是数据源吗？”它会直接找到对应的表。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 反对 上下文工程 benchmark：正确率更高同时成本降低约 33% (转写中成本数字前后矛盾)
回答者：也就是说，它会纳入更多信息，因此能更准确地判断你到底要去哪里，不需要反复阅读 instruction，也不需要处理过大的内容。这里展示的是1个 single benchmark。在配合5的情况下，连续测试3次且3次都正确，我们大概可以达到86%，成本约为3毛钱。如果采用另一种方式，大概是72%，成本是0.14。也就是说，我们的成本会降低33%，同时取得更高的结果。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 口径不齐 Snowflake 把'大量 agent 经 LLM 访问 S3 上由 Snowflake 管理的 Iceberg 表'当成治理问题提出——说明 agent 可能在 Snowflake 计算之外读取数据
数据治理一直是 Snowflake 的强项。问题在于，如果数据位于云端，并不属于 Snowflake，而是存储在 S3 中，当外部系统访问由 Snowflake 管理的 Iceberg 表时，会发生什么？在当前 AI 时代，如果大量 Agent 通过 LLM 进行连接和调用，又会发生什么？如果访问控制和数据特有的策略无法跨引擎保持一致，那么数据在任何引擎上都可能因疏忽或运维风险而暴露。随着多引擎架构的发展，单引擎治理也会成为企业的瓶颈。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 支持 CoCo 在 CLI 中按用户有权访问的数据直接构建数据管道——管道建好后持续在 Snowflake 上运行，机制上拉动核心计算 (无量化)
回答者：用户可以在自己熟悉的任何操作界面中使用 Coco，例如 CLI。我在 Coco 的 CLI 中要求它创建一条新管道，并根据它已知我有权访问的数据执行操作。Coco 会进一步询问用户希望创建哪种管道等问题。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 支持 CoCo 可替不会建模的用户构建 ML 模型、设置数据变化时自动重训，并部署到 agent 供业务用户调用——新增持续性的计算负载
回答者：Coco 可以通过机器学习工程构建机器学习模型。即使用户不了解如何构建模型，Coco 也可以根据模型的构建方式完成相关工作。当数据发生变化，或模型随着时间推移变得可用时，用户还可以自动执行相应工作。这些机器学习模型可以部署到不同的智能体中，并通过 Core 使用。Core 是面向业务用户的界面，用户可以通过提问获取洞察。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 支持 企业把新增预算更多投向数据底座，Snowflake/Databricks 增速 20%+ (数字有误)
回 答 者：客户会维持一些较为传统、老旧的软件，重点则放在数据库底座上。可以明显看到，Snowflake和Databricks截至第2季度末的增速都在20%以上。两家的收入规模相近，一家是70亿，另一家是67亿，表现都可圈可点。
来源背景：未具名专家，纪要无正式简介。内容围绕 Agentforce 渗透率、credit 计费、POC 免费 10-20 万 credit、实施项目经验 ，推测为 Salesforce 内部或 Salesforce 实施伙伴的 Agentforce 交付/解决方案人员。视角局限：Salesforce 视角，对 Snowflake 只是顺带；所称 'Snowflake、Databricks 增速 20% 以上、收入 70 亿与 67 亿' 与公开数据不符 (Snowflake 产品收入增速 30%+)，可信度低。 - 支持 BFSI 客户开放 co-work/Cortex agents 给 200+ 网点人员后，该用例每月增加约 3 万美元消费，总额 4 万→7 万美元
随着 Cortex agents 和 co-work 的引入，超过 200 名分支层级的代理获得了访问权限，提升了采用率和 Snowflake 用户数量。该具体用例大约增加了每月 30,000 美元的消费，使总额从 40,000 美元上升到 70,000 美元。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 该用例 PoC 时约 5 千美元/月，生产后按交易量估算约 3 万美元/月 (6 倍)
最初的每月 40,000 美元消费覆盖的是传统的 SQL 查询和 ML 建模。随着 agentic AI、co-work 以及更多用户的访问，过去几个月消费增加了超过 30,000 美元。POC 初期每月成本约为 5,000 美元；随着用例向生产化推进，基于处理的交易量和应用数量，估算的月成本上升到 30,000 美元。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 co-work 让业务用户直接问数据、替代 BI 报表，把 Snowflake 从 DBA 工具扩到业务用户，带来用户数扩张
这大幅提高了 Snowflake 在更广泛用户群体中的采用率。此前，Snowflake 主要是 DBAs 和其他技术用户使用的技术工具。借助 co-work 等工具，业务用户也能访问该平台。访问权限扩展会增加采用率，因为用户会使用 LLM 进行调研、内容生成、应用开发等活动。这是 AI 为 Snowflake 带来的另一种价值形式。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 Coco Desktop 生成的查询经过优化，比 Codex/Claude Desktop 消耗更少 token——AI 让同一任务更省
当客户在没有 Snowflake 数据平台的情况下使用 LLM 时，模型缺乏上下文，无法保留相关关系，也无法理解客户之前与 LLM 的对话。Snowflake 解决了这些限制。将 LLM 与 Snowflake 一起使用还可以降低 token 消耗，因为 Coco Desktop 生成的查询经过优化，消耗的 token 比通过其原生界面（如 Codex 或 Claude Desktop）使用的可比工具要少。这是 Coco Desktop 与第三方 AI 工具之间的关键区别。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 中性 数据平台的存量消费扩张来自更多数据、更低延迟和更多团队/agent 访问
PM: What drives the expansion is 2 different big items. One is expansion of the existing use case or new use cases. In terms of new use cases, it's sky's the limit depends on the company and how much they want to move things to Databricks, so it's quite open. In terms of how you expand an existing use case in terms of consumption is mostly around more data, either more data, or they're going to lower latency so the data consumption, the ETL becomes more real-time that drives more DBU in terms of the cost. And then more access to more teams like dashboards, data sharing, that governance layer can expose the data to more outside parties or even expanded party within the corporation, whether that's business intelligence, data rooms, and nowadays obviously agents. So I think the biggest driver of what expands the existing use case is some of this AI, business intelligence, and again as I said maybe data sharing, and then what drives new use cases depends on the vertical and how much that initial adoption has been successful to be able to go to more use cases. Generally, big enterprises have at least 10 plus use cases running on Databricks, but some even much higher. Hopefully that answers the question.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 自然语言问数最终转化为 SQL 执行，直接推动数据库用量增长 (Databricks SQL 的增长很大程度归功于此)
AI技术简化了数据访问流程，是推动业务增长的关键因素。以往，数据查询需要编写SQL语句，而现在用户可以直接通过自然语言进行提问，例如询问“上周最大的客户是谁”或“上周哪个产品用量增长最快”。这种数据访问的“民主化”使得公司内部任何人都能轻松查询数据，从而增加了用户数量。这些自然语言查询最终仍会转化为SQL语句来执行，直接驱动了数据库用量的增长。以Databricks为例，其Databricks SQL业务的同比大幅增长很大程度上归功于AI带来的便捷查询场景。Snowflake新推出的Cortex产品也吸引了大量新客户，官网数据显示新增客户超过7,000个，其用量也相当可观。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 支持 BFSI 客户引入 Cortex Agents/CoWork、向 200+ 分行人员开放后，月 Snowflake 消费从 $40K 增至 $70K
回 答 者：在 Cortex Agents、Coco 及这些 Co-work 功能推出之前，客户主要进行传统机器学习建模，例如梯度提升预测、回归分析或 ARIMA 预测等。使用传统机器学习模型时，每月消耗约为 40,000 美元，主要由中央 IT 团队驱动，其他分支机构用户无法使用 Snowflake。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 同上案例：开放给 200+ 分行代理人员后增量约 $30K/月，仍在增长
现在，通过 Cortex Agents 和 Co-work，我们已向超过 200 名分支机构代理人员开放权限，提升了 Snowflake 的使用率。与过去相比，目前使用 Snowflake 的用户更多。仅这一特定用例就带来了约 30,000 美元的增量，而且这一数字仍在增长。目前，每月消耗已从 40,000 美元增至 70,000 美元。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 单个 AI 用例从 POC 到生产，月消耗 $5K → $30K (约 6 倍)
最初，概念验证阶段的消耗约为每月 5,000 美元。随着项目逐步进入生产阶段，按照客户希望处理的交易数量和应用数量估算，目前消耗已增至 30,000 美元。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 汽车零部件制造商仅 2-3 个 AI 用例，Snowflake 消费同比 +35-36%，目标翻倍
回 答 者：该客户的 consumption 增长约为35%、36%。目前我们只确定了2、3个用例。客户在这2、3个用例中看到成效后，将会拓展更多用例。我们希望其 consumption 实现翻倍，目前同比增长为35%。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 反对 CoCo 生成的查询经过优化，Token 消耗明显少于客户直接用 Codex/Claude Desktop——AI 让单次调用更省
此外，客户结合Snowflake使用LLM时，消耗的Token数量少于通常的使用方式，因为Coco Desktop生成的查询经过了充分优化。与客户直接通过Codex或Claude Desktop等工具的原生界面使用相比，其Token消耗量明显更少。这就是使用Coco Desktop与使用第三方AI工具之间的区别。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 业务人员不再依赖 Power BI，通过 CoCo 直接提问，Snowflake 用户从 DBA 扩展到业务人员
现在，客户不再需要使用 Power BI 或其他 BI 工具。业务人员可以根据自身权限直接针对数据提出问题，并通过 Snowflake Coco 即时以图形或图表形式获取所需数据。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 中性 Snowflake 内置 MCP 服务器，客户可用自有 IDE/Claude Desktop/VS Code/A2A 连接——外部 agent 也走 Snowflake
Snowflake还内置了模型上下文协议服务器，可以将相关能力开放给Jewel或客户希望使用的其他工具。我们可以将自身作为MCP服务器对外提供服务，客户则可以使用自己的原生IDE，例如Cloud Desktop、VS Code，或通过A2A进行连接。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 支持 采用 AI 产品 (主要是 CoCo) 的客户核心消费更多 (此处未给数字)
And so, when we talk about AI and the performance, it's based on SKUs. And then the core is -- the core business, data warehouse, and storage, analytics, ingestion, what you know is Snowflake as a whole. And so, one of the things that we're seeing that is extremely encouraging and really goes to the variability of the business is we're seeing that when people adopt our AI products primarily CoCo, because we have the most data around CoCo, they're actually consuming more of the core. And so, it may be more personas. And so, you're reaching more people within a business.
- 反对 CoCo 里的成本优化技能是前 10 大技能之一，公司鼓励客户优化
Like -- that's like not sustainable. One of the things that we actually have -- is my mic on? One of the things that we have that is awesome is we have a skill within CoCo to do cost optimization. And it's one of the top 10 skills that's used. And so, we actually encourage and want our customers to do optimization. We encourage and want our customers to use CoCo, but we don't want them to do it in a unhealthy way. And so, part of the sales execution excellence that I talked about with JB and the sales team is really developing those use cases and selling those outcomes.
- 支持 agent 数量增加推动 Snowflake/MongoDB/Datadog 等基础设施与云数据库采购增长 (无量化)
另外，从数据基础设施的角度看，新的 AI Agent 模型带动了相关需求。正如我前面提到的 Datadog，以及 Snowflake、MongoDB 等公司，其基础设施产品和云数据库的采购量都在增长。基础数据规模的扩大，或者说基础数据底座的扩容，正在受到 AI 赋能的推动。
来源背景：未具名专家，推测为 LVMH 集团内部数字化/IT 相关人员 (纪要多次以'集团'自称并举 LV、Tiffany 门店应用为例，,)，职位年限未披露。视角局限：甲方单一集团，Snowflake 采购为统一 AI infra 口径、金额不清；对 Snowflake 无具体用量数据。 - 反对 成本优化技能是前 10 大新技能之一，公司不想让客户对账单感到意外 (避免“定时炸弹”)
And so there's two different things that we're using it for. One of the things that comes up – prior to me getting here, one of the things that the company talked about was optimizations impacting the revenue. And one of the top 10 new skills is a cost optimization skill. And so in a consumption business, it's really – the last thing we want a customer to do is be surprised about how much consumption they have. We want the customer to know exactly how much they're consuming, what they're getting out of it. And we really need to drive that sort of positive ROI for the customer.
- 支持 AI 使用越多，warehouse 节点用量越多；用户与大模型交互不断调整 prompt，比人工精准写 SQL 更费钱
Snowflake与其他厂商合作时，肯定会有一个成本价。Snowflake用户使用ANTHROPIC模型时，Snowflake会向用户收取费用，同时也需要支付给ANTHROPIC等合作方一定的成本。具体来说，Snowflake每收多少million的token，会收多少钱，再支付给ANTHROPIC或微软等多少钱。内部资料并没有明确写出具体比例，所以到底给了百分之多少并不确定，这部分非常机密，内部资料完全没有涉及。现在来看，主要涉及四笔费用。在AI时代，这些费用取决于客户如何使用Snowflake的AI服务。Snowflake的AI主要是Cortex Code和Co-work这两大类。Cortex Code系列，客户用得越多，产生的费用也会越高。如果Code用得很多，AI token的数据量就会上升，收费也会变高。除了token费用高，最常见的使用场景是连接Snowflake数据库，通过大模型在数据库中运行SQL或读取数据。AI用得越多，使用warehouse资源节点也会越多，传统warehouse资源节点的费用也会大幅增加。这两部分费用随着用户使用量增加而增长，Snowflake因此获利更多。借助AI，用户与大模型交互时会不断调整prompt，这意味着资源节点的使用会更多。如果没有AI，用户会更精准地写SQL语句，相对来说会更省钱。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 口径不齐 结构化数据有机增长月环比约 6% (年翻倍)；非结构化数据一次性导入高峰已过，增长趋于平稳
这种增长在初期经历了一个加速阶段，尤其是在Databricks推出向量数据库和Agent等功能时，客户需要一次性地将存量的非结构化数据导入平台，导致数据量激增。目前，这一波导入高峰已过，数据增长速度已趋于平稳。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。
E03 AI 收入的质量
判断：AI 层的收入质量偏低，这一点管理层自己承认了 (毛利率指引下调)。最好的情况是 Snowflake 用开源模型自己做推理、把 AI 毛利拉回来；最坏的情况是 token 继续降价、客户继续换便宜模型，AI 收入一边增长一边通缩。
毛利：CFO 在 Q1 说 AI 产品毛利低于核心平台，75% 的指引是靠 AWS 新合同省下的带宽成本守住的；Q2 把指引降到 74%。Funda 的渠道估 CoCo 毛利率约 52–55%；一位专家说 AI 负载毛利 34%；北欧 SI 说 token 收入约 85–90% 付给模型厂商，"是必备功能而不是利润来源"。偏乐观的是 Funda 的另一位合作伙伴：CoCo 毛利 60% 以上，2–4 个季度可能到 80%，理由是 Snowflake 加价很多 (CoCo 比 Claude Code 标价贵约 40%)。
通缩：Dan Niles 说 token 价格自 5 月以来下降一半，份额转向开源模型；一位专家说中国开源模型只要 Anthropic/OpenAI 的 1/20 且效率相当；CEO 主动把便宜的小模型 (Mistral) 路由进产品；Cortex AI Gateway 的卖点就是按成本自动选模型。管理层把这说成好事 (开源模型自己跑，毛利更好)，但对收入的含义是 AI credit 的单价在往下走。
定价表印证了这一点：按 2026-10 的 credit 消耗表，CoCo 调用 Sonnet 4.6 的 token 价格只比 Anthropic 标价高约 10%，CoWork 高约 30% (见 s22)。Snowflake 在 Q3 FY26 宣布的 Anthropic $2 亿合作，CFO 说明是 Snowflake 向 Anthropic 购买的承诺；OpenAI 的 $2 亿也是同样结构。所以 token 层的毛利主要取决于模型厂商降价有多少被 Snowflake 留下。
管理层的取舍：CEO 在高盛说"单纯转售大模型容量不创造价值，是假装有业务"，他的偏好排序是先卖 CoCo/CoWork，其次卖模型容量，最后才是当 Claude 的数据后端；但他也说如果 CoCo 大规模普及压低毛利，他愿意接受，增长优先于毛利。CFO 把顺序说得更直白：先做产品、再要采用和收入，毛利率之后再说。
回 答 者：Snowflake需要向Anthropic支付大致相当的金额，将其作为转嫁成本，因此Snowflake可能需要将约85%-90%的成本支付给模型提供商，在Token业务上的毛利率并不高。对Snowflake而言，这更多是一项必备功能，因为许多客户目前都在使用Genie、Azure Foundry或AWS Bedrock，希望基于自身的结构化数据和企业业务数据使用大语言模型。因此，Snowflake需要提供大语言模型，帮助客户构建智能体、与数据交互，并基于数据开展智能体应用。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。
1. Model inference is billed by token, at approximately 12 credits per million tokens, plus data warehouse credits for SQL and Python. 2. Now available as an MCP server and Claude Code plugin, improving developer workflows. 3. Customer feedback is positive, but use cases remain unclear. Customers need specific guidance on when to use Cortex Code and when not to. 4. Cortex Code gross margin is estimated at approximately 52–55%, versus approximately 70–75% for Snowflake overall 5. Snowflake has announced a planned acquisition of Natoma, expected to help drive Consumption growth.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。
第一方 agent 的采用也非常可观。许多客户将 Snowflake Coco 用作编码和数据 agent，而不是使用诸如 Acrobat 或 Codex 等工具。Snowflake Coco 在第一季度覆盖了约 1,700 个客户，较去年增加了近 1,000 个账户。在所述具体季度中，超过 2,500 个客户采用了 co-work，体现了 11% 的同比增长。Snowflake 的营收指引也显示，纯粹归因于 AI 的收入有了显著增长，同时与 AI 工作负载相关的毛利率为 34%。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。
+ “Snowflake is marking up Cortex Code aggressively, so I don’t think the gross margin is low.” + “I estimate Cortex Code gross margin is north of 60%, and could reach 80% in 2–4 quarters.” + “Cortex Code is not expensive for Snowflake, but it’s expensive for end customers. If you’re an end customer running queries with Cortex Code, the cost can be punishing. This is the fundamental issue Snowflake needs to address.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。
1. **Databricks CEO Ali Ghodsi on Never Wanting the Job, Beating Snowflake and Why Databricks Stays Private** *Long Strange Trip: CEO to CEO with Brian Halligan, and Ali Ghodsi, co-founder and CEO of Databricks* The last round raised $5B against roughly $20B of demand, the anti-Snowflake playbook (open format, AI, a claimed 3x TCO edge), why he says Databricks would be worth more public, and the one-bottleneck operating model. 2. **Dan Niles: This Isn’t the Top, It’s a Speed Bump** *RiskReversal with Dan Nathan, and Dan Niles, founder of Niles Investment Management* The AI trade’s leading indicators: Texas’s data-center moratorium, the labs’ own slow-down talk, token prices down by half since May with share shifting to open-weight models, a 5% ten-year, and where he still wants to be (Meta, Google, Microsoft, Apple). 3. **Palantir CEO Alex Karp on AI Safety, Liability and Nationalizing the Frontier Labs** *CNBC’s Squawk on the Street with Carl Quintanilla, David Faber and Sara Eisen, and Alex Karp, co-founder and CEO of
来源背景：TMTB Podcast Intel (由 Ethmos 研究 agent 生成的播客摘要，明示)。相关内容为 Databricks CEO Ali Ghodsi 在 Brian Halligan 播客 'Long Strange Trip' 中的说法，以及 Dan Niles (Niles Investment Management 创始人) 的观点。视角局限：只有一句话摘要，正文是图片未读到；Ghodsi 是直接竞争对手，有利益偏向。
此外，行业壁垒也在发生变化。目前，Qwen、DeepSeek等中国模型的成本仅为Anthropic和OpenAI基础模型的1/20，但效率相当。因此，对客户而言，模型之间的差异化正在减弱。很多客户开始转向开放权重模型或开源模型，并自行掌控数据。目前，Anthropic和OpenAI面临的问题在于，数据和模型运行在其他国家。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。
And so we feel very good about being able to do that. Plus there is a lot of innovation that we are driving within these coding agent products themselves. As I said, they handle very complicated tasks, but not everything is complicated. If you want to summarize, for example, I mean I did something a couple of days ago to summarize Slack threads, and perfectly small models from Mistral are enough for that. You don't need the latest and greatest Opus models for summarizing Slack threads. We are building that – those kinds of capabilities natively into Snowflake, so that it can be efficient in what kind of models that it uses. But my short answer to your question is they're creating incredible value, but we are not resting on that. We are creating the controls that one needs in order to keep costs manageable as things continue expanding. Then I'll let Brian take the AI and margin question.
So strategic value of selling inference pass-through into your installed base versus potentially a lower gross margin. How do you think about that trade-off? I mean, first of all, it depends – I hate to start like this, but it depends on what inference is. If it is merely reselling undifferentiated capacity from a large supplier, you're not creating any value. It's like fake news on the part of people that are doing this, trying to pretend that they have a business. On the other hand, if you say, I have a gateway that actually can do a meaningful job of helping my customers optimize spend. In other words, I am creating value on top of my suppliers, and it has some amount of stickiness and redeeming value. That's a meaningful new category.
It's important for us to understand what our strengths are. Our strength is as a data platform, and inference as a component of a modern data platform absolutely makes sense to us. And we also like to sell at the highest value creating point. In other words, my order of preference is always if I can convince a customer to use CoCo or CoWork directly, that's what I want them to be using. If they say, no, all I want is model capacity from you. I'm going to run my own harness. Yes, we will do it. If they say, I want neither of those. I just want the data platform, and it can be a back end to Claude, we will do that as well, somewhat more reluctantly. It's important that you have your Maslow's hierarchy of where you are creating value and acting according to that. And Brian and I are super aligned on what are the business outcomes that we want to drive. Let's say, if CoCo adoption were to go up massively and it has an impact on our gross margins, I'm happy to come and explain that to you all day along.
Yeah, thank you, Brian. Going back to when we develop products, the number one thing is we want to develop a great product. That is the key thing that we want to do. Secondly, we want to make sure that we have massive adoption through use cases and driving benefit to then in turn drive revenue. And then, we'll work on sort of the margin implication. Sridhar and I are very committed to driving overall operating margin leverage in the business. And so, you saw our non-GAAP product gross margin go down to 74% because we've increased our guidance so much. And so, the mix between our AI products and the course changed a little. But we're still committed, as we guided, to increase in our overall operating margin. And so, as we go through and do model choice and use different models, the best thing for us right now is to give our customers the best answer with the best business outcome, and then we'll continue to work on margins as we go forward. But we're committed to driving operating leverage in the model.
Sridhar Ramaswamy: The $100 million is primarily the product suite, but it's the whole stack. Sridhar Ramaswamy: Cortex Search and Cortex Analyst, which are our unstructured and structured data products, respectively. Brian G. Robins: The $200 million is a buy side that we're buying from Anthropic. Sridhar Ramaswamy: partnerships with Anthropic, which include a buy, but also a broader go-to-market motion Brad Reback (Stifel): The op margin obviously down-ticked on the first half sales and marketing investment. Sridhar Ramaswamy: We clearly invested pretty heavily in our sales and marketing teams in the first two quarters Sridhar Ramaswamy: I don't think of this as an either/or. Mike Cikos (Needham): Is it fair to think that 3Q here was just a strong execution quarter Sridhar Ramaswamy: We've consistently told all of you that we view a 3% beat, as a very good beat Sridhar Ramaswamy: I view the Q3 beat as actually still a very solid beat at some 2.5%.
Sanjit K. Singh (Morgan Stanley): congrats on reasserting 30% product revenue growth in Q4. Brian G. Robins: it's really based on the high stable growth that we see in our core business. Brian G. Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type. Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake. Brian G. Robins: we guide based on the Observe customer behavior up until really the point of earnings. Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
展开本节全部 105 条原文 (支持 10 / 反对 45 / 其他 50，按时间)
- 口径不齐 AI 收入 run-rate 达 $1 亿，相对约 $46 亿产品收入只占约 2%
Sridhar Ramaswamy: Remaining performance obligations totaled $7.88 billion, with year-over-year growth accelerating to 37%. Sridhar Ramaswamy: Our net revenue retention remained stable at a very healthy 125%, and we added a record 615 new customers this quarter. Sridhar Ramaswamy: $100 million in AI revenue run rate, achieved one quarter earlier than anticipated, Sridhar Ramaswamy: Because we operate as a consumption-based business, this number reflects real-world enterprise usage. Sridhar Ramaswamy: It's a direct signal of how customers are using our AI capabilities in production to create value today. Sridhar Ramaswamy: In Q3, more than 7,300 accounts are using our AI capabilities every week.
- 口径不齐 $1 亿 AI run-rate 不只是 AI 产品，口径包含整个栈 (可能含 AI 带动的仓库算力)
Sridhar Ramaswamy: we had four, nine-figure deals this quarter. Patrick Colville (Scotiabank): could you just remind us what does that $100 million actually include? Sridhar Ramaswamy: The $100 million is primarily the product suite, but it's the whole stack. Sridhar Ramaswamy: Cortex Search and Cortex Analyst, which are our unstructured and structured data products, respectively. Brian G.
- 反对 Anthropic $2 亿是 Snowflake 向 Anthropic 购买的承诺——token 成本端，指向转售型 AI 收入
Sridhar Ramaswamy: Cortex Search and Cortex Analyst, which are our unstructured and structured data products, respectively. Brian G. Robins: The $200 million is a buy side that we're buying from Anthropic. Sridhar Ramaswamy: partnerships with Anthropic, which include a buy, but also a broader go-to-market motion Brad Reback (Stifel): The op margin obviously down-ticked on the first half sales and marketing investment. Sridhar Ramaswamy: We clearly invested pretty heavily in our sales and marketing teams in the first two quarters Sridhar Ramaswamy: I don't think of this as an either/or.
- 中性 Q3 非 GAAP 产品毛利率 75.9%，全年指引 75%
Brian G. Robins: Our product revenue grew 29% Brian G. Robins: Together, these factors contributed to a stable net retention rate of 125%. Brian G. Robins: Financial services and technology verticals led growth in Q3. Brian G. Robins: Our Global 2000 customers now total 776 Brian G. Robins: each of these accounts spending, on average, $2.3 million on a trailing 12-month basis. Brian G. Robins: We achieved strong booking results, signing four nine-figure deals. Brian G. Robins: This represents a record number of large deals signed in a single quarter. Brian G. Robins: Q3 non-GAAP product gross margins was 75.9%. Brian G. Robins: Non-GAAP operating margin expanded more than 450 basis points year-over-year to 11% Brian G. Robins: Non-GAAP adjusted free cash flow margin was 11%. Brian G. Robins: In Q3, we used $233 million to repurchase 1 million shares at a weighted average price per share of $223.35. Brian G. Robins: We still have $1.3 billion remaining on our original authorization for $4.5 billion through March of 2027. Brian G. Robins: We ended the quarter with $4.4 billion in cash, cash equivalents Brian G. Robins: For Q4, we expect product revenue between $1.195 billion and $1.2 billion, representing a 27% year-over-year growth. Brian G. Robins: We expect non-GAAP operating margin of 7%. Brian G. Robins: We are raising our FY 2026 product revenue guidance. Brian G. Robins: We now expect product revenue of approximately $4.446 billion, representing 28% year-over-year growth. Brian G. Robins: We expect non-GAAP product gross margin of 75%, non-GAAP operating margin of 9%
- 支持 CoCo 用客户合同里已买的同一批 credit 计费，比外部工具单独订阅简单——token 收入被装进 credit
现在，通过在Snowflake内原生使用Cortex Code，客户在合同下已购买的同一批credits也可以用于Cortex Code。从计费和成本角度来说，这对他们来说更为简单易行。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 客户通过 Snowflake 使用 OpenAI/Anthropic 而非直连，每次查询由 Snowflake 管 token、扣 credit
在这种架构中，数据保留在Snowflake内；数据不会离开Snowflake环境。像Snowflake Horizon目录、RBAC和数据掩码这样的治理要求也可以应用于Snowflake中的数据。对于每一次查询，Snowflake管理令牌、消耗Snowflake credits，并且执行仍然在Snowflake内进行。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 OpenAI 合作是直接合作而非经 Azure，多年期
首先值得注意的是，这个合作是直接与 OpenAI 的合作，而不是通过 Azure 或任何中介。据我了解，这是一个多年期的合作，可能会随着时间演进。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 截至 2025Q3，Databricks AI 收入 run-rate 超 $1B，Snowflake 约 $100M，10 倍差距
这是解释一切的核心数字：截至2025年Q3，Databricks AI收入run-rate已超$1B，Snowflake AI收入约$100M——整整10倍的差距。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 中性 2026 年 2 月 Snowflake 与 OpenAI 签 $200M 多年期协议，GPT-5.2 原生嵌入 Cortex AI
**路径一**的优势是最低集成复杂度，代价是切换成本几乎等于重建。**路径二**通过Databricks的多云架构和Bedrock的多模型设计降低了两层的锁定，但数据层仍有数据重力。**路径三**（Snowflake）的差异化在于其跨云部署能力和数据市场（Marketplace）生态——2026年2月Snowflake与OpenAI签署的$200M多年期合作协议，使GPT-5.2等前沿模型原生嵌入Cortex AI，覆盖其12,600家全球客户在三大云上的部署；但Cortex AI的模型层成熟度仍落后于Databricks的Mosaic AI，且$100M vs $1B+的AI收入差距意味着该路径的AI深度尚需时间验证。**路径四**享受GCP在AI模型上的竞争力（Gemini + TPU），但Google在企业级产品上的执行连续性（产品频繁调整、服务中断案例）是行业普遍关注的风险因素。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 反对 AI 产品毛利率低于核心业务，靠核心效率抵消
Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R.
- 反对 靠核心业务效率抵消 AI 低毛利
Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type.
- 中性 与 OpenAI 扩展 $2 亿合作 (与 Anthropic 同量级)
Sridhar Ramaswamy: Snowflake Openflow, now generally available, makes it easier than ever to bring in structured, unstructured Sridhar Ramaswamy: Now generally available, Snowflake Postgres is a world-class operational database Sridhar Ramaswamy: This transforms Snowflake from a system you analyze with into a platform that you build on. Sridhar Ramaswamy: And our recent acquisition of Observe, a market-leading observability platform, Sridhar Ramaswamy: Our deepened partnership with Anthropic is already helping customers like Intercom see significant impact. Sridhar Ramaswamy: We also recently announced a $200 million expanded partnership with OpenAI. Sridhar Ramaswamy: customers now have access to the latest Gemini models natively within Snowflake Sridhar Ramaswamy: AI-powered workflows built on Snowflake with partners like Elementum are replacing the traditional software systems Sridhar Ramaswamy: replacing the traditional software systems used for processes like software license and invoice management. Sridhar Ramaswamy: the company is moving away from legacy infrastructure onto a platform built for scalability Sridhar Ramaswamy: we believe Snowflake is uniquely positioned to become the control plane for the agentic era.
- 中性 FY26 非 GAAP 产品毛利率 75.8%；FY27 指引 75%
Robins: We expect Observe to contribute approximately 1 percentage point of product revenue growth in FY 2027. Brian G. Robins: We expect FY 2027 non-GAAP product gross margin of 75%. Brian G. Robins: and FY 2027 non-GAAP operating margin of 12.5%.
- 反对 管理层承认新 AI 产品利润率不如核心业务，靠核心效率抵消
* Margin触发了上季度的抛售，我们在Preview中特别提到我们不担心Margin的一个原因是观察到SNOW对Customer Success等部门进行了裁员，以及未观察到Memory涨价对Margin的影响 * “我想强调这一点。就在第四季度，我们借助AI获得了很多收益，我们进行了小规模的裁员，公司约200名员工受到了影响。因此，如果你看我们第四季度的净员工增加人数，我们只增加了37人。所以，AI真正改变了投资增长的框架，增长不再与员工人数挂钩。” * "我们的服务交付团队完成客户项目的速度提升了最多5倍，响应准确率提高了超过25%……我们构建了帮助销售人员优先处理客户、自动化研究和生成个性化触达的智能体能力，预计今年将节省相当于90名全职工程师的生产力。” * “我们推出新产品的目标首先是打造出色的产品，其次是让产品易于使用，再次是推动营收增长。达到这些目标之后，我们才会着眼于优化利润率。我们推出了大量新的AI产品，这些产品目前的利润率不如核心业务高，但我们正在通过提升核心业务的效率来抵消这一影响。我们会做出正确的决策来推动增长，并在营业利润率层面实现整体平衡。” * “利润率的改善将同时体现在毛利率层面和公司整体层面……因此，AI在毛利率线以上和营业利润率层面都有很多可以提升的空间。坦率地说，关键在于如何排列优先级，因为机遇实在太多了。”
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 中性 客户关心的是 token 经济而不是数据位置，20% 客户在评估开源模型；数据仍留在 Snowflake，用不用 Cortex 都可以
BJ: That is such a great question because I think it's not associated with Snowflake or Databricks in general. It's just all of them right now. A lot of our clients, especially as they start thinking about agent tech and what that could mean for them across the organization, we are going with the point of view of composable enterprises, and that leads to composable architectures. We need to have, and let's just say the framework, if the framework is data decision delivery of some kind, it's better to decompose your core capability in a way where if the models change, you're able to switch. If... and that has happened so many times in the space, that, you know, the same prompt, same workflow that we had, is no longer valid with the new update of the model. People and our clients are also saying, "The same model is not costing me a lot, same workflow is costing me a lot more with the new model that's been in there, it's more expensive for me to run that, so what do I do?" In that case, we are going... at least 20% of my clients are going into, "Can we do open source now in the same workflow?" So that...
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 数据留在 Snowflake 但模型调用可在 Cortex 之外——Cortex token 收入不是必然
BJ: I believe this is where Snowflake, and I think that's where the advantage goes. And bear in mind, a lot of this is token economics. They're not worried about the data, right? So that still allows Snowflake to keep the data in their environment. And if a Cortex or non-Cortex happens, it's fine, right? But that's where the current concerns are. So composable architecture would be my parting thought here. And good advice will always build that.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 CoCo 价格高，客户主要阻力是成本：一家大金融客户单查询成本达 $5,000，比同等 Claude Code 贵约 33%
+ Cortex Code generates high credit consumption + One large financial-services customer has seen single-query costs as high as $5,000 + A key 2Q26 priority is partnering with Accenture and Deloitte to build solutions that lower the cost + ~33% more expensive than equivalent Claude Code usage * **Cortex Code contributed little in 1Q26 but more in 2Q26**
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 伙伴 C 估计 CoCo 毛利率 60% 以上，2-4 个季度可能到 80%，因为 Snowflake 大幅加价
+ “Snowflake is marking up Cortex Code aggressively, so I don’t think the gross margin is low.” + “I estimate Cortex Code gross margin is north of 60%, and could reach 80% in 2–4 quarters.” + “Cortex Code is not expensive for Snowflake, but it’s expensive for end customers. If you’re an end customer running queries with Cortex Code, the cost can be punishing. This is the fundamental issue Snowflake needs to address.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 反对 伙伴 B 怀疑 CoCo 毛利率可能很低
* “Cortex Code is the fastest-growing product I have ever seen at Snowflake — its trajectory far exceeds Snowpark’s at the same stage.” * “Cortex Code is essentially a pixel-perfect replica of Claude Code’s capabilities, and because Snowflake locked up a batch of GPUs ahead of time, Cortex Code actually runs faster than Claude Code.” * “Snowflake should host the Anthropic model itself rather than calling APIs through AWS Bedrock. Snowflake’s sales team has told us they will do specific inference optimizations.” * “That said, I suspect the gross margin on Cortex Code may be quite low.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 Funda 认为 CoCo 毛利率低于数仓，但因加价、自托管模型、部分 GPU 成本前期已确认，本季毛利率不会受明显冲击
**2.** While Cortex Code’s gross margin is undoubtedly below traditional Data Warehouse, given (i) Cortex Code’s meaningful price markup vs. Claude Code, (ii) Snowflake hosting the model itself rather than relying on AWS Bedrock, and (iii) part of the GPU costs likely already recognized in prior periods, we do not expect a material impact on this quarter’s gross margin.
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 反对 CFO 明确承认 AI 产品毛利率低于核心平台；75% 产品毛利率指引维持是靠 AWS 新合同降低带宽成本等抵消，而非 AI 自身毛利高
Karl, thanks for the question. You're absolutely right. Our AI products have a lower gross margin than our core platform. The one thing that we want to do with our AI products when we launch a new product like CoCo is make sure that we develop a great product, that we get massive adoption, and we've seen really good adoption with CoCo. We're up to roughly about – a little over 7,000 accounts have adopted CoCo. With that said, we're offsetting that and keeping the same product gross margin at 75% for the full year, and lower bandwidth costs, i.e., I talked about the AWS contract, and so we're offsetting it there. So that's how we're able to do that. We're committed to find efficiency to be able to maintain that 75% gross margin.
- 中性 与 AWS 签 5 年 $6B 合同，较 FY2023 合同翻倍以上，换取 AWS 更多 GTM 投入；已计入指引
We have approximately $800 million remaining of our original $4.5 billion repurchase authorization. We ended the quarter with $4.4 billion in cash, cash equivalents, short-term, and long-term investments. During the quarter, we entered into a five-year, $6 billion contract with AWS, more than doubling our prior contract signed in FY 2023. With this agreement, AWS has committed to an expanded go-to-market investment and collaboration. This agreement marks an important milestone in our ongoing partnership with AWS, and its impact is fully incorporated into our outlook. Moving to our outlook.
- 中性 Q1 宣布与 OpenAI 扩大 $200M 合作
Just today, we announced an expanded collaboration with AWS through a new $6 billion multiyear agreement to accelerate enterprise AI adoption globally, leveraging Graviton compute and AI services. The announcement comes as Snowflake surpassed $7 billion in lifetime AWS Marketplace sales, reflecting the growing demand for AI and data workloads running on Snowflake. During the quarter, we also announced an expanding $200 million partnership with OpenAI. And just recently, we brought the joint capabilities from our landmark partnership with SAP to general availability, enabling customers to unite mission-critical business data across their core data systems within our AI Data Cloud. Before I close, I want to acknowledge our Co-Founder and Chief Architect, Benoit Dageville, who will be stepping away from day-to-day operations in mid-June and continuing as a member of Snowflake's board of directors.
- 口径不齐 管理层承认 CoCo 毛利率较低，靠 AWS 合同降低带宽成本，全年产品毛利率维持 75%
“We’re offsetting that through lower bandwidth costs (i.e., the AWS contract I mentioned), and keeping product gross margin flat at 75% for the full year.” Cortex Code does carry a lower gross margin, but only modestly so. This is highly consistent with the key call in our Preview. We argued in our Preview that Cortex Code should be more like a 60%-gross-margin product rather than a 20%-gross-margin wrapper product, because Cortex Code is in fact priced quite high.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 Funda 认为 CoCo 是约 60% 毛利率产品，而非 20% 毛利率的套壳产品，因为定价高
“We’re offsetting that through lower bandwidth costs (i.e., the AWS contract I mentioned), and keeping product gross margin flat at 75% for the full year.” Cortex Code does carry a lower gross margin, but only modestly so. This is highly consistent with the key call in our Preview. We argued in our Preview that Cortex Code should be more like a 60%-gross-margin product rather than a 20%-gross-margin wrapper product, because Cortex Code is in fact priced quite high.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 中性 开源模型变强对 Snowflake 有利：自租 GPU 跑，毛利结构不同于向大模型厂商采购 (暗示现在采购模型的毛利更低)
And we didn't get as much into it, but we spend a fair amount of time making sure that both CoCo and CoWork work effectively across all models and as others. You saw the partnership with Space X. But we also watch the open source models carefully where if their performance rises up, that's actually – it's very positive for us because we rent GPUs from the hyperscalers and we have excellent infrastructure teams that can help us run that at scale. And it obviously produces just different margin profiles than working with large model makers. It's pretty early for all of us. Okay.
- 中性 应对 token 优化浪潮：子 agent 用小模型、Snowflake 托管开源模型、把技能编译成确定性代码 (token 单价通缩会传导到 AI 产品收入)
The thing that's actually executing the plan is so very important. One of the things that our engineers and I co-developed together a few weeks ago was a skill that would generate like a plan for how do you solve a complex problem. And part of what you can do when you do things like that is you can have subagents work with smaller models. And similar to the other point back to our advantage, if there's an open source model that's perfectly great at some job that also happens to be hosted by Snowflake, we can use one of those models. You don't always have to use like the marquee name models for every job. In fact, we make these models available in every model garden and people run a lot of jobs using much smaller models as well. Similarly, I think techniques like skill completion, skill is an English language recipe.
- 中性 CoCo 里可选多家模型，Anthropic 输入/输出 $5/$25，GPT-5.4 $5/$30，开源模型很便宜 (token 成本按模型走)
此外，Cloud Code 只能访问 OpenAI 或 Anthropic 的模型，而 Snowflake Cortex Code 除了 Anthropic 之外，还提供了更多样化的大语言模型选择，包括 GPT 5.4、5.3、Deep Seek 以及来自 Llama、Miral、Mixtral 的模型。客户可以根据愿意支付的成本选择任何大语言模型。目前，Anthropic 模型的输入和输出 token 成本分别为 5 美元和 25 美元，ChatGPT 5.4 的成本是 5 美元和 30 美元，而像 Deep Llama 这样的模型则非常便宜。这为客户提供了根据不同用例选择模型的灵活性。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 反对 CoCo 收入是对 Snowflake 消费的 pass-through；随用量上升、微调和蒸馏小模型，40% 价差会大幅缩小 (token 收入会通缩)
Cortex code as well, right? That is something you have to keep in mind. It's all a pass through for Snowflake consumption is higher. And today perhaps the cost unit might not be 40% more expensive, but as the volumes increase, as we do more and more fine tuning and distilling of smaller models, the difference will come down dramatically. 嗯,Okay.他说而且就是其实现在用这个Cortex Code的用户越来越多了,其实对Snowflake也有好处,因为它之前其实都是一些开发人员去用Snowflake,现在也可以看到很多的这个业务分析师啊等等,就更多的业务这块的人员在用Snowflake,然后在用这个Cortex Code,然后随着使用量的增加的话,就对于整个算力啊,然后存储这块的支出也会增加嘛,所以公司的这个营收的话也会就是水涨船高。呃然后他说这个,而且现在是贵40%,但是随着他们就是更多的进行微调啊等等,就这个成本也会降下来,然后到时候定价的差距也会降下来。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 反对 CoCo 毛利率没有披露，但远低于传统存储/计算约 70% 的水平
回 答 者：Snowflake 并未公布 Cortex Code 的具体毛利率，因此没有明确的公开信息。可以确定的是，其毛利率远低于传统存储和计算业务约 70% 的水平，但具体数值尚不清楚。关于其大致范围，同样也属于非公开信息。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 中性 Snowflake 刚与 AWS 签 60 亿美元协议；客户每在 Snowflake 花 1 美元，会在 CSP 上带来 3-10 美元额外消费
回 答 者：这些云服务提供商（CSP）拥有自己的产品，例如 AWS Redshift、Google BigQuery 以及 Azure Fabric 或 Synapse。然而，它们与 Snowflake 和 Databricks 之间是竞合关系。在竞争层面，虽然 CSP 有自家的数据库产品，但在性能上与 Databricks 和 Snowflake 存在一定差距。Snowflake 和 Databricks 的另一个核心优势是跨云能力。如果一个客户最初使用 AWS Bedrock，后续希望迁移到 GCP，他们无需迁移数据，因为 Snowflake 和 Databricks 在各大云平台均可使用，从而避免了高昂的数据迁出成本和额外的工具迁移费用。同时，CSP 也欢迎 Snowflake 和 Databricks 成为其客户。例如，Snowflake 近期与 AWS 签订了价值 60 亿的订单。这是因为 Snowflake 和 Databricks 的客户越多，所消耗的算力和存储资源就越多，从而增加了 CSP 的收入。因此，双方是典型的竞合关系。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 中性 Claude Code 每位开发者每月约 $200-250；CoCo 可能贵 40%，因为按 token 计费、经 Snowflake 调用 Claude Sonnet/Opus 等高级模型 (说明 CoCo 收入里有转售的 token)
单从纯粹的编码助手功能来看，Claude Code 通常是更便宜的选择。根据使用情况，其部署成本大约在每位开发人员每月 200 到 250 美元之间。而 Cortex Code 的价格可能高出 40%，因为它包含了 Snowflake 的 AI 服务消耗，基于 token 计费，并且其工作负载通过 Snowflake 调用了如 Claude Sona 或 Opus 等高级模型。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 支持 Snowflake 把 CoCo 看作间接收入驱动，而不是直接利润来源 (即使 AI Credits 降价也不改变)
此外，Snowflake 在 4 月（即第二季度初）推出了 AI Credits，这在一定程度上将 AI 定价与版本定价进行了解耦。但即便 Snowflake 通过 AI Credits 机制在一定程度上降低了价格，整体情况也不会改变。Cortex Code 的核心价值在于，它能帮助团队更快地生成数据管道，从而推动整体计算资源的消耗。这才是 Cortex Code 的真正目的：作为一种工具，确保团队提高生产力，进而增加平台消耗。因此，我认为 Snowflake 更多地将其视为一个间接的收入驱动因素，而非直接的利润来源。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 口径不齐 中东客户嫌 Qlik AI 太死板，改用 Snowflake Cortex 作为 AI studio 并自带 LLM——Snowflake 得到工作负载，但模型来源未必经 Snowflake 计费
**Expert:** Uber basic. The Qlik AI story is just for those uber-basic pieces. The challenge is that it's all black box, not glass box. What I mean by that is Databricks, Google, all of these folks have a glass box approach where you can actually go into a Cortex or the Microsoft suite, Azure Studio, to actually configure and tweak and swap out LLMs and build your own story around that, whereas Qlik is, you get what you get. You get your Claude, you get this, you get that, you can't really tweak too many parameters, which has proved not to be palatable for anything apart from really basic use cases. We've had customers in the Middle East, actually, my last customer I spoke to, they wanted this, they wanted that, and they wanted to tweak it and have this. It was just too rigid. They said, "Right, fine. We're just going to use, I think it was Cortex, as the AI studio as part of Snowflake." They brought their LLM and they just effectively built their own piece and they could then own that and then tweak it as they needed.
来源背景：Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 之后重复一遍，以前半部分行号为准。 - 反对 新客户 80-90% 收入来自 AI 产品本身，如 MiniMax 主要通过 AI Gateway 调用多家模型，数据产品贡献很少 (新客户收入以 token 转售为主)
在过去12个月，Databricks的机器学习与GenAI相关产品的年度同比增长率为184%。从存量客户来看，新增收入由数据和AI两大产品线大致平分。AI产品增量最高的来自AgentBricks，该产品帮助客户构建智能体，其收费模式包括调用大模型的Token费用以及构建Agent本身消耗的计算资源（Model Serving）。以太古可口可乐（Swire Coca-Cola）为例，作为AgentBricks用量最大的客户之一，该公司已使用Databricks的数据产品（如ETL、Data Warehouse）三年。其主要应用场景是构建了两个智能体：一个面向内部员工，作为知识库助手（Knowledge Assistant），约有两三百名活跃用户；另一个规模更大，供销售部门使用，通过嵌入在移动应用中的Agent，支持拍照上传、查询库存、销售额及促销活动等功能。仅后一个Agent的Model Serving费用每月就达到约2万美元。为支持这两个智能体，该公司的数据ETL频率从按周批次处理提升至每日一次，使其数据方面的收入也实现翻倍。对于新增客户，其收入的80%至90%来自AI产品本身。这类客户，尤其是数字原生企业（Digital Native Business, DNB），希望快速利用Databricks平台调用各类大模型。一个典型的新增客户是Minimax，作为Databricks全球第三大AI客户，其主要使用AI Gateway和Smart
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 反对 通过 Databricks 调用 Claude/GPT 的价格与直连模型厂商相同，Databricks 利润空间很薄
Databricks与Anthropic、OpenAI之间是典型的竞合关系。在合作层面，Anthropic和OpenAI都是Databricks的战略合作伙伴。客户可以通过Databricks平台直接访问和调用Anthropic的Claude系列模型以及OpenAI的GPT系列模型，调用价格与直接从模型提供商处获取完全相同，Databricks在此过程中的利润空间非常薄。Databricks的价值主张在于，帮助客户将通用的大模型智能与企业自身的私有数据相结合，通过在自有数据上向大模型提问，实现更具体、更符合业务需求的“数据智能”。在竞争层面，三方都在争夺成为客户首选的“可信赖解决方案提供商”（trusted
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 反对 token 转售约 85-90% 成本付给模型提供商，Snowflake 在 token 上毛利很低，是'必备功能'而非利润来源
回 答 者：Snowflake需要向Anthropic支付大致相当的金额，将其作为转嫁成本，因此Snowflake可能需要将约85%-90%的成本支付给模型提供商，在Token业务上的毛利率并不高。对Snowflake而言，这更多是一项必备功能，因为许多客户目前都在使用Genie、Azure Foundry或AWS Bedrock，希望基于自身的结构化数据和企业业务数据使用大语言模型。因此，Snowflake需要提供大语言模型，帮助客户构建智能体、与数据交互，并基于数据开展智能体应用。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 当前模型使用约 70% Anthropic/OpenAI、30% 开源；Snowflake 与 Anthropic/OpenAI 签了约 $200M 协议换取首日接入
回 答 者：目前使用的模型中，70%是Anthropic和OpenAI模型，30%是开放权重模型。主要原因是Anthropic模型目前已部署在美国和欧洲的数据中心，并且均为全托管模式。Snowflake在3或4或5个月前与Anthropic和OpenAI签署了一项价值200 million美元的协议。根据该协议，Snowflake可以在Anthropic和OpenAI的所有领先大语言模型发布的第1天立即接入。模型一旦面向公众发布，就会立即在Snowflake中上线，使客户能够使用最新模型和相关创新。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 反对 客户 POC 后因 ROI 不足从昂贵前沿模型转向 Kimi/Moonshot 等开源模型 (token 单价通缩)
不过，在完成初步试点和概念验证后，许多客户发现，使用OpenAI、Opus 4.8或GPT 5.4等基础模型并未带来足够价值，主要原因在于成本较高。客户在这些昂贵的大语言模型上投入数千美元，但未获得相应的投资回报。与此同时，他们发现Moonshot和Kimi等开放权重模型能够以低得多的价格，提供相同的质量、输出和准确率。因此，许多客户也开始在Snowflake上转向这些开放权重模型。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 目前 AI 应用规模不大，模型调用还不是主要成本因素；5-8 年后可能增长 100-1000 倍时才会精细优化模型成本
这项功能的重要性会不断提升，但不会带来根本性改变，因为在大多数情况下，底层模型并不是限制因素。目前相关应用的规模还不够大，模型使用量尚未成为主要成本因素。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 Cortex Code 有成本控制，单用户成本不超过 100 美元 (仅见于摘要稿)
Snowflake 的一个策略是，与其让客户为 Anthropic 或 OpenAI 购买基于用户的许可，不如他们直接购买 Snowflake 积分，并给每位员工分配 Cortex Code Desktop 的访问权限，允许任何人提出任何问题。我们也有成本控制措施以确保单用户成本不会超过 100 美元，从而实现成本管理。
来源背景：Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。 - 反对 Cortex Code 下拉列表有 20+ 模型，token 是 pass-through：客户消耗 $1,000 token，约 85-90% 付给模型商
A:是的。如果你使用过 Cortex Code 或看过演示，就会注意到有一个下拉列表，里面有 20+ 可供使用的 LLM。优势在于 token 成本对 Snowflake 来说是过账（pass-through）。如果客户在 Snowflake 下消耗了 1,000 美元的 tokens，Snowflake 大约需要将其中 85%–90% 支付给模型提供商，因此 token 的毛利并不高。
来源背景：Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。 - 中性 受监管客户只用闭源模型，Cortex Code 底层更多用 Azure OpenAI，不考虑开源
我们公司目前更多使用 Azure OpenAI，因为从企业角度看，它在安全性和治理方面更有保障。未来也可能开发和使用一些开源模型，但目前仍处于较早阶段。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 约 25% 客户以某种形式采用开源模型；开源厂商开始建渠道
回 答 者：开源模型的采用率毫无疑问正在提升，原因在于其采用较为便捷。从今年的一些统计数据来看，大约有 25% 的客户以某种形式采用了开源模型。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 支持 直接的 Cortex 消费今天仍相对较小，CoCo 拉动的间接消费在战略上更重要 (支持第 1 层是主体)
这就是 AI 在扩大工作负载和组织内用户基数方面的原因。我想强调的是，直接的 Cortex 消费在今天仍相对较小。Cortex code 导致的消费在战略上比直接消费更为重要。对于那些重度采用 Cortex Code 的客户，我估计其消费同比增长约为 45%，而对于使用传统 Snowflake 工作负载的可比客户，增长可能大约为 30%。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 反对 开源模型压缩了 Cortex 专有推理的经济性，客户没那么多理由为 LLM 访问付溢价
大型客户已经在朝这个方向迁移，因为开源模型可以降低推理成本并改善数据控制。对 Snowflake 来说，我看到两个影响。负面是开源模型对通过 Cortex 提供的专有推理的经济性带来压力，客户没有那么多理由为访问 LLM 支付溢价。但 Snowflake 的回应是变得更加模型不可知：Cortex 支持像 Llama 这样的开源模型并与专有模型并行，Snowflake 也对这些由 Llama 驱动的模型进行了优化。积极方面是，更便宜的模型可能意味着更多的 AI 使用。例如，如果推理成本下降 70%，客户不一定会减少他们的整体 AI 预算——他们可以部署更多代理并处理更多工作负载，同时仍需要对这些代理和数据进行治理。这就是为什么 Snowflake 在模型层面可能失去一些经济利益，但可以在数据和计算层面获得显著的消费增长。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 模型服务来自第三方，Unity AI Gateway 可把简单查询路由到开源模型，但目前仍全部路由到大模型
我们有一款名为Unity AI Gateway的产品，客户可以通过它集中管理所有LLM使用和模型服务，但模型始终来自第三方。客户通过Databricks使用这些服务的优势在于，相关使用也可以纳入治理体系，并且能够设置限制、进行监控，最终还可以调整查询路由。如果查询难度不高，可以将其路由至开源模型。目前，查询仍然全部路由至规模更大、成本更高的模型。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 反对 Foundation Model API 利润很低，需支付 Google/OpenAI/Anthropic 费用
AI爆发阶段核心产品的收入和增长，增长最快的是Agent Bricks，今年（2026年）Q1刚开始收费，所以有爆发式增长。第二个爆发式增长的是Foundation Model API。这一部分利润很低，因为要付给Google、OpenAI和Anthropic费用。第三大产品是Databricks Apps。到了Q2环比增长不高。另外，稳定阶段主要贡献的产品，包括是vector search， batch inference和knowledge assistant，客户上传文档后Databricks立刻build一个bot，客户不用自己读文档，直接问bot公司内部SOP、HR政策等。现在主要给客户卖Agent Bricks和apps，客户如果想用foundation API会主动找Databricks，Databricks帮客户build agent，卖Agent Bricks。如果客户有更高要求，比如agent要贴合企业需求，有统一的app look and fill，还要做scale down、scale in，就会包装成Databricks Apps。背后会调用Lakebase存context。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 反对 token 收费与客户直购第三方模型价格一致，只靠规模折扣赚差价；smart router、token 优化、GenAI 规则免费
基本上还是和 Databricks有关系，并不是单纯转移到客户手上。客户是通过 Unity AI gateway 调用这些 API。Databricks提供的服务中间有一个 smart router，也就是智能分流，比如复杂的推理问题会用 Claude Code，做 summary 用 DeepSeek，多模态生成图片用 GPT，数学问题用 Claude Code 或 GLM5.2，Databricks会帮客户做分流。目前 foundation API 用量最大的是 Gemini 3.2 FLASH，因为在 Databricks这边，Databricks的分流策略有 token optimization，Gemini 3.2 FLASH 是做批量 token 优化，优化比最高。用户的用法不是简单调用 API 后返回原系统，一般是在Databricks这边搭建 agent，有一步是问大模型，并且用 Databricks提供的 GenAI功能，在 GenAI里定义规则，比如公司不能谈竞争对手产品优劣，回复消息不能包含公司敏感信息，包括领导名字等，客户会定义 GenAI规则放在 Databricks上，Databricks提供 GenAI、smart router 和 token optimization 都是免费的。Databricks收费的项目是按 token 收费，和客户直接购买第三方模型价格一致。Databricks有规模优势，会拿到折扣，从中赚取差价。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 反对 Databricks Foundation API 用量最大的是 Gemini Flash，因 token 优化比最高 (路由到便宜模型，token 收入通缩)
基本上还是和 Databricks有关系，并不是单纯转移到客户手上。客户是通过 Unity AI gateway 调用这些 API。Databricks提供的服务中间有一个 smart router，也就是智能分流，比如复杂的推理问题会用 Claude Code，做 summary 用 DeepSeek，多模态生成图片用 GPT，数学问题用 Claude Code 或 GLM5.2，Databricks会帮客户做分流。目前 foundation API 用量最大的是 Gemini 3.2 FLASH，因为在 Databricks这边，Databricks的分流策略有 token optimization，Gemini 3.2 FLASH 是做批量 token 优化，优化比最高。用户的用法不是简单调用 API 后返回原系统，一般是在Databricks这边搭建 agent，有一步是问大模型，并且用 Databricks提供的 GenAI功能，在 GenAI里定义规则，比如公司不能谈竞争对手产品优劣，回复消息不能包含公司敏感信息，包括领导名字等，客户会定义 GenAI规则放在 Databricks上，Databricks提供 GenAI、smart router 和 token optimization 都是免费的。Databricks收费的项目是按 token 收费，和客户直接购买第三方模型价格一致。Databricks有规模优势，会拿到折扣，从中赚取差价。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 口径不齐 专家称 CoCo + CoWork 合计约占总消耗 15%——远高于 2 月专家 (<5%) 和 8-26 专家 (狭义 AI 负载 15-20% 但含核心计算) 的口径，很可能把 AI 拉动的仓库消费也算进去
Many of our customers, um, um, adopting snow Snowpark is increasing, you know, that is around, um, I would say another 15 to 20%. And then, you know, so data warehouse, Snowpark, these are all, um, natural workload expansions, more applications coming on board, things like that, right? Now, Cortex Code and Cortex co-work, um, are the ones that are accelerating, but they're still small portion of the consumption, right? So I would say between, you know, Co-Co and co-work, um, it is probably around, you know, 15% or so. Um, but we expect that to increase as many of our customers start, you know, adopting Co-Co, um, and then, you know, as many of our customers start, you know, using more queries and things like that. Um, we expect um, the consumption for for Co-Co and co-work, um, to start increasing over the next couple of quarters.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 中性 CoCo 有按用户的基础许可，但真正的支出来自查询次数、多模型调用、多步骤流程带来的 token/计算消耗——即收入中有 token 转售成分
And also, you know, Coco is only going to be used by a a set of developers, right? Um so if there are, you know, um 500, 600 named users that are using um Snowflake, probably only 30, 40% of them uh are going to use Coco. So there is a base um license, user-based license for Coco, but the real consumption is the more number of queries they make, and also if the queries are using multiple models, if the queries have um, you know, multi-step uh processes so that uh token or consumption increases, that is where uh customers are spending. So I would say for those customers that have um meaningful, you know, uh investments already, um consumption, at least in my view, is is going up somewhere around, you know, 15 to 20% right now. It's not, it's not still, you know, very big, but it's starting to, you know, um most of the incremental spend um is is coming on um on Coco and Co-work.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 中性 Snowflake 除 Anthropic/OpenAI 外也支持自带模型和开源模型，并支持 RAG、A2A、MCP 集成
A:Snowflake 起初主要在 AWS 上得到支持，现在也可在 Azure 和 GCP 上使用，因此随着许多客户转向多云，Snowflake 可在三大超大云上被访问。除了 Anthropic 和 OpenAI 等基础模型外，Snowflake 也支持自带模型和开源模型。通过支持更多模型，Snowflake 正在支持检索增强生成、agent-to-agent 协议以及 MCP 服务器集成。所有这些需求都在增加，Snowflake 将从中受益。
来源背景：标题'北美软件渠道商离职专家'；与 2026-08-19 Merit 音频转写为同一场访谈，身份局限见该条 - 反对 专家认为单独的 AI credit 主要是 Snowflake 为了控制自身成本，客户更想要固定费率
BJ: I think that I understand where it's coming from, but it's more, respectfully, where I'm sitting. I think it's more to contain their own costs. My buyers would, my clients would rather have a flat rate.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 模型路由只是标配，不会是高毛利网关业务
BJ: At one point, I thought the world only belonged to OpenAI... we're joking... but the reality is that all the best models and the workflows and the costing, it just made a lot of sense, right? Then Google and Google Gemini was also doing a really good job for the knowledge worker. Databricks and Snowflake were powering that infrastructure, but now we're seeing a lot more, you know, when it comes to agent and agent governance, it's becoming more neutral and it's more designed because the costing of the most expensive Ferrari model is just too much, right? Even if you have access to it, if you can't use it, what's the purpose? Because it's just too expensive. And that's where CFOs and the mandate, and I do think that's where we're going to see a lot more innovation. And this routing alone will be just a key checkmark. It's not going to be the main... I don't think it's a high margin gateway offering by any means. But what you learn from it and the patterns might be where we're going to see a lot more solidified growth, more domain specific growth. I say this because Snowflake, for instance, has been investing a lot because they didn't have the application layer, right? Databricks by default was the choice. It's a choice for data science folks.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 客户在不太复杂的工作流上积极使用开源模型 (含中国模型、Nemotron)，推动 token 单价下行
BJ: Great question. And I think we are all moving towards a future where it's multi-model environments, right? And a lot of weight towards the open source models. I think absolutely the Chinese labs have proven that they can do it more efficiently, of course, distilling the outputs from the larger models as well to make that happen. But at the same time, the aggressiveness and how they're sharing it in open source and MIT license accordingly does encourage at least the naysayers to pay attention to the models and not see them as a threat, right? Which has led to even adoption of open source models from Meta and NVIDIA as well. NVIDIA's own Nemotron, and Nemotron is also something in consideration as well. My clients are actively leveraging it for not so complex workflows and it's working effectively and I think we're going to see a lot more of that. I think both Snowflake and Databricks will benefit from that because at the end of the day, the agent needs an agent-ready data layer and I think that's where there's a lot of work to be done. We're not looking at just dashboards anymore; we have to act on that dashboard and there's a lot of work and growth expected.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 中性 每替代 1 名 FTE 约产生 1 万欧元 AI OPEX，其中 20-25% (约 2,000-2,500 欧元) 流向 Snowflake/MongoDB 等数据平台
回 答 者：坦率地说，我没有完整的数据，但可以提供一个估算。按照每名FTE约10,000欧元的AI运营支出计算，其中约20%至25%可能用于Snowflake、MongoDB等数据仓库相关服务。具体比例会因企业的架构配置而略有不同，但这可以作为一个有参考价值的大致范围。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 AI OPEX 所有组成部分都会被审视，开源模型使用会增加——模型降价压低 token 转售收入
回 答 者：成本以及模型实验室的成本都受到严格审视，因此开源模型的使用可能会逐渐增加，企业在选择底层模型时也会更加审慎。基本上，AI运营支出的所有组成部分都会受到更严格的审视。随着业务扩张，这将成为一项持续多年的成本控制工作，而不是一次性举措。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 口径不齐 Snowflake 将让 CoWork/CoCo 做模型路由降低客户消耗；Merit 认为间接消费大于直接收费——即 token 层通缩、靠核心层补
未来，Snowflake将尽可能让cowork和Codex Code选择模型并进行不同路由，从而降低客户在相关场景中的消耗。我们认为，这对整个S并非负面影响，而是较大的正面影响。Codex Code为Snowflake带来的间接消费将大于直接收费，因为它缩短了软件开发周期，并扩大了用户范围。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 反对 CoCo 按 token 计费约每百万 token 12 credits，另加数仓 credit；毛利率估约 52-55%，低于公司整体 70-75%
1. Model inference is billed by token, at approximately 12 credits per million tokens, plus data warehouse credits for SQL and Python. 2. Now available as an MCP server and Claude Code plugin, improving developer workflows. 3. Customer feedback is positive, but use cases remain unclear. Customers need specific guidance on when to use Cortex Code and when not to. 4. Cortex Code gross margin is estimated at approximately 52–55%, versus approximately 70–75% for Snowflake overall 5. Snowflake has announced a planned acquisition of Natoma, expected to help drive Consumption growth.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 CoCo 毛利率约 52-55%
1. Model inference is billed by token, at approximately 12 credits per million tokens, plus data warehouse credits for SQL and Python. 2. Now available as an MCP server and Claude Code plugin, improving developer workflows. 3. Customer feedback is positive, but use cases remain unclear. Customers need specific guidance on when to use Cortex Code and when not to. 4. Cortex Code gross margin is estimated at approximately 52–55%, versus approximately 70–75% for Snowflake overall 5. Snowflake has announced a planned acquisition of Natoma, expected to help drive Consumption growth.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 多数消费来自流量和 token 而非席位；token 消费占比可能在 3-4 个季度从 10% 升到 30-40% (意味着第二层转售收入成为增长主体)
1. For example, within a given enterprise, only some users may use Snowflake, and an even smaller share may use Cortex Code. More users may use Snowflake CoWork, while fewer may use Cortex Code. However, because Cortex Code runs queries, token consumption per user may be higher. Cortex Code may therefore have fewer seats but higher usage as query volume increases. This is a hypothetical example, not actual customer data. 2. General users may run more queries and have more seats, but data application use cases consume more tokens. Even with fewer Cortex Code users, token consumption remains healthy. Most consumption comes from traffic and tokens, not seats. If token consumption is currently 10%, it could rise to 30% to 40% over the next three to four quarters as customers expand usage. This is directional, not a precise forecast or company guidance. 3. Is It Still a Data-Environment Use Case?
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 token 成本会受采购和财务部门更严格审查，类似 5-10 年前的云支出治理
1. Enterprises are paying more attention to AI agent usage and token spending, which is already a common issue for CoCo, OpenAI, GitHub Copilot, and Claude Code. 2. Within Snowflake, the additional value from being tied to the ecosystem supports pricing slightly above market levels, but outside Snowflake, customers may also connect directly to model providers. 3. As usage grows, token costs will face greater scrutiny, even if the impact doesn’t necessarily show up next quarter. Procurement and finance teams are introducing stricter cost tracking and savings measures, similar to how enterprises managed cloud spending 5-10 years ago. This cost-structure shift could be larger because agents can take on work previously done by full-time employees. 2. More customers are beginning to consolidate data into Snowflake
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 产品毛利率指引从 75% 降到 74%，原因是低贡献利润率的 AI 工作负载收入占比上升
In Q3, we expect product revenue between $1.588 billion and $1.593 billion, representing 37% to 38% year-over-year growth. Turning to margins. For FY 2027, we now expect 74% non-GAAP product gross margin. This revised outlook includes a higher revenue mix from fast growing AI workloads, which carry a lower contribution margin today. We're delivering continued operating margin expansion as we offset growing cloud costs with slowing head count expense. We are increasing our FY 2027 non-GAAP operating margin guidance from 13.5% to 14.5%.
- 反对 CFO 把顺序说得很清楚：先做产品、再要采用和收入，毛利率之后再说；指引上调太多导致 AI/核心组合变化、毛利率降
Going back to when we develop products, the number one thing is we want to develop a great product. That is the key thing that we want to do. Secondly, we want to make sure that we have massive adoption through use cases and driving benefit to then in turn drive revenue. And then, we'll work on sort of the margin implication. Sridhar and I are very committed to driving overall operating margin leverage in the business. And so, you saw our non-GAAP product gross margin go down to 74% because we've increased our guidance so much. And so, the mix between our AI products and the course changed a little. But we're still committed, as we guided, to increase in our overall operating margin. And so, as we go through and do model choice and use different models, the best thing for us right now is to give our customers the best answer with the best business outcome, and then we'll continue to work on margins as we go forward.
- 口径不齐 开源模型由 Snowflake 自己做推理，经济性不同；'auto' 模式自动匹配模型，给了优化空间——毛利改善潜力但也意味着 token 单价降
And this is also where open source models come in. There's obviously been several generations of these open source models, and we support many of them within Snowflake. And yes, we have pretty different economics when it comes to open source models, since we run the inference ourselves. So that offers a lot of potential for future optimizations. And within our harnesses, many of the requests that we get from customers come in this mode that we call auto, where we can pair up the task with the model that is most appropriate for that particular task. And that gives us a lot of leeway in being able to optimize tasks for our customers. Yeah.
- 中性 前沿模型厂商一直在降价，使成本相对用量可控——token 通缩是现实
And so, as we go through and do model choice and use different models, the best thing for us right now is to give our customers the best answer with the best business outcome, and then we'll continue to work on margins as we go forward. But we're committed to driving operating leverage in the model. And one more thing on this one, Karl, which is even the frontier models have been revising prices down on a regular basis and have been introducing additional models to their families which have kept costs somewhat in check relative to the usage of organizations. Thank you. Thank you.
- 中性 客户越来越想在模型间切换以优化成本，模型中立被视为优势
And all of you know this at least as far as the frontier labs go, there used to be somewhat of a dichotomy where Anthropic was available extensively on AWS while the OpenAI models tended to be more on Azure. A material change that's happened is that both the companies are deploying substantial capacity of their own, but it's also the case that they are available in other clouds than the ones that they started with. And we're absolutely seeing a lot of interest in being able to switch between different models and also to optimize cost. And this is also where open source models come in. There's obviously been several generations of these open source models, and we support many of them within Snowflake.
- 反对 FY27 产品毛利率指引 74%，因 AI 工作负载占比上升且贡献利润率较低
> “This revised outlook incorporates the rising mix of rapidly growing AI workload revenue, which currently carries lower contribution margins.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 反对 管理层优先产品和采用，之后再考虑利润率
> “We want to ensure broad adoption across use cases and deliver real value that drives revenue growth. Then we will consider the margin implications.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 口径不齐 抵消因素：模型中立、自动路由、开源模型自托管、前沿模型持续降价 (也意味着 token 收入随模型降价通缩)
> “Even frontier models have continued to decline in price. This has helped keep costs under control relative to organizations’ usage.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 中性 Snowflake 不训练前沿模型，保持模型中立，Arctic 模型用于受限任务
> “We have not trained models to enter the frontier-model space.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 CEO：单纯转售大模型容量不创造价值，是假装有业务
How do you think about that trade-off? I mean, first of all, it depends – I hate to start like this, but it depends on what inference is. If it is merely reselling undifferentiated capacity from a large supplier, you're not creating any value. It's like fake news on the part of people that are doing this, trying to pretend that they have a business. On the other hand, if you say, I have a gateway that actually can do a meaningful job of helping my customers optimize spend. In other words, I am creating value on top of my suppliers, and it has some amount of stickiness and redeeming value.
- 中性 同时向 OpenAI 和 Anthropic 采购容量，并能自己跑开源权重模型
That's a meaningful new category. And so we look at inference, for example, in the context of, can I offer my customers choice? Because we buy capacity in both – both from OpenAI and from Anthropic, and we have the capacity to do things like run open weight models ourselves. In that context, inference becomes a little more interesting. We also generally take the lens off.
- 口径不齐 价值层级：最想卖 CoCo/CoWork，其次卖模型容量，最后勉强只做 Claude 的数据后端
Our strength is as a data platform, and inference as a component of a modern data platform absolutely makes sense to us. And we also like to sell at the highest value creating point. In other words, my order of preference is always if I can convince a customer to use CoCo or CoWork directly, that's what I want them to be using. If they say, no, all I want is model capacity from you. I'm going to run my own harness. Yes, we will do it. If they say, I want neither of those. I just want the data platform, and it can be a back end to Claude, we will do that as well, somewhat more reluctantly. It's important that you have your Maslow's hierarchy of where you are creating value and acting according to that. And Brian and I are super aligned on what are the business outcomes that we want to drive.
- 反对 若 CoCo 大规模普及压低毛利，CEO 愿意接受并解释：增长优先于毛利
It's important that you have your Maslow's hierarchy of where you are creating value and acting according to that. And Brian and I are super aligned on what are the business outcomes that we want to drive. Let's say, if CoCo adoption were to go up massively and it has an impact on our gross margins, I'm happy to come and explain that to you all day along. That's not an issue because it will drive a meaningful acceleration in our overall business. And I can also tell you, as we grow and scale, as we do these things, we get better at optimizing. We get better at running open source models, which will have much better margins.
- 支持 改善 AI 毛利的路径：规模化优化、开源模型、批量采购 (AWS $6B 合同)
That's not an issue because it will drive a meaningful acceleration in our overall business. And I can also tell you, as we grow and scale, as we do these things, we get better at optimizing. We get better at running open source models, which will have much better margins. We obviously buy a bigger quantity from the suppliers like we do with AWS and the $6 billion contract, which gives us better economics. There are good answers to things like gross margin, but it needs to make strategic sense. What I have little appetite for is being a blind reseller of someone else's intelligence.
- 口径不齐 被问 CoCo 毛利护栏时，CFO 没给数字，只说会提前一两个季度沟通
We also know, although we're bringing that down and we also build models, understand what this consumption is going to be. And so, it's really Sridhar and I don't want to surprise folks. And so, if we were seeing that massive CoCo adoption beyond what we're already seeing, we would have the ability to communicate that to you within a quarter or two to manage that. I want to ask a technical question on agents. So, Sridhar, there is a school of thought that the systems and architectures of today are not going to scale for real-time agents because agents have so much more volume, for lack of a better word.
- 支持 Snowflake 自训专用小模型 (Text-to-SQL 8B) 在 Text-to-SQL 上精度超过更大的通用模型——意味着部分 AI 推理可由自有低成本模型承担，不必全部转售前沿模型 token
Snowflake 的 Text-to-SQL 在该 task 的测试中，Text-to-SQL8B 在我们使用的对比中只是一个大约 32B 的模型，但与其他很多大模型相比，能够取得更高的 accuracy。同时，它的模型大概只有 300万参数，30，小于 30亿参数。当然，这并不意味着 scale 不重要。Scale 非常重要，但模型本身或模型大小本身并不能解决所有 domain data 问题。当 evaluation、dataset 和实际使用能够真正反映 workload 时，专门的数据和训练方法可以把一个最小的模型转变为更强的企业级应用。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 支持 SQL 自动补全已上线，由 3B 模型后训练成 4B 专用模型，precision、接受率提升、时延降 30%——自有模型承担高频轻量任务
在该 task 上，我们最初的模型是 1个 3B、3B 的模型，之后通过后续训练获得了 1个 4B、更 customized 或 specialized 的模型。使用这个更 specialized 的模型后，我们提升了所有指标。在 1个内部 precision test 中，我们看到我们的……
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 口径不齐 CoWork 背后同时包含'市面上最顶尖的 AI 模型'(第三方前沿模型)，Snowflake 强调价值在企业上下文而非模型本身——没有说明 token 如何计费、谁付模型成本
最重要的是底层的上下文。AI 大模型非常聪明，我们也包含了市面上最顶尖的 AI 模型。但是，无论 AI 模型多么聪明，如果脱离企业上下文，就无法发挥其 100% 的能力。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 口径不齐 CoCo 模型默认 auto，由 Snowflake 为用例选择模型，也可手选——Snowflake 控制模型路由，可借此调节 token 成本/毛利，但未披露底层模型和计费
Here's a great example. You can pick and choose the model you want for your use case. Uh, by default it takes auto, which is what we use to select the best model that is suitable for the use case from the application we are building. How are we doing on the time? Good? Should we give them more time to work on the projects? Um, let me see, I want to see if uh, the streaming data is coming in. I don't want to just go ahead and visualize without actually visualizing it on my stream. We also announced the Google data stream, uh, a few months ago, a couple of months ago, that actually takes away all the technical details.
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 口径不齐 AWS 讲者称 Snowflake 客户中约 70% 与 AWS 相关，亚太 75%——Snowflake 对 AWS 的依赖度高 (主语在转写中缺失)
先介绍几个数据。目前，在Snowflake的客户群中，相关客户占比为70%，亚太区的比例更高，达到75%。另一个数据是，双方相关客户在marketing等方面，相关数据从原来的31亿，以及从85亿提高到96亿。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 反对 专家称 AI 工作负载相关毛利率为 34%，远低于产品整体 ~75%
第一方 agent 的采用也非常可观。许多客户将 Snowflake Coco 用作编码和数据 agent，而不是使用诸如 Acrobat 或 Codex 等工具。Snowflake Coco 在第一季度覆盖了约 1,700 个客户，较去年增加了近 1,000 个账户。在所述具体季度中，超过 2,500 个客户采用了 co-work，体现了 11% 的同比增长。Snowflake 的营收指引也显示，纯粹归因于 AI 的收入有了显著增长，同时与 AI 工作负载相关的毛利率为 34%。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 中性 Snowflake 与 Anthropic 和 OpenAI 都有大规模独家合同，新模型几天内上架
A:Snowflake 对模型持中性态度，并通过单一平台为客户提供他们想要的前沿模型访问。Coco Desktop 通过单一界面和接口将 Anthropic、Gemini 和 OpenAI 的模型集中接入，这是一个关键差异化点。Snowflake 与 Anthropic 和 OpenAI 均有大规模的独家合同，这有助于确保公开发布的模型能在几天内提供给 Snowflake 客户。客户可以在 Snowflake 的上下文数据之上使用他们偏好的任何模型。Coco Desktop 目前提供包括最新的 GPT、Anthropic 和 Gemini 模型在内的 12 多种模型。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 中性 客户不必单独买 Anthropic/OpenAI 订阅，在 Snowflake 内访问模型——即 token 经 Snowflake 转售
客户可以不用单独购买 Anthropic 和 OpenAI 的订阅，而是在 Snowflake 的平台内访问相关模型。Snowflake 的价值在于提供数据上下文、数据基础以及多表和多对象之间的关系。客户可以使用任意 LLM 来查询这些数据。如果 GPT-6 在几天前才发布，它也应能在很短时间内出现在 Snowflake Cortex Desktop 中，使客户能够将其与 Snowflake 数据一起使用。这就是 Snowflake 的主要商业主张。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 token 成本是客户主要关切，客户采用自动路由把任务分给 GLM、Qwen、DeepSeek 和小模型——压低每任务 token 收入
token 成本是客户的主要关切之一。使用 GPT-6 或 Claude Opus 5 可能导致成本大幅上升。因此客户正在采用 Snowflake 的自动路由功能，为每个用例选择合适的模型。可选模型还包括 GLM 5.2、Qwen、DeepSeek 以及更小的语言模型。客户希望为每个用例使用最合适的 LLM，而不是对所有事情都使用 Opus 或 GPT-6。自动路由有助于为每项任务提供最适合的模型访问。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 中国开源模型成本约为 Anthropic/OpenAI 的 1/28，侵蚀基础模型差异化
A:Anthropic 和 OpenAI 目前处于亏损状态，因为它们的推理收入低于推理支出。它们在训练模型上花费了数十亿美元，而 GPU、数据中心和推理成本持续上升。像 Qwen、DeepSeek 和其他中国模型的成本仅为可比 Anthropic 和 OpenAI 基础模型的大约 1/28，但提供了类似的效率。这正在侵蚀基础模型提供方的差异化优势。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 两家都用 Anthropic/OpenAI 第三方端点，推理方面没有差异
database quite fast and governed pretty easily. And I think that's another advantage that Databricks had. In terms of... And early on, I think Databricks built their own model that I think under the purchase of MosaicML definitely helped. But then Mosaic, I think, focused more on the training and for inference, I think the partnership mattered. And I think in that front, they're pretty similar, I would say. How much Databricks or Snowflake go buy their own GPUs and serve open source models, I don't know the contrast there, but they both obviously work with third-party Anthropic and OpenAI. In that regard, I don't think there's any difference. They just use their endpoints. When it comes to fine-tuning and some of the training stuff, Databricks has some innovation around reinforcement learning, but that's not much of a revenue driver, but it's definitely a driver for credibility and the Mosaic AI team has built lots of capabilities around applying reinforcement learning and supervised fine-tuning post-training stuff to your data. I don't think there's much of an adoption there, but I think even that front for some advanced digital native companies, Databricks' AI is a bit more attractive and easier to access. The areas that are pretty common is like batch inference, like AI functions to SQL; they're pretty identical, maybe governed a bit differently, but both of them have it using AI functions and exposing it in LLM and using SQL to run batch inference on massive amount of data. Yeah, go ahead.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 反对 成本驱动下，懂行客户直接选开源/中国模型，不依赖平台路由；不懂行客户才让平台决定
a decision making on which product they actually use. That cost is the driver because they know if you do the prompt correctly, if you have the right harness thing, then for the most case these cheaper models will do the same business and you don't need to rely on Databricks to make a decision for you. But I think a lot of customers who are more naive or they don't know exactly what this sovereign AI means and all that stuff, they trust, they have a contract with Anthropic or whatever, then they let Databricks decide and the routing of Databricks. Some of it I would say would be cost-based decisions and some of it would be more contract dollar value decisions as to which AI they would use.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 口径不齐 客户从 Databricks 迁入 Snowflake 的原因之一是借 Snowflake 与 AWS 的绑定获得稀缺 GPU (印度客户 Alliance 微调模型)
客户迁移至Snowflake主要有两个原因。第一，Snowflake与AWS的深度绑定关系能帮助大客户获取稀缺的GPU算力资源。例如，印度的一家大型客户Alliance，为了进行大规模数据训练和对K3、K2.7及GLM-5.2等模型进行微调（fine-tune），需要大量GPU资源，因此选择通过Snowflake来确保从AWS获得所需的GPU capacity。第二，Snowflake推出的新产品，如Hybrid Table和Postgres Page，在功能和性能上表现出色，获得了客户的认可。尽管Databricks的AI产品Genie目前仍在促销期，尚未正式收费，但上述因素依然构成了客户选择迁移的充分理由。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 反对 Snowflake 会主动接触以前没用过其产品的 AI 需求客户；Databricks 认为直接转卖 LLM 利润率低、流失风险高
Snowflake在推出新产品时，更像是在开辟一个新的阵地，其新产品与老产品之间的关联性不如Databricks紧密。Snowflake会主动接触有AI需求的客户，即使这些客户之前没有使用过其产品。相较之下，Databricks更强调“Data Intelligence”，认为如果没有对客户业务背景和数据的深入理解，直接提供通用智能服务（如转卖LLM）的模式利润率低，且客户流失风险高，因此不倾向于采用这种低成本但粘性差的获客途径。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 反对 客户关注 Token 成本，Snowflake 自动路由到 GLM/Qwen/DeepSeek/小模型，不希望所有场景都用 Opus
许多客户还会基于基础模型，按照自身需求对模型进行微调。目前，客户主要关注的问题之一是Token成本。如果客户使用Astra 6或Claude Opus 5，Token成本会大幅上升。因此，客户开始选择Snowflake的自动路由功能，针对不同用例自动匹配适合的模型。我们还提供GLM 5.2、Qwen、Deep Sea以及更小的语言模型。客户现在可以针对适合的用例使用所需的任意大语言模型，而不希望所有场景都使用Opus或Astra。我们的自动路由能力可以确保客户针对不同用例使用最合适的模型。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 反对 中国开源模型成本仅为 Anthropic/OpenAI 的 1/20 且效率相当，模型差异化减弱——token 单价通缩
此外，行业壁垒也在发生变化。目前，Qwen、DeepSeek等中国模型的成本仅为Anthropic和OpenAI基础模型的1/20，但效率相当。因此，对客户而言，模型之间的差异化正在减弱。很多客户开始转向开放权重模型或开源模型，并自行掌控数据。目前，Anthropic和OpenAI面临的问题在于，数据和模型运行在其他国家。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 中性 Snowflake 与 Anthropic/OpenAI 签大合同，客户无需单独订阅模型、通过 Snowflake 合同统一使用 (即 token 走 Snowflake 计费转售)
Snowflake与Anthropic和OpenAI都签订了大型独家合同，确保模型公开发布后，几天内即可进入Snowflake平台，供客户使用和部署。客户可以根据自身需求，在Snowflake的上下文数据基础上使用任意模型。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 口径不齐 专家把 74% 毛利率指引说成“完全源于 AI 工作负载”，实际含义应是 AI 工作负载拉低毛利率 (76%→74%)，其表述因果不清
从首席财务官给出的收入指引来看，人工智能为Snowflake的收入带来了显著增长。公司给出的毛利率指引为74%，这完全源于AI工作负载。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 口径不齐 AI 收入按 SKU 计算，不做归因：AI 相关 SKU 是 AI，其他 (数仓、存储、分析、摄取) 都是核心
Yeah, absolutely. When we put our revenue together, there's no -- and I know a number of companies do attribution to saying, AI revenue is this or maybe they're selling a bundle, and they have to -- for accounting purposes, they have to break it out. Ours is down to the SKU level, and so we have a series of products that are AI related. And so, when we talk about AI and the performance, it's based on SKUs. And then the core is -- the core business, data warehouse, and storage, analytics, ingestion, what you know is Snowflake as a whole. And so, one of the things that we're seeing that is extremely encouraging and really goes to the variability of the business is we're seeing that when people adopt our AI products primarily CoCo, because we have the most data around CoCo, they're actually consuming more of the core. And so, it may be more personas.
- 反对 全年非 GAAP 产品毛利率指引 74%；AI 产品毛利天然低于核心
You seeing a really big decline there. You're seeing us hold gross margins with the increased revenue that you see. And we guide 74% non-GAAP product gross margins and you can see the strong cash flow generation. And so, we're doing that with less people around the enterprise. One of the things I talked about on earnings, if you look at sort of year-to-date hiring for first and second quarter this year versus last year, we added roughly 330 people this year. Last year, we added about 950 people. Of the 330 this year that we added, 180-ish of those more or less came from the Observe acquisition. And so, the amount of people that we're adding with the acceleration of revenue, I think, is really a testament of how times are changing. Yeah. On gross margin specifically, the company has done a really good job over the years of basically getting leverage and scale with our providers once we sort of reach a certain threshold. And so, as we launch all these new AI products, AI products naturally have a lower gross margin than our core business. The good news is it does drive our core business so that's helping out. But we feel that there's things that we can do over time to maintain a strong gross margin. Yeah.
- 中性 客户用 Snowflake 承诺额在多模型间选择，基础问题用开源或便宜模型 (token 单价通缩)
They want to be able to have model choice. And so, one of the things that we're able to do is people make a Snowflake commitment for CoCo and consumption and be able to use the right model for sure, the right question or the right activity, the right query to basically optimize the cost they're paying for that. And so, you don't want to use the most advanced, most technical, most expensive model for a very, very basic question. And so, you'll use maybe an open source model or a model that cost much less. Right? Right. Makes sense. One of the other questions we get just in terms of some of the recent performance is this sort of distinction between AI driven and the core. Right? And I think it gets a little confusing because AI can benefit the core. Like if you see new use cases, you need more data. But can you just sort of talk about those two dynamics and how do you kind of see the growth of the core business trending just given these new use cases that are coming on top of it? Yeah, absolutely. When we put our revenue together, there's no -- and I know a number of companies do attribution to saying, AI revenue is this or maybe they're selling a bundle, and they have to -- for accounting purposes, they have to break it out.
- 反对 Dan Niles：token 价格自 5 月以来下降一半，份额转向开源模型 (AI 产品 token 收入通缩)
1. **Databricks CEO Ali Ghodsi on Never Wanting the Job, Beating Snowflake and Why Databricks Stays Private** *Long Strange Trip: CEO to CEO with Brian Halligan, and Ali Ghodsi, co-founder and CEO of Databricks* The last round raised $5B against roughly $20B of demand, the anti-Snowflake playbook (open format, AI, a claimed 3x TCO edge), why he says Databricks would be worth more public, and the one-bottleneck operating model. 2. **Dan Niles: This Isn’t the Top, It’s a Speed Bump** *RiskReversal with Dan Nathan, and Dan Niles, founder of Niles Investment Management* The AI trade’s leading indicators: Texas’s data-center moratorium, the labs’ own slow-down talk, token prices down by half since May with share shifting to open-weight models, a 5% ten-year, and where he still wants to be (Meta, Google, Microsoft, Apple). 3. **Palantir CEO Alex Karp on AI Safety, Liability and Nationalizing the Frontier Labs** *CNBC’s Squawk on the Street with Carl Quintanilla, David Faber and Sara Eisen, and Alex Karp, co-founder and CEO of
来源背景：TMTB Podcast Intel (由 Ethmos 研究 agent 生成的播客摘要，明示)。相关内容为 Databricks CEO Ali Ghodsi 在 Brian Halligan 播客 'Long Strange Trip' 中的说法，以及 Dan Niles (Niles Investment Management 创始人) 的观点。视角局限：只有一句话摘要，正文是图片未读到；Ghodsi 是直接竞争对手，有利益偏向。 - 口径不齐 Snowflake 向 Anthropic 等支付的 token 成本比例在内部资料中也没有写明，非常机密
Snowflake与其他厂商合作时，肯定会有一个成本价。Snowflake用户使用ANTHROPIC模型时，Snowflake会向用户收取费用，同时也需要支付给ANTHROPIC等合作方一定的成本。具体来说，Snowflake每收多少million的token，会收多少钱，再支付给ANTHROPIC或微软等多少钱。内部资料并没有明确写出具体比例，所以到底给了百分之多少并不确定，这部分非常机密，内部资料完全没有涉及。现在来看，主要涉及四笔费用。在AI时代，这些费用取决于客户如何使用Snowflake的AI服务。Snowflake的AI主要是Cortex Code和Co-work这两大类。Cortex Code系列，客户用得越多，产生的费用也会越高。如果Code用得很多，AI token的数据量就会上升，收费也会变高。除了token费用高，最常见的使用场景是连接Snowflake数据库，通过大模型在数据库中运行SQL或读取数据。AI用得越多，使用warehouse资源节点也会越多，传统warehouse资源节点的费用也会大幅增加。这两部分费用随着用户使用量增加而增长，Snowflake因此获利更多。借助AI，用户与大模型交互时会不断调整prompt，这意味着资源节点的使用会更多。如果没有AI，用户会更精准地写SQL语句，相对来说会更省钱。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 反对 产品毛利率过去三个季度收缩，主因 AI 工作负载上升及推理和云基础设施成本增加；credit 价格一般不调
Snowflake产品毛利率在过去三个季度确实有收缩，主要原因可能是AI
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 口径不齐 客户使用 Snowflake 时更多选择 Claude 模型，并要求尽早用最新版本 (token 层依赖 Anthropic)
Snowflake 和 Anthropic 的合作只会越来越深入。根据我的观察，客户在使用 Snowflake 时，更多会选择Claude的模型，很多用户也更愿意用 Anthropic 最新的模型。比如说，Anthropic 假设推出了 FABLE 这样的 LLM 模型，而 Snowflake 官方可能只支持到 5.0 版本，但有些客户非常关注模型的版本，甚至在 Snowflake 还没有来得及推出 5.1 版本时，就会主动询问能否用最新的 FABLE 5.1 或 Opus 5.1 版本。这些版本虽然在 Snowflake 官方还没有上线，但由于 Snowflake 和 Anthropic 合作紧密，也会通过 public preview 或 private preview 的方式，让客户提前体验到 Anthropic 的最新模型。从这个角度来看，Snowflake 和 Anthropic 的合作会越来越密切。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 反对 AI 模型粘性最弱，换模型供应商只需改 URL 和 endpoint；数据平台因数据引力粘性高
数据公司的用户粘性相对较好，因为存在“数据引力”，客户数据量增多后迁移成本较高。相比之下，AI模型的粘性最弱，客户更换模型供应商（如从Anthropic更换）仅需修改URL和endpoint，非常便捷。对于数据公司而言，无论是Snowfla ke还是Databricks，客户粘性都比较高。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。 - 口径不齐 Snowflake Cortex AI 在商业合作上与 Anthropic 绑定较深，中立性不如 Databricks
此外，在生态合作方面，Databricks更具中立性和灵活性。它可以在AWS、Azure、GCP三朵云上销售，并与Anthropic、OpenAI、Google等主流大模型厂商均建立了战略合作伙伴关系，为客户提供了更广泛的云厂商和模型选择。而Snowflake目前主要在AWS上售卖，其Cortex AI在商业合作上与Anthropic绑定较深，尽管也能通过External Model连接Azure AI和Gemini，但在中立性上有所不同。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。
E04 单价和竞争
判断：锁定在变弱，而且是 Snowflake 为了赢新客户主动让出来的。今天整体迁走仍然少，Snowflake 的产品力也在变强；但新预算更多流向 Databricks，同一份数据可以被别的引擎读写以后，计算的单价迟早要拼。
开放的代价：Snowflake 让其他引擎读写 Snowflake 上的数据，"鼓励客户选最便宜/最好的引擎"；Polaris 支持 9 个外部引擎读取，开始支持写入；CEO 说托管 Iceberg 表让其他引擎可查询以后，"价值必须往上游移"(治理、灾备、agent、可观测)。一个演示里，Databricks 通过目录读 Snowflake 的 HR 表时，Snowflake 定义的脱敏策略仍然生效：数据治理留在 Snowflake，计算可以在 Databricks。
Databricks 和云厂商：Databricks 增速约是 Snowflake 的两倍 (分析师在 Investor Day 当面问 CEO："低 30% 的增速感觉不够")；新平台预算约 60:40 偏向 Databricks；前 Databricks 销售说去年约 200 家客户迁移。三家云厂商都在用数百万美元补贴迁移。按每 TB 处理成本，Snowflake 比 Databricks 贵约 40%，但免运维。
Snowflake 守得住的地方：两家之间整体切换"非常少见"；PB 级客户迁往 Databricks 要花数百万美元并有业务中断风险，通常不迁；正面交锋 Snowflake 胜约 60%、Databricks 约 30%；Snowflake 托管 Iceberg 存储 GA 后，把存储收入拿回来一部分，管理层说 Iceberg 的影响"比原先建模的好"；Adaptive Compute 性能大幅提升，但按收入中性定价 (不像 2024 年的性能改进那样直接变成收入逆风)；管理层说 Gen2 等性价比改进已经被量和新负载抵消。Databricks 的 Genie 原本免费，后来收费再降到半价，对手也在为 AI 层拼价格。
We are steering the iceberg standard but we're also being among the first at implementing it. Right now, we're the broadest in terms of the implementation of the V3 spec and we're steering the V4 spec. I mentioned this morning, we integrated all the rest catalog APIs into Horizon to make sure that we can interoperate with data regardless of where it sits even if it's on Databricks. We can read and write data of other engines with glue and others. They can read and write data that sits in Snowflake. So this whole notion of I'm locked in and I put my data like, that's not excuse customers. Please use whatever gives you the best experience or the best performance, the best economics. And okay, this one is super important for all of you in this room. When we introduce Iceberg, I think some of us regrets how much noise it cost. But one of the things that was factual was with Snowflake, you store in our format and you pay for a storage.
You can crush them for saying that. So we support open format. We want to increasingly make it painless for our customers to deal with open formats. So we have now something called Snowflake-managed Iceberg tables, which is a fancy way of saying, you cannot have your cake and eat it, too, which is you can store data with Snowflake but have it be queryable in Iceberg format by any other engine. And the place where we create value has to be further upstream. It has to be in, do we provide better governance? Do we provide better disaster recovery? Is it easier to create and run agents on top of Snowflake? Do we provide a better observability solution? And so, there is this whole stack of things on top of the data platform that is open that we have to be providing. And to a large extent for a lot of these, my attitude is bring it on. It's not that easy to create the platform that Snowflake is.
There's a microphone right behind you. Brian, I like your socks. Just I'v egone with the swag store. Sridhar, if all these products take off, low 30% growth doesn't really feel right. I mean, it feels like this is a 40% market. Your primary competitor is growing twice as fast as you. So, when you think about ultimately what you think these new solutions can do to help accelerate growth, I know you're not giving guidance here but it doesn't feel like you're at cruising altitude from where the rest of the industry is out right now. I don't know what to say. Absolutely, we aspire for more but showing is the new doing.
回 答 者：此前比较的是季度环比增速，但从整体同比增速来看，Databricks的增长确实更快。Databricks的整体同比增速为80%；以2025年第1季度和2026年第1季度相比，同比增速同样为80%，而Snowflake约为35%，因此Databricks的增速超过Snowflake的2倍。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。
BJ: Hyperscalers are very aggressive in going after this business in the data cloud and compute side of the business. They're heavily subsidizing. Their incentives here, we're talking millions of dollars can be offered if the business case is there from our side to subsidize all the engineering work and migrate off Snowflake or Databricks into their environments. And that is concerning to me because then if you think about it, that is being subsidized to client impressions. That's exactly how Snowflake grew in the beginning, by the way, right? But that fundamentally changes the perception in today's environment. I feel like if we don't strategically advise our clients on the pros and cons of this, they're going to be paying for it later as the subsidies expire. So that I'm more worried about from a market dynamics standpoint of view. All 3 of them are offering really highly incentivized subsidies for migrations.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。
EXPERT: So if I take as a baseline price, I will put $22, let us say $25 per terabyte of data for processing at Databricks, and around 35-40 at Snowflake. So if you compare that way, 35-40 in Snowflake. So if I compare that terabyte to terabyte costing, you pay at least 40% higher on Snowflake, and you don't have any maintenance on Snowflake, by the way. While in Databricks, you need very highly skilled engineers. So that's why, on kind of just number basis, Snowflake is 40% costlier than Databricks.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。
much because actually at the data warehouse level, they both Databricks and Snowflake have come to market with federation capabilities where if you have some tables of data in SAP or in some other kind of technology that sits outside of a Databricks or outside of a Snowflake, through these federation capabilities in Databricks and Snowflake, you can actually still bring that data into Snowflake. So today, I don't see that as a competitive differentiator or something that AtScale really should be calling out, and it was never something that we made a point. But when we were selling, customers that asked the question, hey, if 5 years from now, 10 years from now, we switch from Snowflake to Databricks, what would happen to our semantic layer investment with AtScale? We would answer, honestly, say, well, nothing, you know, because your semantics are built in AtScale, and those are easily portable into a Databricks or Snowflake. So it gave customers confidence that they weren't locked in and they had this ability to change. But the reality is it's very rare that someone switches from Snowflake to Databricks or vice versa. So I don't think that's a strong differentiator anymore.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。
已经使用 Snowflake 的客户通常认为没有必要迁移至 Databricks，因为两者性能相近，而 Snowflake 速度更快、成本更低。迁移本身需要完成一系列工作。对于拥有 TB 级和 PB 级数据的大型企业部署，迁移至 Databricks 可能需要花费数百万美元，同时还会带来业务风险和业务中断，因此客户通常不会迁移。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。
A:在多数交易中 Snowflake 通常会获胜。在正面交锋时，Snowflake 赢得 60% 的机会，Databricks 赢得 30%。总体上 Snowflake 确实赢得更多对手的机会。
来源背景：MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。
There's a lot more during the conference on how do we help govern agents, which is security policy, identity for agents, data movement, exfiltration protection, all of those. We also talked about adaptive compute. I put it in here just because there's something I want all of you to be clear on. Massive performance improvement but we're doing the exact same thing that we did with Gen 2. We priced it in a way that we're aiming for revenue neutrality. So our customers get the benefit that is materially faster but none of you need to go change your models. We're still good even though we're trying to push very hard for the adoption of all of this. Interactive analytics for responsive experiences. You can say this gets into the click house type of work workloads.
When we introduce Iceberg, I think some of us regrets how much noise it cost. But one of the things that was factual was with Snowflake, you store in our format and you pay for a storage. With Iceberg, we always said it's customer managed storage but there's no reason for that trade-off. So we introduce and is generally available here at Summit, Snowflake managed storage for iceberg. So you can still be interoperable but we'll do the management of the storage. We'll give you economics. So, that I think is going to be in an even better tailwind relative to how we thought and modeled the adoption of Iceberg. Sharing – I think all of you know, I'm personally passionate about the network effect, personally passionate about the unsiloing of data and helping organizations connect. I am very excited that we're finally breaking out of the – it's two parties. You need directional, how do we do multiparty collaboration and symmetric.
And yet your role is to also abstract value and price one level above the core components of that, so you can capture gross margin. Maybe, Brian, just tell us a little bit about how that philosophy is coming into play in 2026 with things like Gen2. Yeah. I mean, part of the business is you got to be better from a price performance perspective than the last generation was. And you always have to give your customers the ability to get more out of your product. And so, we price that in, but we're seeing the offset of that with volume and new jobs coming into the company. And so, we carefully – as we go and price stuff, we carefully take that into consideration. So there's not big step downs in revenue, but we also want to get our customers get the benefit and the value of these performance enhancements that we're doing on the platform. I'll perhaps add on a quote from one of my previous bosses and mentors. Revenue solves all known problems. Fantastic.
Sanjit K. Singh (Morgan Stanley): congrats on reasserting 30% product revenue growth in Q4. Brian G. Robins: it's really based on the high stable growth that we see in our core business. Brian G. Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type. Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake. Brian G. Robins: we guide based on the Observe customer behavior up until really the point of earnings. Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
展开本节全部 181 条原文 (支持 44 / 反对 80 / 其他 57，按时间)
- 中性 Zero Copy 主要是和 SaaS 厂商的数据互通，管理层称双赢
Sridhar Ramaswamy: we are 15% to 20% of the way through kind of on-prem legacy migrations. Raimo Lenschow (Barclays): Zero Copy comes up a lot in the conversation. Sridhar Ramaswamy: Zero Copy generally comes out in the context of SaaS vendors, Sridhar Ramaswamy: Zero Copy agreements just make data flow more smoothly Sridhar Ramaswamy: We see these as a win-win. Ardy [ph] for Mark Murphy (JPMorgan): explicitly stated that Snowflake budget is now tied to their AI budget, and they're kind of broadening Sridhar Ramaswamy: The strongest pattern that we have had to work hard and earned this year is to be that genuine player Sridhar Ramaswamy: no amount of talking can make you that, you need project, products that produce the magic Matthew Martino (Goldman Sachs): what about the Snowflake platform is allowing customers to really accelerate their AI journeys? Sridhar Ramaswamy: with things like Snowflake Intelligence and Streamlit, which we are very heavy users of Sridhar Ramaswamy: investments like Openflow or even Postgres are going to expand the aperture Alex Zukin (Wolfe Research): the momentum that you're describing is showing up in bookings Sridhar Ramaswamy: The virtuous cycle of a Snowflake customer is one in which they sign a deal. Sridhar Ramaswamy: they don't have to make any pre-commits towards AI Sridhar Ramaswamy: bookings are an indicator of how much a customer thinks they're going to spend in the coming years. Sridhar Ramaswamy: we had four, nine-figure deals this quarter. Patrick Colville (Scotiabank): could you just remind us what does that $100 million actually include? Sridhar Ramaswamy: The $100 million is primarily the product suite, but it's the whole stack. Sridhar Ramaswamy: Cortex Search and Cortex Analyst, which are our unstructured and structured data products, respectively. Brian G. Robins: The $200 million is a buy side that we're buying from Anthropic. Sridhar Ramaswamy: partnerships with Anthropic, which include a buy, but also a broader go-to-market motion Brad Reback (Stifel): The op margin obviously down-ticked on the first half sales and marketing investment.
- 支持 一旦全押 Snowflake 等平台并在上面跑数百个 agent 流程，切换成本高，每年涨价 10% 也难以触发切换——定价权支持
They will then be subject to price increases. Because the switching cost will be quite high, even a 10% per year rise or more is annoying but often not enough to justify switching once you have taken a bet.
来源背景：自述'We are a global systems integrator with approximately 300,000 employees'，在其 AI 业务中做 Palantir 渠道/实施 (Palantir 占其 AI 业务 <5%)，与 Microsoft/Databricks/Snowflake 等都有合作；面向企业客户，熟悉欧洲公共部门。主题是 Palantir，Snowflake 只是顺带。视角局限：SI 视角、看重'锁定'叙事；无 Snowflake 具体数字。 - 中性 Snowflake 和 Databricks 在数仓和大规模数据集成上更强，数据侧各平台解决的问题大致相同
On the data side, companies take a bet on one of these ecosystems and align to it. Palantir is strong in data fusion, while Snowflake and Databricks are likely stronger in data warehousing and very high-volume data integrations, although it can be argued both ways.
来源背景：自述'We are a global systems integrator with approximately 300,000 employees'，在其 AI 业务中做 Palantir 渠道/实施 (Palantir 占其 AI 业务 <5%)，与 Microsoft/Databricks/Snowflake 等都有合作；面向企业客户，熟悉欧洲公共部门。主题是 Palantir，Snowflake 只是顺带。视角局限：SI 视角、看重'锁定'叙事；无 Snowflake 具体数字。 - 反对 用户从 Snowflake 迁移到 BigQuery 或 Databricks 是挺明确的趋势，尤其是数据全在一个云上的公司；但整体净增长不是零和
用户从Snowflake迁移到BigQuery或Databricks是挺明确的趋势，特别是一个公司的data全存在一个cloud上，但如果跨cloud的话就不好说了。但整体上还是净增长，并不是一个零和游戏，因为AI agent在2B的场景他总要有一些。Multi cloud现在还是很大的趋势。有些厂商是不得不选择某个云，比如必须用TPU或者特定的Office上的应用，这些情况下可能就会选择单一的cloud。北美很多厂商都有自己的multi cloud策略，但具体还是要看他的策略。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。 - 支持 Snowflake 底层为多云设计，跨云数据分析优势非常明显，BigQuery 建议客户复制数据到 GCP
这个肯定是有差距的，BigQuery最初设计并不是为了multi cloud，没有做特别native的优化，而Snowflake从底层就是natively为multi cloud去用的。跨云数据分析方面，Snowflake优势非常明显，像数据需要在不同的cloud之间的表去做join和分析。所以BigQuery通常会建议用户将数据至少复制一份到GCP，这样查询速度会快很多。不过现在这点是在优先级上，因为GCP意识到要想增长现在需要靠GCP以外的增长。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。 - 中性 未来两年数仓份额获取潜力基本是 BigQuery 和 Snowflake，Redshift 无明显优势
即使在 AWS 上，Redshift 与 Database 或 Snowflake相比，并不好说就有特别明显的优势，所以现在基本上还是BigQuery 和 Snowflake。但BigTable 的增速预计比 Snowflake 更快。现在做很多政府相关的项目，如果项目必须部署在 GCP 上，那其他厂商就很难有优势。这块还是由模型能力决定。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。 - 支持 Snowpark 直接读 Delta Lake，代码兼容性 90-98%，与 Databricks 比成本便宜 2-5 倍、性能高约 40%，推动从 Databricks 迁入
我们大幅推动了客户采用，包括从Databricks的迁移。在许多概念验证、试点和与Databricks的基准测试中，Snowflake显示出2到5倍的成本优化。我们还观察到大约40%的性能提升。正因为如此，我们为客户采用开辟了新的机会和用例 ，包括使用Snowpark。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 Iceberg 已成为默认开放表格式，客户不希望被平台锁定
A:Iceberg已成为我们的主要开放表格式。它现在是默认格式。Iceberg的性能和支持显著提高；客户现在可以查询、读取并与Snowflake Horizon目录同步。以前的限制已被解决。许多客户不希望平台锁定，这种方法有助于朝着那一方向发展。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 反对 Iceberg 导致 Snowflake 内部存储减少：客户把数据放 S3，只在 Snowflake 上跑计算
在Snowflake内部有一些存储减少，因为一些客户没有使用Snowflake存储他们的数据；他们将数据保存在外部对象存储中，例如AWS S3存储桶，同时在Snowflake上运行计算。这平衡了存储减少的影响。与访问外部存储相比，我们在查询性能上优于像BigQuery和Redshift这样的其他供应商。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 开放格式让 Snowflake 能覆盖以前流向 Trino、ClickHouse 的市场
借助开源格式，我们能够覆盖以前转向Trino、ClickHouse或其他独立供应商的市场。客户现在可以原生利用Snowflake。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 内存涨价未传导到客户价格，因为 Snowflake 不按虚拟机/内存计价
我们不把内存作为特定的定价指标。客户选择小、中或大实例，Snowflake并不明确披露实例具有多少RAM或CPU。到目前为止，内存成本上涨并没有向客户传导。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 公开费率可协商，价格变化通常在购买时才发现；没有政策性涨价
A:定价因地区和云而异。我对任何涨价政策没有可见性。公开费率通常可协商；任何变化通常在客户购买时被发现。就我个人而言，我没有任何关于涨价的信息，我也不认为会有政策驱动的涨价。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 中性 Iceberg 增长 40-45% 但只占总消费约 5%，2026 年到不了 10%
我估计现在大约占总体消费的 5%。它增长更快，但不会快到 5% 会大幅变化。它可能上升到 6% 或许 7%，但我不认为它会在 2026 年达到 10%。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 中性 Snowflake 未发出列价上调信号，折扣存在但遵循列价
A:与我合作的其他一些平台公司不同，据我所知，Snowflake 尚未发出任何列价上调的信号。当然有折扣，但它们仍然遵循列价，所以我们会看到的。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 客户越来越多请合作伙伴帮忙优化存储分层、运行时和许可成本——成本敏感
客户越来越多地要求我们帮助他们优化存储分层并在基本情况下改善运行时间，更有效地使用他们的许可占比。这并不意味着 Snowflake 做了什么不对；只是因为更多的计算和更多的存储被放到了 Snowflake 上。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 支持 客户锁定足够深后 Snowflake 可能提价，但目前在做相反的事——争份额
还有一点：我的印象是 Snowflake 正在尝试把客户吸引进它的生态系统，一旦你在 Snowflake 堆栈中足够深，它可能会上调价格。如果你看看这些 IT 公司历史上的运作方式，一旦他们知道你的切换成本足够高，他们就会开始提高价格。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 Databricks SQL 数仓收入 run-rate 在 2025Q3 突破 $1B，较 2024 年底 $600M 近乎翻倍，进攻 Snowflake 腹地
Databricks不仅在AI主场领先，还在Snowflake的腹地——数据仓库——发起了进攻。Databricks数据仓库业务（Databricks SQL）的收入已于2025年Q3突破$1B run-rate（据Databricks 2025年12月官方披露），较2024年底的$600M实现近乎翻倍。"Lakehouse"架构——将数据湖的灵活性与数据仓库的性能合二为一——正在赢得不想维护两套系统的企业。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 反对 Fabric 捆绑在 M365/Azure 订阅中，边际成本极低，直接压缩 Databricks 和 Snowflake 的定价空间
**1）定价降维**——Fabric捆绑在Microsoft 365/Azure订阅中，对已有微软生态的企业而言边际成本极低。这直接挤压了Databricks和Snowflake独立收费的定价空间。
来源背景：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 - 反对 支持 Iceberg 为一等公民，让客户放心其他引擎可读同一份数据——开放格式降低锁定
Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide.
- 反对 为避免账单冲击将推出 Snowflake Intelligence 每用户上限——管控消费
Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go.
- 反对 Snowflake 在端到端 AI/ML 和 agentic analytics 上不如 Databricks，导致落选
It performs well from an analytics perspective, and it's very familiar to enterprise buyers. The separate separation of storage and compute is helpful, and it's a clean ecosystem for lack of better words. Where I felt Snowflake was less compelling for our use case specifically was around end-to-end AI and ML workflow depth. It has been investing heavily in ML and AI features, but for us, Databricks sort of felt more natural for data engineering plus ML experimentation. Snowflake is excellent as a warehouse, but I saw that it's less flexible for the kind of agentic analytics that we're looking to do from custom AI workflow. We also learned and looked at Azure ML and AI data stack. Azure ML is broader, got good enterprise security, integration with Microsoft and Access and OpenAI.
来源背景：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 - 中性 Snowflake 预计月成本 1-2.5 万美元，低于 Databricks 2.5-3 万美元；两者都是最贵档
EXPERT: So using $1,000 a month as a simple baseline for self-hosted style workflow layer, I would think about pricing tiers rather than exact quotes. But to give you an idea, because each product matters, meters differently, I would say, directionally for a similar early-stage Untrap workload, the comparison would probably be something like this. So if, again, if self-hosted n8n is $1,000 lowest cost, mostly infrastructure plus engineering, managed vector databases, Pinecone, Quadrant, et cetera, would probably be about $2,000-$5,000, or about 100%-400% above baseline. So it's more expensive than self-hosted, but narrower, mainly retrieval, I'd say embeddings and semantic search. Then, if I look at Azure and AWS services, I'd probably say that's between $5,000-$15,000 or 400%-1400% above baseline. I'd say costs rise once you include inference, storage orchestration, especially monitoring and networking as well. But the inference is, does eat away quite a bit. Snowflake, by cost comparison, would be about $10,000-$25,000 or, I would say, ish. It's strongly strong governance analytics, like I said, so more of an entire end-to-end process or a tool, and then Databricks would probably be around $25,000-$30,000. And it's probably the most expensive just because of the central lake house and ML platform, but it also replaces more of the data engineering and ML stack as well that we would have to run otherwise. So.
来源背景：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 - 中性 评分 Databricks 8、Snowflake 7.5，差距来自 AI/ML 需求
EXPERT: Reflecting both on quality and value for money, I'd say Databricks, 8 out of 10. I'd say the best overall platform for money in our opinion, especially if you're serious about building scalable applications. Like I said, I've used this tool at the bank before, so had some familiarity with it. Strong for Lakehouse, very good data engineering processes, governance is built in, strong workflow for ML, and also any kind of pipeline and production AI. So, that's kind of the reason why I would give it an 8 out of 10. For Snowflake, 7.5, maybe. Again, excellent value for governed analytics, warehousing, good enterprise reporting,
来源背景：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 - 支持 Snowflake 易用、对企业买家熟悉、存算分离
So it's great if you're serious about building a scalable data platform, but maybe overkill if you need a lightweight analytical engine. So we also looked at Snowflake. Snowflake was strong from a data warehousing and governance perspective. It's easy to use. It performs well from an analytics perspective, and it's very familiar to enterprise buyers. The separate separation of storage and compute is helpful, and it's a clean ecosystem for lack of better words. Where I felt Snowflake was less compelling for our use case specifically was around end-to-end AI and ML workflow depth.
来源背景：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 - 反对 按每 TB 处理成本 Snowflake 比 Databricks 贵约 40%，但 Snowflake 免运维
EXPERT: So if I take as a baseline price, I will put $22, let us say $25 per terabyte of data for processing at Databricks, and around 35-40 at Snowflake. So if you compare that way, 35-40 in Snowflake. So if I compare that terabyte to terabyte costing, you pay at least 40% higher on Snowflake, and you don't have any maintenance on Snowflake, by the way. While in Databricks, you need very highly skilled engineers. So that's why, on kind of just number basis, Snowflake is 40% costlier than Databricks.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 - 反对 最终选 AWS+Databricks：非结构化数据预处理需求 + 客户给 AWS 3 年承诺换得合理价格
EXPERT: So we are using, in the given scenario, AWS and Databricks combination because we had a huge preprocessing on unstructured data for our data science and ML use cases. At the same time, we wanted to use Redshift, S3, and build agentic solution using Bedrock. So all this was under one roof, and our client was able to give a 3-year commitment. We got very reasonable pricing from AWS. So AWS plus Databricks comes out as a winner.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 - 中性 Snowflake 适合数据分析师/BI 报表负载；Databricks 适合大量非结构化数据的 AI 负载
EXPERT: We have evaluated multiple platforms, including AWS, GCP, and Azure. What I feel is, in ballpark terms, if you are building any high-end or complex, demanding application or platform, then AWS or GCP are the best bet. If you are coming from a Microsoft background, then Azure is your best bet. If you are more like a data scientist kind of team and you want to process a lot of unstructured or structured data for AI workloads, then Databricks may work well. At the same time, Snowflake is very good if you have more of a data analyst kind of focus and a lot of BI and reporting, but you don't want to get into an easy-to-use kind of platform with heavy concurrency and scaling performance.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 - 反对 工程能力强的组织用 GCP+Databricks 是最便宜组合
cost-efficient, but you have to commit for 1-3 years. So if you are not committing, then you will find that AWS will be 20%-30% costlier than GCP. Azure, depending on your relationship, will be somewhere on the kind of 15%-20% costlier than GCP, sometimes nearer to AWS, so slightly lower. Now, coming to Databricks and Snowflake, I think Snowflake is slightly costlier, maybe in the range of $30-$40 per petabyte kind of data. But depending on your workload and if you are more toward data science history across a lot of unstructured data, then Databricks will be cheaper for you. They have DPU, they have spot instances to make it cheaper. So depending on the use cases, a midsize organization can find it, but if they are like engineering-heavy organization and they know how to use it, maybe GCP and Databricks will be the cheapest combo.
来源背景：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 - 支持 Snowflake 与 Databricks 互相切换非常少见
much because actually at the data warehouse level, they both Databricks and Snowflake have come to market with federation capabilities where if you have some tables of data in SAP or in some other kind of technology that sits outside of a Databricks or outside of a Snowflake, through these federation capabilities in Databricks and Snowflake, you can actually still bring that data into Snowflake. So today, I don't see that as a competitive differentiator or something that AtScale really should be calling out, and it was never something that we made a point. But when we were selling, customers that asked the question, hey, if 5 years from now, 10 years from now, we switch from Snowflake to Databricks, what would happen to our semantic layer investment with AtScale? We would answer, honestly, say, well, nothing, you know, because your semantics are built in AtScale, and those are easily portable into a Databricks or Snowflake. So it gave customers confidence that they weren't locked in and they had this ability to change. But the reality is it's very rare that someone switches from Snowflake to Databricks or vice versa. So I don't think that's a strong differentiator anymore.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 支持 企业对数仓是 5-10 年的大投资，不会频繁更换
DG: Yes. Okay. I understand. Before, the reality is most companies make 1 big investment. You know, they're going to be with Databricks for 5-10 years, they're going to be with Snowflake 5-10 years, and so on and so on. So they're not going to chop and change their chosen data warehouse, you know, every year. Where AtScale's warehouse agnostic benefit used to be strong. It's actually more strong just from a sales proposition standpoint. The client doesn't care so
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 反对 AI 时代语义/context 层更核心，离数据最近的数仓会赢，但专家押 Databricks
We'll see. It's a very big space, you know, I think there's a lot of room for a lot of players, but I think the ones that are going to win are the ones who are closest to the data. And the ones that are closest to the data are the data warehouses, and I think Databricks has a much stronger position here because they really have been born out of the AI machine learning world. They've done a lot better than Snowflake in that space. So my money will be on Databricks to be the one to watch out for in this semantic layer and context layer space.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 中性 客户多数同时有 Snowflake 和 Databricks；Snowflake 卖点是把 AI 放在受治理的数据旁边 (Cortex)，Databricks 赢工程导向文化
BJ: Yes, I think. Snowflake... and I'm also seeing a lot more Databricks at the moment, right? If I think of what we talked about, Snowflake is still resonating in my client space. And by the way, my clients have both environments for the most part. They're not centralized on one today. But on Snowflake, the pull is more around keeping AI close to governed data through Cortex, Cortex Search, Cortex Analyst, and Snowflake just released Cortex Code as well. So I believe that is the angle they're taking. Those cultures become more governed and it actually appeals to a lot of financial services organizations accordingly. What I see is, there's more engineering-led cultures in place. Databricks takes over. Faster pipeline modernization, migration acceleration, less manual data engineering, and also that persona loves to build on Databricks.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 自称偏 Snowflake 的专家承认最近是 'Databricks 的秀'
BJ: My bias has been rightfully checked lately. It's been a Databricks show for me. However, I'm more worried about the hyperscalers.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 云厂商为把客户从 Snowflake/Databricks 迁出提供数百万美元补贴，三家都在做
BJ: Hyperscalers are very aggressive in going after this business in the data cloud and compute side of the business. They're heavily subsidizing. Their incentives here, we're talking millions of dollars can be offered if the business case is there from our side to subsidize all the engineering work and migrate off Snowflake or Databricks into their environments. And that is concerning to me because then if you think about it, that is being subsidized to client impressions. That's exactly how Snowflake grew in the beginning, by the way, right? But that fundamentally changes the perception in today's environment. I feel like if we don't strategically advise our clients on the pros and cons of this, they're going to be paying for it later as the subsidies expire. So that I'm more worried about from a market dynamics standpoint of view. All 3 of them are offering really highly incentivized subsidies for migrations.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 补贴额度从 10-20 万美元升到包揽整个迁移
BJ: I was concerned when I see that millions are being offered before it used to be $100,000 was the offer, $200,000. Now it's like, no, we'll take care of the whole migration. And I was like, interesting.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 专家认为云厂商补贴对 Snowflake 业务有实质影响，且已有一例 AWS 迁出
BB: Yeah. And what's the implication then? I mean, do you think that has a real knock-on effect in terms of Snowflake's business? BJ: Yes, absolutely.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 支持 Snowflake 里花时间整理的一方数据迁出门槛高，除非出了问题才会考虑
BJ: And I think that is a bit of a threat. With that being said, a good advisor will always help the client think through it holistically versus immediate, 1 or 2 years kind of review. And I think there's a case to be made that your analytical use cases... let's forget about automation and agentics... they'll fundamentally drive the majority of the Fortune 500 insights, and that will continue to support Snowflake as well. And I believe a lot of the early traction and the biggest hurdle to all of this is that your first-party data that sits within Snowflake that you've spent a lot of time curating to migrate it off in the same manner. Something really needs to be off for someone to consider that. And again, sometimes the ROI, fees, experience from a sales rep can dictate a lot of this too.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 支持 无复制的数据共享仍是 Snowflake 的杠杆，尤其面向客户的业务
BJ: I go back to use cases and the original decision because mind you, Snowflake sometimes was also the recipient of the migrations of the hyperscalers as well or other environments on-prem or whatnot. So there was a decision made. There's efficiency in mapping it out. For Snowflake, in my view, Snowflake was when any time you needed to share data without replication and those use cases, particularly in the customer-facing side of the business, still gives a lot of leverage for Snowflake. I think it's just the short-term navigation of the hyperscalers that are going after that business versus Databricks. I believe that that is more of a concern for them. I think the buyer persona for Snowflake still remains the same, not the same as Databricks, which is more AI-driven, a data science look and feel versus data engineering and then more SQL and business insights BI profiles. So there's a bit of that they have to do. They have to know, I believe, Snowflake has to attract the deep data scientists. That was always Databricks.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 中性 Snowflake 买方画像仍是 SQL/BI，需要吸引深度数据科学家 (Databricks 的地盘)
BJ: I go back to use cases and the original decision because mind you, Snowflake sometimes was also the recipient of the migrations of the hyperscalers as well or other environments on-prem or whatnot. So there was a decision made. There's efficiency in mapping it out. For Snowflake, in my view, Snowflake was when any time you needed to share data without replication and those use cases, particularly in the customer-facing side of the business, still gives a lot of leverage for Snowflake. I think it's just the short-term navigation of the hyperscalers that are going after that business versus Databricks. I believe that that is more of a concern for them. I think the buyer persona for Snowflake still remains the same, not the same as Databricks, which is more AI-driven, a data science look and feel versus data engineering and then more SQL and business insights BI profiles. So there's a bit of that they have to do. They have to know, I believe, Snowflake has to attract the deep data scientists. That was always Databricks.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 支持 分析型用例将继续驱动大多数财富 500 强洞察并支撑 Snowflake
BJ: And I think that is a bit of a threat. With that being said, a good advisor will always help the client think through it holistically versus immediate, 1 or 2 years kind of review. And I think there's a case to be made that your analytical use cases... let's forget about automation and agentics... they'll fundamentally drive the majority of the Fortune 500 insights, and that will continue to support Snowflake as well. And I believe a lot of the early traction and the biggest hurdle to all of this is that your first-party data that sits within Snowflake that you've spent a lot of time curating to migrate it off in the same manner. Something really needs to be off for someone to consider that. And again, sometimes the ROI, fees, experience from a sales rep can dictate a lot of this too.
来源背景：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 - 反对 开放格式成为主流，Snowflake 仍偏专有但已部分开放；客户整体转向开放架构
HS: See, if you look at the recent thing, everybody started taking on open architecture. Like Delta was not open earlier, but 2 years back, Databricks made it open. In fact, Databricks is now on Iceberg as well, right? So similarly, Snowflake. It is still proprietary, but they have made some extent of Snowflake as open. So I think people are moving towards open architecture because of the 1 key thing, which is adoption. The resistance to adoption has to come down. However, when you compete with somebody like a Snowflake, which is a, I would say a different tool, which is a data warehouse, right? The use cases are very different. The workloads are very different. I'll put Redshift, Snowflake, Synapse, or any data warehouse for that matter in the same bucket. The use case, what's running on Databricks will be very different from what is running on Snowflake. So it has to be modeled accordingly as per the use case or the requirement of the customer to position Dremio on that front. But towards openness, I think all of these customers are moving towards a more open architecture, which was probably not true 2 years or 3 years back.
来源背景：Harsh Singla，标题栏写 'Former Regional Sales Director - Dremio' ，但主持人介绍为前 Dremio APAC 企业客户经理 (约 2 年多)，现任 Confluent 区域销售总监约 3 年 ；更早在 Qubole。即离开 Dremio 约 3 年。视角局限：APAC 销售，离开 Dremio 已久；对 Snowflake 只有泛泛比较；Dremio 立场天然推开放湖仓、反数仓。 - 反对 SAP 客户 30-40% 的数据在 SAP 之外，SAP+Dremio 可借此挤压 Snowflake/Databricks
HS: Yeah, I will talk both ways. The opportunity for Dremio which SAP brings to the table is the great muscle where they have a reach to those customers which were probably unheard of or unthinkable for Dremio to reach. SAP knows the CIOs of any large Fortune 500 companies; they are speaking every day, but Dremio was not. So that is a big opportunity for Dremio to reach out to those customers and then pitch their product and propositions to acquire them along with SAP. Conversely, the opportunity for SAP, as I spoke earlier, is to basically increase their wallet share within their customer base because Dremio will have a very small customer base compared to SAP. But within that customer base, Dremio will provide an expansion opportunity because now they have a lever to combine or join the data which is sitting outside SAP. I feel that is a big massive opportunity and if I have to put a number on that today, probably even a customer who is a pure SAP shop, 30% to 40% of the data even sits outside of SAP which can be brought to the knowledge of SAP to understand the potential. So this definitely gives great expansion opportunities and kills those CSP natives or someone like Snowflake or Databricks. Of course, it depends on the kind of use case that they are implementing, but there is definitely an opportunity to go and pitch those customers to encapsulate within the SAP, especially those pet accounts that we call for the SAP ecosystem that they have.
来源背景：Harsh Singla，标题栏写 'Former Regional Sales Director - Dremio' ，但主持人介绍为前 Dremio APAC 企业客户经理 (约 2 年多)，现任 Confluent 区域销售总监约 3 年 ；更早在 Qubole。即离开 Dremio 约 3 年。视角局限：APAC 销售，离开 Dremio 已久；对 Snowflake 只有泛泛比较；Dremio 立场天然推开放湖仓、反数仓。 - 反对 Microsoft Fabric 在 RFP 中与 Snowflake 正面对比，客户对微软 AI 层感兴趣
FC: This current quarter, we have... let me think... we have about 8 projects in the pipeline that we are hoping to close on and working directly with Snowflake as well as our customers to refine what those projects would look like. So all in all, it's looking good. It's looking solid. We have a good pipeline. We just need to execute and kick these projects off and really win these. Yes, a couple of them are sort of in an RFP stage where they are evaluating and specifically from a hyperscaler perspective, Microsoft Fabric as a direct BB: Yep. FC: competitor to Snowflake as well as one that we're just hearing more and more as, hey, we need the power of Fabric. We also are really interested in the AI layer of Microsoft, and therefore we want to throw their hat in the ring and do a head-to-head comparison directly against Snowflake. So time will tell which will win, but I still think Snowflake has a very strong offering that should win the pot here.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 KeyBanc 称很多企业客户中 Snowflake 相对 Databricks 的战略地位在上升
KeyBanc says recent SNOW surveys were mixed ahead of earnings, with April consumption trends remaining healthy/in-line while forward spending intentions deteriorated meaningfully. The firm flagged weaker NTM partner growth expectations, potential F2Q guide risk (~24% vs Street 26%), and softer-than-expected Cortex Code adoption data versus management commentary. However, KeyB also notes Snowflake continues gaining strategic relevance versus Databricks among many enterprise customers, while core consumption trends into F1Q remain constructive, maintaining OW and $200 PT.
来源背景：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。 - 反对 大型金融客户优化导致消费低于计划
* “The main reason for the cut is one very large financial-services customer whose consumption is below plan as they go through optimization. This is offsetting some of the upside from Cortex Code and Snowflake Intelligence.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 反对 在只涉及数据工程和 AI/ML 的 greenfield 招标中，Snowpark 对 Databricks 胜率不高；Iceberg/开放格式是后续竞争因素
* Strong traction within the existing customer base * For complex AI/ML or MLOps requirements, Snowpark is not as strong as Databricks * In greenfield RFPs scoped only around data engineering and AI/ML, Snowpark does not have a strong winning chance against Databricks * Iceberg tables, open data formats, and other dynamics are upcoming competitive factors * “Snowpark is an extension of core capabilities — it lets you run native Python, Java, and Scala inside Snowflake, similar to what Databricks can do. Snowpark exists to compete with Databricks, ensuring Databricks does not encroach on or cannibalize Snowflake’s business.”
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 拥抱 Iceberg 让客户把非 SQL 工作负载留在 Snowflake 计算引擎上，Snowpark 增速升至 mid-50s 到 mid-60s
* Snowpark has grown in the mid-50s to mid-60s YoY over the past two quarters (vs. mid-40s previously) — a clear acceleration * “Snowpark is growing because when Snowflake decided to embrace Iceberg (about two years ago), there were widespread concerns that Snowflake was giving up storage. The reason Snowflake was willing to do this is that they wanted customers to focus on and leverage their compute engines. The net effect has been Snowpark expansion — your code continues to execute inside Snowflake, your non-SQL workloads (Java, Python, etc.) are handled by Snowflake. You don’t need to move the data. It’s like an Oracle database sitting in Azure — Oracle has interconnects with the CSPs, you don’t physically move the data, you execute it where the data lives. This is meaningfully improving Snowflake’s posture in the developer community.” * **Snowpark Connect is also a clear contributor:**
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 中性 Iceberg 增长但不是最快的产品，用例主要是与 Databricks 共存，属于联合销售的一部分
* Growing, but not the fastest-growing product * Use cases remain limited, primarily centered on Snowflake/Databricks coexistence * Iceberg is not a standalone solution; it is part of a co-selling motion
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 中性 CoCo 已扩展支持 Glue、Airflow、dbt Cloud 甚至 Databricks——CoCo 可在竞品平台上工作，可能帮客户在非 Snowflake 引擎上做事 (反例风险)
So CoCo is a general-purpose coding agent that has a set of features that are specialized for Snowflake and data platforms. We have published benchmarks that show that CoCo can outperform even the frontier models when it comes to doing operations within Snowflake. And over the past quarter, we've actually expanded it to support other data platforms like Amazon Glue or Airflow or dbt Cloud, and in fact even Databricks so it's incredibly powerful. And in terms of how it impacts our customers' ability, our ability, our partners' ability to get things done faster is any kind of coding transformation, and a migration is one such example, can be made faster with CoCo. We have a migrations team that is busy creating we call them harnesses.
- 中性 竞争问题上管理层强调云厂商合作、多云中立和模型选择 (Anthropic/OpenAI/开源)，未提 Databricks、Iceberg 或价格
We are also about a certain amount of independence from the mechanics of the cloud providers. A Snowflake implementation works fine on AWS, but it can also work on Azure. We have similar, really good partnerships with the leading AI labs, both Anthropic and OpenAI. We collaborate very closely with them to create great AI products, but also to create safe AI products. And similarly, Cortex Code and Snowflake Intelligence provide model choice. We run fine on both the models. We also host a whole series of other models ourselves, and as things like open source models become more important, we always act on behalf of what is right for the customer, which I think positions us in very good stead with all our customers. And I'll just add on to that. Just from a sales execution perspective within the quarter, the achievement was great in all geographies and all industry verticals.
- 反对 Funda 3Q25 看空的理由：Gen2 优化和 Snowpark Connect 同时造成负面影响，前沿实验室在优化 Snowflake 预算 (性能改进是收入逆风的历史例证)
* We were very bullish in 1Q26, and we think Cortex Code is taking in the spillover demand from Claude Code. Because Cortex Code hosts the Anthropic model by itself, it is faster than Claude Code. This is the fastest-growing new product in Snowflake’s history. * We were bullish in 4Q25, because the Gen2 optimization was close to finishing, and Snowpark Connect also turned from a negative impact to a positive one. * We were bearish in 3Q25, because we saw that Gen2 optimization and Snowpark Connect were both having a negative impact on Snowflake at the same time, and Frontier Labs was optimizing their Snowflake budget... * We have been bullish on Snowflake since 3Q24, and we believed that the new CEO was making the right and aggressive organizational changes to Snowflake.
来源背景：作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。 - 支持 Adaptive Compute 性能大幅提升但按收入中性定价，与 Gen2 同样处理，模型不用改
We also talked about adaptive compute. I put it in here just because there's something I want all of you to be clear on. Massive performance improvement but we're doing the exact same thing that we did with Gen 2. We priced it in a way that we're aiming for revenue neutrality. So our customers get the benefit that is materially faster but none of you need to go change your models. We're still good even though we're trying to push very hard for the adoption of all of this. Interactive analytics for responsive experiences.
- 支持 Snowflake 托管 Iceberg 存储 GA：保留开放格式同时把存储收入拿回来，称比原先建模的 Iceberg 影响更好
But one of the things that was factual was with Snowflake, you store in our format and you pay for a storage. With Iceberg, we always said it's customer managed storage but there's no reason for that trade-off. So we introduce and is generally available here at Summit, Snowflake managed storage for iceberg. So you can still be interoperable but we'll do the management of the storage. We'll give you economics. So, that I think is going to be in an even better tailwind relative to how we thought and modeled the adoption of Iceberg. Sharing – I think all of you know, I'm personally passionate about the network effect, personally passionate about the unsiloing of data and helping organizations connect. I am very excited that we're finally breaking out of the – it's two parties.
- 反对 Snowflake 主动开放：其他引擎可读写 Snowflake 上的数据，鼓励客户选最便宜/最好的引擎 (计算可以被分流)
Right now, we're the broadest in terms of the implementation of the V3 spec and we're steering the V4 spec. I mentioned this morning, we integrated all the rest catalog APIs into Horizon to make sure that we can interoperate with data regardless of where it sits even if it's on Databricks. We can read and write data of other engines with glue and others. They can read and write data that sits in Snowflake. So this whole notion of I'm locked in and I put my data like, that's not excuse customers. Please use whatever gives you the best experience or the best performance, the best economics. And okay, this one is super important for all of you in this room. When we introduce Iceberg, I think some of us regrets how much noise it cost.
- 反对 分析师当面指出主要竞争对手 (Databricks) 增速是 Snowflake 的两倍，低 30% 增速不够
Brian, I like your socks. Just I'v egone with the swag store. Sridhar, if all these products take off, low 30% growth doesn't really feel right. I mean, it feels like this is a 40% market. Your primary competitor is growing twice as fast as you. So, when you think about ultimately what you think these new solutions can do to help accelerate growth, I know you're not giving guidance here but it doesn't feel like you're at cruising altitude from where the rest of the industry is out right now. I don't know what to say.
- 支持 Datometry 虚拟化 Teradata 层产品化，加快 Teradata 迁出
I have scripts and applications and reporting systems and changes that that takes time. That's why some migrations Teradata have done are two years, three years. What this virtualization lets us do is, say, if this is Snowflake and these are the apps, we put a layer in between and that layer lets everything else in the enterprise environment think that it is Teradata. It looks like Teradata, it has Teradata SQL, it has Teradata scripts and Teradata is a really rich, functional product. To the rest of the ecosystem, it looks like Teradata. And what it's doing behind the scenes is translating into Snowflake. So, when we did the acquisition, converting it to support Snowflake took a few months. I don't know how long ago that was, six months or so. Now, we're ready to start accelerating migrations through this. So we're excited about the opportunity to help customers move off of Teradata quicker. Then Sridhar mentioned it, CoWork is super important. We do believe that it is the enabler to change how people go about their jobs.
- 反对 Databricks Genie 目前不如 CoCo 成熟，但正在集成更多 LLM，追平“只是时间问题”
不过，Databricks 也在积极追赶，正在其产品中集成更多的大语言模型。因此，双方能力达到不相上下的水平只是时间问题。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 反对 客户为省 Snowflake 存储费把数据放进 S3 上的 Iceberg 表，再用 Cortex 在上面建管道 (存储收入外流，计算仍留在 Snowflake)
目前，市场的一个主要趋势是转向 Iceberg。Iceberg 是 Snowflake 之外的一种大型外部表格式。客户现在正使用 Cortex 创建管道，在 Iceberg 的外部表上执行数据质量、数据集成和数据治理等操作。为了节省 Snowflake 内部的存储成本，客户选择将数据保存在 S3 存储桶中的 Iceberg 表里。在此基础上，客户利用 Cortex 创建数据管道、注入作业和流处理任务，这正成为我们观察到的一个新兴工作负载。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 中性 很多客户同时用 CoCo 和 Claude Code：Snowflake 内的任务用 CoCo，通用软件工程用 Claude Code
综上所述，最核心的决定因素是数据仓库的位置。如果用于大语言模型的大部分数据都存储在 Snowflake 中，那么 Cortex Code 是显而易见的选择。但对于通用的软件工程任务，一些客户会使用 Cloud Code。实际上，许多客户会同时使用这两种工具，并非非此即彼。关键的区别在于上下文感知能力以及我前面提到的几点原因。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 中性 纯编码能力上 Cortex Code 还比不上 Claude Code；数据工作流和 SQL 生成上 CoCo 更好
回 答 者：从我的观点来看，目前 Cortex code 在纯粹的编码能力上还无法与 Claude code 匹敌。如果以代码生成质量、推理深度、调试能力、对代码库的理解以及复杂的工程任务为基准，我认为 Claude code 目前仍具有优势。具体来看，在通用软件工程、复杂调试和多文件推理方面，Claude code 的表现优于 Cortex code。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 护城河不是编码模型，也并非不可逾越；如果 Anthropic/OpenAI/微软/AWS 做更深的集成，Snowflake 优势会缩小
当然，如果 Anthropic、OpenAI、微软、AWS 等公司未来开始构建更深度的集成，Snowflake 的优势将会减弱。所以我认为 Snowflake 的护城河并非不可逾越，但这条护城河的关键不在于编码模型本身。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 Databricks 增速明显快于 Snowflake，同比 >50%，在 AI 原生增长上占优
回 答 者：根据我看到的最新数据，DataBricks 的增长速度显著快于 Snowflake，其同比增长超过 50%。我的理解是，Snowflake 在数据云的采用方面处于领先地位，而 DataBricks 目前在 AI 原生增长方面占据优势。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 Snowflake 和 Databricks 越来越多通过 Iceberg 读同一份底层数据，不再需要物理复制
两个平台正越来越多地通过开放格式（如 Apache Iceberg）访问相同的底层数据。传统上，客户需要物理上复制大量数据，但我认为未来这将不再是必需的。这是否会造成操作上的不便？是的，但我认为相比 3 年前，不便程度已大大降低。当前最大的挑战在于跨平台的治理、成本管理、安全性一致性、数据血缘和技能复杂性。由于 Iceberg 提升了互操作性，实际的数据迁移问题正变得越来越小。Snowflake 和 Databricks 都支持开放表格式，客户也对多平台架构越来越适应。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 中性 竞争从二选一变成预算分配：Snowflake 拿分析/BI/治理/EDP 预算，Databricks 拿 AI/ML/agent/平台工程预算
关于它们是否争夺同一份预算，情况很有趣。目前来看，存在部分重叠和竞争。例如，在 EDP（企业数据平台）、分析现代化、数据仓库或 AI 平台的预算方面，它们之间确实存在直接竞争。但预算也在逐渐分化。我看到的情况是，Snowflake 主要获取的是分析、数据共享、BI、治理和 EDP 相关的预算；而 Databricks 则主要获取 AI、数据科学、ML、智能体（agent）开发和平台工程的预算。因此，我认为讨论的焦点正在从“选择 Snowflake 还是 Databricks”转向“投入多少预算给 Snowflake，多少给 Databricks”。举例来说，在 5 年前，一家大型银行发起数据仓库的招标邀请时，面临的通常是 Snowflake 与其他厂商的竞争局面。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 最大风险是 Databricks 把业务用户分析做简单，那 Snowflake 就脆弱了
我认为最大的风险并非来自直接竞争，而是未来某一个平台是否能在对方的优势领域也做得足够好。例如，如果 Databricks 能够显著简化商业用户的分析流程并降低使用门槛，那么 Snowflake 将变得脆弱。反之，如果 Snowflake 在机器学习和数据工程方面取得显著进步，Databricks 就会面临威胁。就目前而言，双方都还未成功跨越到对方的核心领域。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 反对 没有涨价；Snowflake 在提升效率，和客户的对话围绕怎么少花钱
其次，Snowflake 的发展方向恰恰相反，公司一直致力于提高效率并降低客户成本。过去几年，客户已经体验到更优的仓库效率、查询优化和工作负载管理。因此，与客户的沟通更多是围绕如何减少其在 Snowflake 上的支出，而不是解释价格上涨的原因。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 中性 通用能力上 Claude Code 公认强得多；在 Snowflake 内 CoCo 靠环境感知 (RBAC、schema) 胜出
回 答 者：问题的关键在于“在 Snowflake 环境下”这一前提。Cloud Code 是一款通用的人工智能辅助软件开发平台，在更广泛的平台和用例中，其能力公认要强大得多，这一点无可争议。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 反对 CoCo 里有 Claude 插件，用户可以在同一项目内切到 Claude；如果 Anthropic 移除插件，选 CoCo 会更难 (Snowflake 依赖模型厂商的配合)
如果OpenAI和Anthropic真想构成威胁，他们可以采取一些行动。例如，Cortex Code中有一个Claude插件，这意味着用户可以轻松地在Snowflake环境内使用Cortex Code，并在同一个项目中切换到Snowflake环境外使用Claude进行编码。这种灵活性使得用户在Snowflake环境中选择使用Cortex Code变得非常方便。因此，如果Claude的提供方真想制造麻烦，他们可以移除这个插件，那么用户选择使用Cortex Code的决策就会变得更加困难。但同时，我认为他们不会这么做，因为他们的目标是成为无处不在的模型提供商。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 反对 Databricks 在 AI 规模上约是 Snowflake 的 10 倍
在 AI 规模上，Databricks 遥遥领先于 Snowflake，大约是其 10 倍。这与两家公司的出身有关：Databricks 起源于数据科学和机器学习工作负载，而 Snowflake 最初是作为运行 SQL 分析的数据仓库。因此，这两个平台来自不同的发展方向。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 在他的业务里 Databricks practice 体量更大，但环比增速 6-7%，低于 Snowflake 的 10%
从我们自身业务实践来看，Databricks 的业务体量比 Snowflake 更大，但增长速度略慢。Snowflake 业务的季度环比增长率为 10%，而 Databricks 的增长率在 6% 到 7% 之间。当然，Databricks 的增长是在一个更高的基数上实现的。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 口径不齐 Iceberg 主要服务多云客户，纯 Snowflake 客户没有理由用；硬要估的话约占 consumption 的 10-15%
因此，我认为 Iceberg 只是整体生态系统的一部分，其本身并不直接贡献可观的收入，而更像是一个推动整体消费的赋能工具。所以很难给出确切的占比，但如果要我估计，它本身可能占到 10-15% 左右。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 他认为 Snowflake 定价权强 (切换成本高)，可能对传统产品涨价 8-15%，但目前没有看到迹象
回 答 者：我认为有可能。首先，由于客户的迁移成本很高，Snowflake 拥有强大的定价能力。其次，Snowflake 的价格策略通常与云服务提供商（CSP）保持同步，如果 CSP 涨价，他们很可能会跟进。目前我尚未看到任何明确的涨价迹象。他们推出的 AI credits 并非涨价，而更多是鼓励客户进行 AI 相关的消费。我认为他们可能会针对传统产品进行价格上调。根据我在其他案例中看到的经验，我预计涨价幅度可能在 8% 到 15% 之间。这个幅度足以对公司的营收和利润产生实质性影响，但又不足以高到引发客户迁移至其他平台。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 中性 Databricks 销售愿给前端打折，因为其商业模式核心是计算消费
**Expert:** If I was still in presales, I could probably make up some stories, but very, very limited, unfortunately. As Databricks especially have got the AI BI Genie and their Unity Catalog and their semantic layer, they've pretty much got everything you need, especially with the acquisition of Mozaic Data, where if you don't like the answer you're getting, you can always go in and start developing your own agentic solution. The same very much with Snowflake. Snowflake had a bit of a soft start with Snowsight. That's been deprecated. They're not too heavy on the analytics side. With the Cortex piece, with the semantic Views, which is a bit technical, with Snowflake Intelligence, I think it's called, they pretty much have an end-to-end. I know for a fact, Qlik do partner with both of those folks. I do know for a fact that there are conversations where the Databricks rep is thinking, "Well, hang on a minute, we can do all of this. We can then maybe discount the front end because they're all about using the compute." Unfortunately, I think the Qlik value proposition diminishes particularly strongly, not only in Snowflake and Databricks, but also within AWS as well with Glue, with Bedrock, with QuickSight and QuickSuite. They're pretty strong tools around analytics and the AI story. There's been a few customers, I know have churned to those tools.
来源背景：Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 之后重复一遍，以前半部分行号为准。 - 中性 Qlik 进入 Databricks 客户后，Databricks 随即推 AI/BI Genie 和 ETL 栈，Snowflake 同理
**Expert:** I'm not seeing double-digit growth. They partner very heavily with AWS, Databricks and Snowflake. In more than one customer where they went in with Databricks, they've already got a Databricks instance, not exactly. It didn't work out exactly like this. As soon as Qlik left the room, if you will, Databricks were talking about AI BI Genie and their AI stack and their ETL stack as well. That's going to be a challenge and the same for Snowflake. AWS has got Glue. Microsoft are everywhere. I'm not seeing a huge increase in ARR. The pressure is going to be coming obviously, from AI. AI can replace analytics with a chat agent. The trusted data piece, there's a lot of vendors out there now who can do that. There'll be a lot of pressure from those folks as well. I'm not really personally seeing a double-digit growth for Qlik going forward at all.
来源背景：Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 之后重复一遍，以前半部分行号为准。 - 中性 客户对 Snowflake/Databricks 最大担忧是账单不可预测，下行时可能偏好可预测订阅
**AS** [00:40:50]: That's my understanding. If you look at Databricks and Snowflake in isolation, then what customers really worry about is the unpredictable bill. I know the companies are working on building solutions that can help control that. They have done quite a good job. If there's an economic downturn and people are concerned about their bills, then I feel a personal opinion that they want to go with a solution that is more predictable than change often, and they know how to go revisit it every month.
来源背景：Ashish Shah，前 Cloudera 高级产品经理 (2016 年起在 Cloudera 9 年多，先后任销售工程师、ISV 合作伙伴管理 2018-21、产品管理，最后三年负责 SDX 治理组件，明示)；现任 Dataiku Technical Account Manager 。视角局限：Cloudera 本地/混合视角；对 Snowflake 的认识来自竞争丢单和朋友转述 ('I don't have any direct evidence')；离开 Cloudera 时间约 2025-2026。 - 支持 客户选 Snowflake 因易用、部署快；与 AWS 深度绑定，AWS 销售卖 Snowflake 获同等激励
客户选择Snowflake的原因是多方面的。首先，Snowflake的产品更易于使用，部署周期相对较短。其用户界面非常直观，用户通过简单的图形界面拖拽操作即可快速搭建一个数据仓库。其次，在销售策略上，Snowflake与AWS深度绑定。AW S的销售人员销售Snowflake产品可以获得与销售AWS自有产品（如EMR）相同的激励，这极大地扩展了其销售网络。此外，Snowflake在SQL方面进行了大量优化和创新，开发了许多自有的查询原语和UDF（用户自定义函数），例如“all- in-one”等功能，这些功能对于习惯其语法的用户而言非常便捷高效。因此，对于已经使用Snowflake的客户，迁移至Databricks是一个重大决策。尽管Databricks会为愿意迁移的客户提供包括迁移费用在内的投资支持，但迁移本身仍 存在难度。在从零开始选择的客户中，最终决策往往是案例化的，可能取决于客户对销售团队的偏好或对技术支持的评估，目前两家公司在市场上是直接的竞争对手，不存在绝对的优势方。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 支持 已用 Snowflake 的客户迁到 Databricks 是重大决策，Databricks 会补贴迁移费用但仍有难度；新选型案例化，无绝对优势方
客户选择Snowflake的原因是多方面的。首先，Snowflake的产品更易于使用，部署周期相对较短。其用户界面非常直观，用户通过简单的图形界面拖拽操作即可快速搭建一个数据仓库。其次，在销售策略上，Snowflake与AWS深度绑定。AW S的销售人员销售Snowflake产品可以获得与销售AWS自有产品（如EMR）相同的激励，这极大地扩展了其销售网络。此外，Snowflake在SQL方面进行了大量优化和创新，开发了许多自有的查询原语和UDF（用户自定义函数），例如“all- in-one”等功能，这些功能对于习惯其语法的用户而言非常便捷高效。因此，对于已经使用Snowflake的客户，迁移至Databricks是一个重大决策。尽管Databricks会为愿意迁移的客户提供包括迁移费用在内的投资支持，但迁移本身仍 存在难度。在从零开始选择的客户中，最终决策往往是案例化的，可能取决于客户对销售团队的偏好或对技术支持的评估，目前两家公司在市场上是直接的竞争对手，不存在绝对的优势方。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 口径不齐 单价上 Databricks 约为 Snowflake 两倍，但 Databricks 宣称性能带来更低 TCO；简单计算场景 Snowflake 更便宜
在单价方面，Databricks通常比Snowflake更贵。Snowflake按容量售卖，其一个core的价格约为0.3元；而Databricks的最小计算单元价格大约为0.55元。从表面上看，Databricks的价格几乎是Snowflake的两倍。然而，Databricks强调其性能优势可以带来更低的总拥有成本（Total
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 反对 Databricks 用 Unity Catalog 统一治理，Snowflake 有 Horizon 和 Open Catalog 两套 catalog，治理割裂
尽管产品方向趋同，但Databricks在三个关键领域具备差异化竞争优势。首先是统一的治理能力。Databricks的Unity Catalog能够对数据和AI资产进行统一治理。相比之下，Snowflake拥有Horizon和Open Catalog两个独立的catalog，这种割裂的治理体系源于其Horizon Catalog仅能处理Iceberg这一种表格式（table format），导致其治理能力存在局限。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 反对 POC 失败原因之一是竞争对手价格更有吸引力
POC失败的原因多种多样。一类是技术验证失败，即平台无法满足客户特定的业务需求。例如，曾有一个客户仍在使用基于Spark 3.0以前版本的传统Hadoop平台，而Databricks平台支持的最低Spark版本为4.1，巨大的迁移工作量导致了POC的失败。另一类原因是客户内部优先级的变化。此外，成本因素也是一个原因，即在技术方案相近的情况下，竞争对手的价格可能更具吸引力。
来源背景：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 - 反对 大型企业已有云厂商折扣协议，新采购 Snowflake/Databricks 需证明现有云数据库不行，很难
其中存在两方面挑战。首先，当企业已经拥有相关产品时，很难推动其采购新的产品。企业会质疑，为什么不能继续使用 AWS 中的数据库，而必须另外购买 Databricks 或 Snowflake。推动这类采购需要进行大量说服工作，而很多情况下并不具备充分的合理性。
来源背景：未具名专家，自述曾在泰国一家传统大型企业、一家覆盖英美的中型科技公司 (AI 非唯一业务) 和美国小型初创工作 (明示，)，职位未披露；中型科技公司经历中参与过 RAG 早期建设 。视角局限：样本为个人经历的三家公司；更偏工程/AI 应用，未直接谈 Snowflake 账单。 - 反对 年合同 1-2 万美元的小客户会因折扣/迁移优惠转向 Databricks，双向流动
不过，也有许多小型客户会迁移至 Databricks，因为他们可以获得折扣、更优惠的价格或 Databricks 提供的迁移优惠等。这些小型客户的规模通常在 10k 至 20k，也会在 Snowflake 与 Databricks 之间双向迁移。许多客户也会从 Databricks、Redshift、Fabric、ClickHouse 或 BigQuery 迁移至 Snowflake。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 TB/PB 级大客户迁往 Databricks 需花费数百万美元并有业务中断风险，通常不迁
已经使用 Snowflake 的客户通常认为没有必要迁移至 Databricks，因为两者性能相近，而 Snowflake 速度更快、成本更低。迁移本身需要完成一系列工作。对于拥有 TB 级和 PB 级数据的大型企业部署，迁移至 Databricks 可能需要花费数百万美元，同时还会带来业务风险和业务中断，因此客户通常不会迁移。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 口径不齐 Iceberg：FY26 Q2 超过 1,200 家客户采用，目的是降低锁定；无消耗占比，数据存在 S3/ADLS/GCS
回 答 者：FY26 Q2，超过1,200家客户采用Iceberg作为开放表格式，主要目的是降低供应商锁定。Snowflake并未公布Iceberg消耗增长的具体数据。从客户数量来看，年报披露的唯一信息是，仅Q2就有约1,200家客户采用Iceberg，因此其在Snowflake消耗中占比较大。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 Iceberg 曾被视为生存威胁 (计算引擎绕过 Snowflake)，但 Snowflake 接纳了它；未见大型受监管客户换成自建开源湖仓
回 答 者：我们主要服务于传统企业客户，因此我的看法可能存在一定偏向，但目前尚未看到开源技术带来显著影响。此前，Iceberg 及开放表格式一度被视为生存层面的威胁，因为计算引擎可能借此绕过 Snowflake，但 Snowflake 后来接纳了这一趋势。 在我看来，企业持续购买 Snowflake，不仅是因为存储能力，也因为其能够提供治理、安全等服务，并且企业不必自行运行基础设施。因此，目前尚未看到任何大型受监管客户使用自主管理的开源湖仓一体化技术栈替代 Snowflake。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 从 Snowflake 角度，Iceberg 是防御性举措；对 SI 是辅助模块，无占比数字
回 答 者：坦率地说，我没有确切的数据。对我们而言，Iceberg 更像是一个辅助模块。从 Snowflake 的角度看，推出 Iceberg 更像是一种防御性举措，旨在应对开源技术带来的竞争。对于本季度的具体占比，我无法提供准确数字。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 反对 大企业采购非常在意消费可见性、折扣和更长规划周期；按用量付费的不确定性是当前最大采购难题
在采购方面，我们的大型企业客户已经相当成熟。他们非常重视成本，也清楚在按使用量付费的合同模式下，客户很容易处于不利地位。因此，他们非常关注如何提高 Snowflake 使用量的可见性、获得折扣，并延长规划周期。合同中有不同的机制可以解决这些问题，但按使用量付费的模式几乎总会带来一定的不确定性。我甚至认为，这是目前采购面临的最大难题，而且不仅涉及 Snowflake，也涉及超大规模云服务商和模型公司等。在快速发展的 AI 领域，如何围绕按使用量付费的许可模式建立可见性和规划能力，是一个重要问题。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 支持 切换成本高：仅节省部分成本不足以迁出 Snowflake，数据越多锁定越强
回 答 者：这是一个非常有意思的问题。首先，我认为切换成本相对较高，尤其是Snowflake。一旦将数据迁移至Snowflake，企业就需要有充分的理由才能迁出并更换平台，仅仅节省部分成本并不足以构成迁移理由。迁移全部数据会带来运营风险，因此必须具备更加明确且实质性的收益。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 Databricks 增长更快但基数更小，竞争中互有胜负，没有 Databricks 总能胜出的趋势
回 答 者：Databricks也在快速增长，并且确实与Snowflake形成竞争。Databricks的增长速度更快，但其基数也更小，因此很难直接比较。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 支持 Databricks 已被放弃作为战略平台，只保留专业/遗留场景，增长会明显放缓
回 答 者：我们已经放弃将Databricks作为战略平台。Databricks可以继续用于一些专业化场景或遗留平台。总体而言，Snowflake能够承载的业务都会使用Snowflake；只有Snowflake无法承载的业务才会继续使用Databricks。因此，Databricks的增长会明显放缓。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 整合到单一平台既降 TCO，也提升采购谈判筹码 (客户议价力)
从全球角度看，我们希望实现 standardize 和 consolidate，尽量采用统一环境，减少 multi-window approach。将投资集中在一个平台，对降低总体拥有成本更有利，也能在 negotiation 和 procurement 方面获得更多 leverage。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 研发类高算力、自有模型应用放在 AWS 隔离环境，较少放 Snowflake
各领域都有一些成功案例。至于哪些环境部署在 Snowflake 上、哪些不在 Snowflake 上，情况有所不同。研发领域的一些应用会使用自己的模型，例如大语言模型和小语言模型。由于专业性较强，其数据环境和业务逻辑也不相同，因此通常会采用隔离的数据环境，可能部署在 AWS 上。AWS 的高端计算能力可能更强，因此这类应用较少部署在 Snowflake 上。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 Databricks 收入增速超过 20%
回 答 者：从收入来看，Databricks的表现也很强劲。我们确实看到其增幅超过20%，整体表现非常出色。无论是在市场层面，还是在我们自身的业务中，其增速均约为20%。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 支持 正面交锋 Snowflake 胜 60%、Databricks 胜 30%
A:在多数交易中 Snowflake 通常会获胜。在正面交锋时，Snowflake 赢得 60% 的机会，Databricks 赢得 30%。总体上 Snowflake 确实赢得更多对手的机会。
来源背景：MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。 - 支持 胜负取决于能力、领导地位、商业条件和集成；Snowflake 在云数仓和多云战略上居首，Databricks 紧随
这取决于多个因素：谁具备更多能力、谁在该领域是领导者、谁在商业条件上更有优势，以及谁的集成更好。就领导地位而言，Snowflake 位于顶端，尤其是在云数据仓库和多云战略方面。Snowflake 更能引起他们的共鸣，但 Databricks 紧随其后。
来源背景：MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。 - 反对 Microsoft 因 EA 承诺支出成为重要挑战者，Google 跟进，超大云厂商越来越难对付
我们还有其他竞争对手。由于 EA 协议，Microsoft 是一个重要挑战者——许多客户已与其承诺了支出。我们也看到 Google 成为企业环境中的第三大云提供商，最终他们也会像 AWS 一样有承诺支出。总体来看，超大云服务商变得非常具有挑战性。
来源背景：MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。 - 反对 Databricks 销售称在 agent 领域很少看到 Snowflake，正面竞争基本没输
在与Snowflake的正面竞争中，我们基本没有真正输过。在智能体领域，我们也很少看到Snowflake。我们看到许多Snowflake员工加入Databricks，这也说明了一些问题，反向流动的情况则几乎没有。当然，我们始终需要保持警惕。我尊重Snowflake，他们也有不错的产品，但我认为我们在这方面失去的项目并不多。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 反对 数据仓库正逐渐成为标准化能力，客户“在某个阶段逐步淘汰 Snowflake”更合理，迁移在增加
现在，数据仓库正逐渐成为一种标准化能力，需要被纳入完整的上下文体系中。因此，客户在某个阶段逐步淘汰Snowflake会更加合理。我们确实看到越来越多的客户选择Databricks，也观察到客户从Snowflake迁移至Databricks。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 支持 但从 Snowflake 赢客户周期长，重投入 Snowflake 生态的客户迁移成本高、通常不迁
回 答 者：确实如此。我们也会从Snowflake手中赢得一些客户，但这类项目的周期较长。随着生成式AI真正兴起，特别是智能体AI不断发展，以及上下文的重要性日益凸显，这种迁移趋势正在逐步增强。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 中性 传统数仓层双方各有胜负，现在 Databricks 策略不再争数仓层，而是争整个治理层
例如，在部分客户中，我们会与 Snowflake 竞争，围绕数据仓库展开讨论。我们曾多次赢得竞争，也曾有过几次失败。但目前的策略通常已不再是进入客户账户后争夺数据仓库层，而是围绕智能体、LLM 等能力赢得完整的治理层。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 中性 在独立治理层上 Snowflake 最接近 Unity Catalog，但仍不及
如果使用 AWS，情况基本相同，Google 也是如此。Snowflake 可能最为接近，但我认为仍未达到相应水平。目前还没有其他平台能够真正独立于各个平台，同时像 Databricks 一样充当独立的数据层。因此，在大多数情况下，Databricks 是最符合逻辑的选择。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 反对 Databricks 同比约 80%，Snowflake 约 35%，前者是后者两倍多
回 答 者：此前比较的是季度环比增速，但从整体同比增速来看，Databricks的增长确实更快。Databricks的整体同比增速为80%；以2025年第1季度和2026年第1季度相比，同比增速同样为80%，而Snowflake约为35%，因此Databricks的增速超过Snowflake的2倍。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 反对 去年近 200 家客户从 Snowflake 迁往 Databricks，Snowflake 前 6 大客户也在迁，主要是 AI/ML 与流处理
回 答 者：目前，客户迁移主要是从 Azure 或 Microsoft 平台转向 Databricks；第2类常见迁移是从 Snowflake 转向 Databricks，现场工程团队正在大量支持此类项目。这类迁移主要涉及 AI/ML 和流式处理用例。去年，我们看到接近 200 家客户进行此类迁移。事实上，即使是 Snowflake 的前 6 大客户，也在向 Databricks 迁移。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 反对 客户对在 Snowflake 上建 agent 有顾虑，Genie 比 Snowflake Agents 更准
其次，目前使用 Snowflake 的许多客户对智能体构建存在一定担忧，其实际效果不及 Databricks，尤其是在 Genie 的质量方面。与 Snowflake Agents 相比，Genie 能够提供更加准确的结果。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 反对 Snowflake 客户需维护多份数据副本，同等业务量在 Databricks 上花费更低
从数据重复的角度看，由于客户需要维护多份数据副本，Snowflake 的资源消耗更高。对于相同规模的业务，客户使用 Databricks 所需支付的费用更低，这也是客户迁移的关键原因之一。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 中性 Databricks SQL 相较云厂商数仓低 50-60%，自评 20+ 用例无一节省低于 40% (对比对象是云厂商数仓，非 Snowflake)
我们已经进行了大量评估。我本人评估过超过 20 个用例，没有任何一个用例的 SQL 成本节省幅度低于 40%。这是因为 ETL 移动更少、重复资产更少、治理模型更完善，并且具备无服务器模式。无服务器计算在这方面表现良好。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 中性 Databricks 数据仓库业务本身增速 40-45%，高增来自 Apps (>100%) 和 Lakebase (>95%)
从产品角度看，各产品的增速并不相同。其中，Databricks Apps的增速超过100%；事务型数据库Lakebase的增速超过95%；数据仓库等其他平台则保持40%至45%的常规增速。因此，这一判断是正确的，但指的是同比增速，而非季度环比增速。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 中性 Databricks 推动 3/5 年合同并给折扣、成功积分，客户提前续约——价格竞争工具
3. Databricks会向多年期合同客户提供专业服务、成功积分、折扣、阶梯定价及企业协议权益。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 中性 部分客户想同时保留 Snowflake 与 Databricks，专家认为会造成数据副本与管道冲突
1. 部分客户希望同时保留Snowflake和Databricks，以实现多平台部署。
来源背景：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 - 反对 Databricks 同比 80%，Snowflake 约 35%，前者两倍多
A:如果你和当前的增长比较，整体增长确实更快。如你所说，Databricks 的同比增长曾达到 80%，Q1 2025 对比 Q1 2026 的同比增长是 80%，而 Snowflake 大约在 35%，所以 Databricks 绝对是 Snowflake 的两倍多。从产品角度看，并非每个产品都表现出相同的增长。有几款产品增长更高：Databricks Apps 的增长超过 100%，另一个事务型数据库 LakeBase 增长也超过 95%。其它平台比如数据仓库的增长处于常规的 40% 到 45%。所以你的假设是对的，但那是同比增长，而不是环比增长。
来源背景：MeritForum 标题明示“databrick离职专家”，作者 Lantian ；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering,)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。 - 反对 去年约 200 个客户从 Snowflake 迁移，前六大客户也在迁
A:我们现在看到的前三类迁移主要是从 Azure 或 Microsoft 到 Databricks，其次是从 Snowflake 到 Databricks。Snowflake 到 Databricks 是字段工程工作中非常常见的迁移，主要用于 AI、ML 和流处理用例。我们去年看到大约 200 个客户迁移。即便是 Snowflake 的前六大客户也在迁移到 Databricks。我认为有两点原因。
来源背景：MeritForum 标题明示“databrick离职专家”，作者 Lantian ；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering,)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。 - 反对 Snowflake 消费更高因需维护多份数据副本
首先，我们有一个非常好的工具叫 LakeBridge，支持从 Snowflake 到 Databricks 的评估和转换，因此迁移更容易、摩擦更小。现在使用 Snowflake 的大多数客户对构建 agents 有顾虑，Snowflake 上的 agents 表现不像在 Databricks 上那样好，尤其是谈到 Genie 的质量时。Genie 的结果比 Snowflake 的 agents 更准确。另外，从数据重复的角度看，Snowflake 的消费对客户来说更高，因为他们必须维护多份数据副本。如果客户在 Databricks 上有相同的数据量，其支付成本更低，所以这是客户迁移的关键原因之一。
来源背景：MeritForum 标题明示“databrick离职专家”，作者 Lantian ；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering,)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。 - 反对 微软 FDE 可能屏蔽 Snowflake、向客户追加销售 Azure 竞品
回 答 者：他们甚至可能会屏蔽 Snowflake，也可能发现 Azure 上存在竞争，并向客户追加销售，这一点目前无法确定。
来源背景：微软前沿部署工程师 (FDE) 团队成员 (纪要明示其团队直接服务 14 家 FDE 客户、可观察约 70 家，)，现任，职级与年限未披露。视角局限：微软立场，倾向推荐 Azure 体系；对 Snowflake 的判断是推理而非账单数据。 - 反对 Databricks Q2 大幅促销，为抢 CoCo 吸引的用户对 Genie 打折，属低价竞争策略
Databricks在Q2推出了一些大幅度的促销活动。尤其是Q1推出了Genie，目的是吸引更多用户，也属于低价竞争策略。Q1的Genie表现很好，Snowflake推出了CoCo（即Cortex Code），迅速吸引了大量用户，官网显示有7000多个。Databricks也希望通过折扣抢占这部分用户，因此收入会有所减少。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 反对 Genie 原本每分钟 20 个问题免费，Q1 开始收费，5 月调整为半价，未纳入销售目标
Mosaic AI是一个产品线，Mosaic是Databricks收购的一家公司。看 Mosaic 的话，它更多还是传统的 machine learning。Genie因为在打折，暂时没有实质收入。Genie之前是每分钟内20个问题免费，所以客户一般都控制在一分钟内， Q1开始收费，5月份又调整为半价，因此 Genie现在没有被纳入销售目标，也没有统计相关数据，应该收入还是比较少。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 支持 没有看到 Snowflake 客户整体迁往 Databricks，大企业两者并存
目前并未出现Snowflake客户直接迁移至Databricks的明显趋势。在许多大型企业客户中，Snowflake和Databricks本身就同时存在。客户可能使用Snowflake处理某些工作负载，同时在其上使用Databricks。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 反对 新增负载的争夺在加剧，Databricks 在金融、医疗、制药取得进展
在一些更加重视分析能力的场景中，Snowflake与Databricks对新增工作负载的竞争正在加剧。尤其是在金融、医疗和制药等数据量更大、数据类型更多的行业中，Databricks正在取得一定进展。对于新增工作负载和新项目，Databricks正与Snowflake展开激烈竞争。不过，Snowflake客户完全弃用Snowflake并迁移至Databricks的情况，我认为目前并未在市场中出现。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 反对 新增平台预算 Databricks 约 60:40 占优，且缓慢扩大
BJ: I think it's been pretty split. Honestly, I would say net new is 60-40, Databricks net new versus 55-45, widening slowly, but not where me and you would automatically default to. And it's not because Databricks is winning head-to- head against Snowflake or so. It's that the net new work skews towards engineering heavy agent builds and streaming, which is where Databricks started from the beginning, right? And where the work is more governed analytics and business facing insights and agentic associated with the Snowflake still takes it comfortably. And then I think vendors are calling it out as well, in terms of their growth rate accordingly, right? And then, respectfully, frankly, the architecture is now equal between both of these vendors. So it's just how one got in an account and then how it stayed there and then who the buyer persona is. That's fundamentally where it's going for. But if you think about, again, agent builds, if it's purely engineering, analytics, custom app, Databricks. If it's more BI and business-facing
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 口径不齐 60:40 源于新工作偏工程/流处理，且 Thoughtworks 样本偏工程；全市场他按 50-50 建模
Honestly, I would say net new is 60-40, Databricks net new versus 55-45, widening slowly, but not where me and you would automatically default to. And it's not because Databricks is winning head-to- head against Snowflake or so. It's that the net new work skews towards engineering heavy agent builds and streaming, which is where Databricks started from the beginning, right? And where the work is more governed analytics and business facing insights and agentic associated with the Snowflake still takes it comfortably. And then I think vendors are calling it out as well, in terms of their growth rate accordingly, right? And then, respectfully, frankly, the architecture is now equal between both of these vendors. So it's just how one got in an account and then how it stayed there and then who the buyer persona is. That's fundamentally where it's going for. But if you think about, again, agent builds, if it's purely engineering, analytics, custom app, Databricks. If it's more BI and business-facing agents, Snowflake. For that reason, 60-40 right now. But if I were to bet, I would split majority of my business is gonna, and again, I'll flag that, you know, at ThoughtWorks, the nature of my business is not exactly the same as what it would have been in my Accenture days or Deloitte days. I do get more complex engineering work, which other vendors may not go after. When I was running that practice, I wouldn't go after that business. So it's just nature of the brand equity and what kind of adjacent offerings and service lines we have, right? We do have a lot more of that here. So that results in a lot more Databricks business for me. If I was to zoom out, I would model it at 50-50 and instead of having two separate practices, I have basically merged into one just so that we can maximize our internal people capabilities.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 两家架构已经等同，胜负取决于谁先进入账户和买方画像
BJ: I think it's been pretty split. Honestly, I would say net new is 60-40, Databricks net new versus 55-45, widening slowly, but not where me and you would automatically default to. And it's not because Databricks is winning head-to- head against Snowflake or so. It's that the net new work skews towards engineering heavy agent builds and streaming, which is where Databricks started from the beginning, right? And where the work is more governed analytics and business facing insights and agentic associated with the Snowflake still takes it comfortably. And then I think vendors are calling it out as well, in terms of their growth rate accordingly, right? And then, respectfully, frankly, the architecture is now equal between both of these vendors. So it's just how one got in an account and then how it stayed there and then who the buyer persona is. That's fundamentally where it's going for. But if you think about, again, agent builds, if it's purely engineering, analytics, custom app, Databricks. If it's more BI and business-facing
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 支持 BI/面向业务的 agent 更易治理，风险与安全买方在场时对 Snowflake 有利
BJ: Yes, blurring of the workload, but less important when it comes to agents, right? And at the same time, a third buyer persona is also in the room when it happens to be risk and security when it comes to agents and anyone that can credentialize and manage that governance layer faster. And I think that is where Snowflake's current advantage of business BI orientation in terms of workloads works out better because those workflows are easier to govern and easier to defend. And then those are getting improved at scale. So it is leaning in their favor. With that being said, if your use case is really—and I say this because I have 1 client, a large bank that is currently engaged with us to drive a large conversational AI platform, not a chatbot, an actual conversational agentic platform where you say, pay my bill, it will know which bill and it will pay from the account that you have already specified. That becomes more custom engineering, right? And that is where the interface layer is actually Databricks. So it all depends on where the buyer is and what kind of workflows you are running with. Without being said, that is an edge case, very complex, $1 billion worth of business casing was done on it. The rest of the world is still more BI insight oriented and in that case Snowflake wins.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 成熟 CIO 不会押单一供应商，存储与建模成为 table stakes，护城河向 context/权限转移且 Atlan 等也在争夺
By the way, one thing I'll say, none of my clients are... and I'll say because of how the costing has happened on the modeling side... it will be naive for any mature CIO or CDO or CTO to anchor their future on one vendor at the moment. So we are seeing a lot of demand for composable architectures, modular by design, multi-model, multi-cloud, on-prem to cloud from the get-go. That means that the modeling and the way the data is stored is just table stakes, right? The context and the permissions might be where the moat needs to move towards, but then you also have other players like Atlan and others that want to own the context layer and the governance of it. The cake is getting sliced in multiple ways now and different players have their own point of views to take it away from one vendor and make control plane
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 中性 两家目前增长相当，没有赢家
being the main moat. Now, I say this, but I also see from a go-to-market perspective Databricks is absolutely saying, we want you, and again, their buyer persona is more technical in nature, the control plane, right? They want to basically position themselves as a control plane now. Snowflake, on the other side, from just my observation, is going after the business buyer. Business buyer is also very much worried about the governance and grounding and evaluation because if it doesn't work, they see the impact in their business immediately. And Snowflake has an advantage because in their framing, they have always been the choice of the platform for insights, right? So, I feel like Cortex is built for the interface with the business, and the Databricks Unity AI gateway is more of like the technical leader who understands the nuances associated with agents and the governance associated with it. I think that's where both of them are doing exactly what they do with their buyer personas and ICPs. However, I wouldn't call... I think they're both growing equally respectfully at the moment. I wouldn't call anyone a winner yet. David, do I still have you?
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 支持 AWS 和微软补贴更激进，但客户大多没接受，只有 1 个客户因已有大量 AWS 工作流而迁移
BJ: I think both Amazon and Microsoft have been more aggressive. What's interesting, at least from my vantage point, is that my clients are not just falling for it at the moment, which is interesting, right? Because I have 1 client that did opt into it, but there was a lot more adjacent kind of workflows that were already within AWS to warrant that decision. There's a lot more understanding and maturity that I'm noticing, more disciplined, where the clients are really thinking through their architectures and which vendor do they go to or not. And then the goal is to be more neutral architecture by design.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 采购端会和 Snowflake 争取更好条件：捆绑、折扣、多年期承诺
其次，从采购角度来看，企业可以尝试与Snowflake争取更有利的条件。采购部门总体上已经变得更加成熟和专业，也会重点关注IT支出中的大额成本项目。正如前面提到的，这将日益成为企业未来的重要支出领域。因此，采购部门会研究如何提高合同效率，例如采用捆绑采购、折扣和多年期承诺等方式。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 成本足够大后，企业会在 Snowflake 旁边部署开源数据湖、降低数据检索成本，2-3 年后成为重点
因此，企业需要围绕这一点进行设计。如果企业采用Snowflake战略，并将大部分数据存储在Snowflake中，实际上很难绕过它。不过，鉴于成本将非常高，企业一定会重点考虑如何提高效率。我完全可以预见，企业会在Snowflake附近部署开源数据湖，并通过不同方式降低数据检索成本。这种情况一定会发生。虽然这可能不是当前的重点，但随着整体使用量大幅增加，它将在2至3年后成为重点之一。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 支持 数据大部分在 Snowflake 的企业，AI 很难绕开它
因此，企业需要围绕这一点进行设计。如果企业采用Snowflake战略，并将大部分数据存储在Snowflake中，实际上很难绕过它。不过，鉴于成本将非常高，企业一定会重点考虑如何提高效率。我完全可以预见，企业会在Snowflake附近部署开源数据湖，并通过不同方式降低数据检索成本。这种情况一定会发生。虽然这可能不是当前的重点，但随着整体使用量大幅增加，它将在2至3年后成为重点之一。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 中性 开源替代目前进展有限，多供应商策略会增加
此外，企业通常还会尝试引入开源方案。虽然这种情况未来可能出现，但目前尚未看到太多进展。企业也可以采用多供应商策略。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 支持 没看到客户在主要数据平台之间迁移
回 答 者：从我观察到的情况来看，目前没有太多客户在这些平台之间迁移的项目，客户通常会继续使用现有平台。不过，随着业务规模扩大，企业通常会购买额外容量，或承诺增加消费量。
来源背景：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 - 反对 Databricks 也在让其 agent 连接 Microsoft 365、Jira、Google Drive 等外部源
Databricks 也在朝这个模式发展，允许其生态系统中的 agent 连接到 Microsoft 365、Jira 和 Google Drive 等外部源。对 Snowflake 和 Databricks 来说，这一含义很重要：AI 的增长并不必然按一比一转换为数据平台的消费。这些平台必须通过成为治理上下文和数据的那一层来赢得在 AI 架构中的位置。如果它们不是该架构中的治理层，就不能假定每一次 AI 查询都会通过它们。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 中性 4-5 月 Document AI 和 Iceberg V3 发布让可读数据范围变大，对本季度有一定带动
我们也关注到，部分买方对整体情况的假设更加乐观，但我们没有作出特别激进的假设，主要有以下原因。整个4到5月，Snowflake产品侧的更新主要以文档AI和Iceberg V3的推出为主。这些更新对业务具有一定带动作用，因为企业可读取的数据范围变得更大。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 大中型客户迁出 Snowflake 架构成本很高基本不切换，小型/新成立公司更愿切换
如果要将Snowflake的整体架构迁移到data，对客户而言将是一笔非常高的开支。因此，原有的大中型客户基本不会进行切换。但对于很多小型客户，尤其是刚成立的企业，由于过去没有大量数据积累，在切换方面反而会相对积极。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 反对 增长弹性 Databricks > Snowflake > MongoDB，Databricks 先受益于数据治理
从增长弹性来看，Databricks优于Snowflake，Snowflake又优于MongoDB，这与其产品形态有关。企业AI变革首先需要进行数据治理，其中涉及大量基础设施建设，因此Databricks会率先受益。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 反对 AWS 等云厂商在减少对 Snowflake 的销售资源，转向 Databricks 和 Anthropic
1. Customers are increasingly using Snowflake to consolidate data, and Snowflake is taking share from platforms such as Google BigQuery, AWS Redshift, and Microsoft Fabric. 2. However, we also note that CSPs such as AWS are reducing sales resources dedicated to Snowflake and allocating more sales resources to Databricks and Anthropic, which could affect Snowflake sales going forward.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 客户把数据归集到 Snowflake，从 BigQuery/Redshift/Fabric 夺取份额
1. Customers are increasingly using Snowflake to consolidate data, and Snowflake is taking share from platforms such as Google BigQuery, AWS Redshift, and Microsoft Fabric. 2. However, we also note that CSPs such as AWS are reducing sales resources dedicated to Snowflake and allocating more sales resources to Databricks and Anthropic, which could affect Snowflake sales going forward.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 中性 Databricks 与 Snowflake 多数客户同时使用，不完全可替代；竞争主要影响新工具和扩张工作负载的选择
1. Databricks was historically used primarily by well-resourced customers looking to do more data science and develop new models. Databricks and Snowflake have largely been complementary. Both are now building broader capabilities, but Databricks deployments can be more complex, and its architecture may be overengineered, while Snowflake is easier to use and more developer-friendly. 2. Both are entering the operational data storage market, while Databricks is expanding from data lakes and Delta Lake into data warehousing. Snowflake has more customers, although Databricks may have much higher spend per customer. Both should continue to coexist and may compete for share within large customers, while other customers will continue to keep their respective use cases separate. 3. Databricks has strong momentum and better partnerships with ERP vendors. Snowflake initially had closer ties with AWS but is expanding into Azure and GCP to support multicloud adoption. I do not believe they are fully interchangeable when customers use both. Competition will have more impact on which vendor customers choose for new tools and expanding workloads, causing one vendor’s growth rate to increase relative to the other.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 Databricks 动能强，与 ERP 厂商合作更好；单客户支出可能远高于 Snowflake
1. Databricks was historically used primarily by well-resourced customers looking to do more data science and develop new models. Databricks and Snowflake have largely been complementary. Both are now building broader capabilities, but Databricks deployments can be more complex, and its architecture may be overengineered, while Snowflake is easier to use and more developer-friendly. 2. Both are entering the operational data storage market, while Databricks is expanding from data lakes and Delta Lake into data warehousing. Snowflake has more customers, although Databricks may have much higher spend per customer. Both should continue to coexist and may compete for share within large customers, while other customers will continue to keep their respective use cases separate. 3. Databricks has strong momentum and better partnerships with ERP vendors. Snowflake initially had closer ties with AWS but is expanding into Azure and GCP to support multicloud adoption. I do not believe they are fully interchangeable when customers use both. Competition will have more impact on which vendor customers choose for new tools and expanding workloads, causing one vendor’s growth rate to increase relative to the other.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 Iceberg 和 Snowpark 被视为防御性产品，逐渐成为配角
1. It contributes to overall growth but is more of a defensive product. Marketing efforts have become less intense, and it is gradually maturing. Development work continues, including integrations with certain ERP systems in key regions. Strategically, however, Snowpark and Iceberg are gradually becoming supporting components rather than core parts of the Snowflake story. 2. More agentic and AI workloads will drive consumption, and agentic workloads almost inevitably involve unstructured data, which benefits Iceberg. The logic holds, but I have not seen data supporting it, and I do not think it will fundamentally change Iceberg’s growth. Iceberg remains a supporting product rather than a primary growth driver. 5. Will Any New Products Drive Growth in 2H26?
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 中性 SaaS 厂商变数据库并非新趋势，CIO 不会把数据复制到三个平台；Snowflake 与多家做双向 zero-copy 合作。全程未提 Iceberg/Databricks
Snowflake is only a part of the overall software estate that our customers have. We offer interoperability at multiple levels, but we think our flagship products are very important to our future. Yeah, I'll add maybe that the notion of some of these application providers becoming database players is not a new trend. And what we hear consistently from CIOs and CTOs is if I use three applications, I'm not going to copy my data into three different platforms. It's easier to consolidate in a single central platform like Snowflake, which is why we have bi-directional zero-copy partnerships with many of them, and we see a lot of customers aligning their data estates with Snowflake. Yeah, and our investments in which Christian has pioneered and spearheaded with the team for a very long time around being able to host applications in Snowflake, small and big, also positions us exceptionally well for many applications, not just analytic ones, but also systems of record, operational ones that can be built right on top of Snowflake.
- 中性 CEO 承认 Snowflake 只是客户软件资产的一部分，必须'play nice'提供多层互操作
So, I've been very, very consistent for now two plus years in my conviction, in our convictions that owning the user experience is critical. And we see CoCo and CoWork as fundamental to our future because they demonstrate to us and to our customers what is possible. But on the other hand, we understand that we live in a world where we have to play nice. Snowflake is only a part of the overall software estate that our customers have. We offer interoperability at multiple levels, but we think our flagship products are very important to our future. Yeah, I'll add maybe that the notion of some of these application providers becoming database players is not a new trend. And what we hear consistently from CIOs and CTOs is if I use three applications, I'm not going to copy my data into three different platforms.
- 反对 CoCoEvolve 可迁移 Databricks/dbt/Airflow 工作流到 Snowflake；Databricks 也推出了迁移 agent 反向迁移 (AI 降低双向迁移成本)
关于Snowflake的CoCoEvolve，我觉得它是一个比较特别的产品。原本有CoCo，现在又有CoCoEvolve。CoCoEvolve可以自学习，把agent的工作变得更智能，能够理解业务并提升质量。我认为每个产品其实都有这个目标，比如Databricks的genie one，Databricks也在不断提升智能程度和自学习能力。如果只是为了这个目标推出新产品，我觉得有点奇怪。CoCoEvolve有迁移功能，可以迁移Databricks、dbt、airflow等系统的工作流到Snowflake上，并立刻创建业务context，帮助客户迁移和落地。这不仅仅是像Claude Code那样完成特定任务，而是带着Snowflake的清晰意图，专门为Snowflake做事。Databricks很多客户喜欢拿Genie Code和CoCo比较，但Databricks的销售策略是让客户用Agent Bricks和CoCo或CoCoEvolve对比。Genie Code非常聚焦于执行具体任务，没有data intelligence的能力，而Agent Bricks特别是有了Omnigent后，是在客户的商业场景和context下帮助完成任务。针对 CoCoEvolve，Databricks还是希望用 Agent Bricks 与其竞争。Databricks也推出了 migration agent，可以迁移到 Databricks 上。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Databricks 现任员工，负责 GCR/APAC 区域 (给出 GCR/APAC 约八成项目仍在 PoC，了解内部 Omnigent '零号用户' 使用情况与销售策略)。视角局限：竞争对手内部人员；对 Snowflake 是外部观察 ('不太确定 Snowflake 的情况')，尤其是中国市场的 Snowflake；样本为亚太。 - 反对 CEO 直说：迁入变快，迁出也一样变快
We've also talked in the last year about things like standardization of data tables and data graph is becoming flat, grave (00:30:00), for lack of a better word. How do you think about the longer-term implications from switching costs potentially going down in the data infrastructure world? I mean, just bluntly, I told my teams, I'll admit this to you, if migration into Snowflake can be made a whole lot faster, migration out of Snowflake can be made a whole lot faster. That's the world we live in. And it applies to data platforms. It applies potentially to consumer software.
- 反对 Snowflake 托管 Iceberg 表让其他引擎可查询，价值必须往上游 (治理、DR、agent、可观测) 移
So we support open format. We want to increasingly make it painless for our customers to deal with open formats. So we have now something called Snowflake-managed Iceberg tables, which is a fancy way of saying, you cannot have your cake and eat it, too, which is you can store data with Snowflake but have it be queryable in Iceberg format by any other engine. And the place where we create value has to be further upstream. It has to be in, do we provide better governance? Do we provide better disaster recovery? Is it easier to create and run agents on top of Snowflake? Do we provide a better observability solution? And so, there is this whole stack of things on top of the data platform that is open that we have to be providing. And to a large extent for a lot of these, my attitude is bring it on.
- 口径不齐 Gen2 等性价比改进已计入定价，被量和新作业抵消，没有大幅收入下台阶 (未给 6% 等数字)
Maybe, Brian, just tell us a little bit about how that philosophy is coming into play in 2026 with things like Gen2. Yeah. I mean, part of the business is you got to be better from a price performance perspective than the last generation was. And you always have to give your customers the ability to get more out of your product. And so, we price that in, but we're seeing the offset of that with volume and new jobs coming into the company. And so, we carefully – as we go and price stuff, we carefully take that into consideration. So there's not big step downs in revenue, but we also want to get our customers get the benefit and the value of these performance enhancements that we're doing on the platform. I'll perhaps add on a quote from one of my previous bosses and mentors. Revenue solves all known problems.
- 反对 500 毫秒以下新鲜度的数据目前做得不好 (实时 agent 场景短板)
Fast forward to now, you're not going to hear that from any of our customers. Not only is the offering really good, we have also, thanks to technology like CoCo, massively accelerated the process of migrating, let's say, from whatever set of notebooks that you have on to Snowflake or being able to create not one but dozens of machine learning models as part of an experiment that you are running. Making sure that we deal with data that, say, has 500 milliseconds or less of freshness requirement is not something that we do really well right now. There's a team that's actively at work on this. It basically comes down to things like what are the trade-off that you want to make in terms of cost and efficiency.
- 反对 Snowflake 中国区 SE 承认 Snowflake 过去不开放，近几年转向开放是被激烈市场竞争推动的
如果了解 Snowflake 的历史，可能会知道 Snowflake 以前并不是非常开放的平台，但近几年发生了非常大的转变。这也得益于激烈的市场竞争和技术进步，由此推动了技术与平台层面的质变和迭代。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 反对 Snowflake 的开放愿景是'把计算选择权交还客户'，并承认如果 use case 和性能没有优势就会被淘汰——计算环节变成按性能竞价
Snowflake 希望传达的愿景是打造开放平台，将计算的选择权交还给客户。我们需要在客户的 use case 和相应性能方面保持优势，否则就一定会被淘汰。这也是实现双赢的选择，也是开放如此重要的原因。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 反对 Polaris 已支持 9 个外部引擎读取、现在又支持外部引擎写入——工作负载可以在不搬数据的情况下转到别的引擎
2024 年，Polaris 已经捐献给 Apache 社区，目的是为客户提供更多选择。到 2025 年，我们发布了 Polaris Catalog，并将其集成在 Polaris 和 Iceberg 的 RAF 上。随着时间推进，我们支持了 9 个外部引擎读取 Polaris Catalog，现在又支持外部引擎写入 Polaris Catalog。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 反对 Snowflake 明确宣传客户可以借开放架构迁移甚至彻底离开 Snowflake，没有 vendor lock-in
Snowflake 的开放架构建立在 Apache Polaris、Iceberg 和 REST Catalog 之上。Apache Polaris 由 Snowflake 于2020年捐献给 Apache 基金会，并于去年成为 Apache 基金会的顶级项目。它能够真正实现双向互操作，不存在 vendor lock-in，并提供开源、开放的能力。任何第3方 Lakehouse Engine 都可以无缝接入。即使未来因技术迭代需要调整，企业客户也可以在必要时选择迁移，甚至彻底离开，从而避免任何 vendor lock-in。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 口径不齐 演示：Snowflake 通过 Catalog Link to Database 直接读写 Databricks 管理的 Iceberg 表，Spark 也能写同一张表并改 schema，全程无 ETL、同一份数据——互操作是双向的 (Snowflake 也能吃 Databricks 上的数据)
通过这种开放的数据架构，不同平台和引擎之间不再存在数据孤岛，数据流动和协同变得非常高效，因为整个过程中不存在任何 ETL，所有平台操作的都是同1份数据。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 支持 Snowflake 的防守点：外部引擎查询经 Horizon Catalog 时，ABAC/RBAC/脱敏策略仍强制执行——治理层留在 Snowflake
在开放的数据云平台中，最难解决的问题并非管理Snowflake内部的数据，而是在外部引擎调用相关数据时提供保护。无论采用ABAC策略，还是Snowflake传统的RBAC策略，都会通过Catalog进行强制执行。因此，当外部引擎查询表时，获取的是经过策略过滤的数据。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 支持 Snowflake 的论据：纯开源自建初看便宜，但要规模化需要数十到上百人的团队运维——这是托管平台的价值所在
由于历史原因，很多企业可能做出了并不理想的选择。第1种选择是完全依靠开源技术搭建自己的数仓，但搭建过程，尤其是运维过程，会变得非常复杂，治理也较为分散。在行业内实践多年后，很多人已经形成共识：纯开源方案起初看似便宜，但如果要实现真正的企业级应用和大规模推广，运维成本会非常高。如果没有数10人乃至100人以上的团队，就很难成功推广。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 反对 AWS 讲者强调企业现实中不可能 all in 单一平台，会同时用其他数据厂商，Snowflake 通过 federation/catalog 与之结合
回 答 者：可以先看方案，也可以通过 Snow 的 catalog，以及其他 catalog 进行选择。从企业角度看，现实中不可能全部 all in，因此企业也会选择其他数据厂商或数据方案。通过 federation 和 catalog 技术，可以将 Snowflake 与其他平台结合；即使涉及迁移或使用其他引擎，也可以便捷地整合在一起。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 反对 AWS 讲者提出数据可留在 S3 上以控制成本，再按需选引擎——云厂商在推'存储留在自己这、计算可选'
如果数据仍然希望保留在S3上，并考虑成本，已有相应选择。如果今天有提升需求，同时还需要其他特性，也有相应选择。过去，我们还会将加工后的数据放到不同系统上，今天仍然可以实现。因此，它在表类型及快速处理方面，已经超出了原来的考虑，能够解决原来遇到的困难。这也是今天想重点分享的内容。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 中性 Data Cloud 与 Snowflake 经 Zero Copy server-to-server 直连，不会消耗太多 credit/API 额度
在这种情况下，客户不愿意承担额外费用。Anthropic采用server-to-server方式调用，或者Data Cloud与Snowflake通过server-to-server方式实现Zero Copy，都不会消耗太多Credit或API调用额度。
来源背景：未具名专家，纪要无正式简介。内容围绕 Agentforce 渗透率、credit 计费、POC 免费 10-20 万 credit、实施项目经验 ，推测为 Salesforce 内部或 Salesforce 实施伙伴的 Agentforce 交付/解决方案人员。视角局限：Salesforce 视角，对 Snowflake 只是顺带；所称 'Snowflake、Databricks 增速 20% 以上、收入 70 亿与 67 亿' 与公开数据不符 (Snowflake 产品收入增速 30%+)，可信度低。 - 中性 AI 已商品化 (Gemini Enterprise、Bedrock、Copilot)，差异化在数据基础——同时也意味着超大云在提供同类能力
AI 本身已经商品化。GCP 提供 Gemini Enterprise，AWS 有 Bedrock，Azure 提供 Copilot 和 Azure Machine Learning Studio。这些平台都提供功能大体相当的 LLM。关键的差异化是数据基础：数据驻留在哪里，AI 在哪里被实际使用。因此数据基础是客户价值的核心来源。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 Databricks 在非结构化数据上更原生，Snowflake 在 AI 之前把大部分数据当作结构化数据
PM: Yeah, I think when it comes to Snowflake versus Databricks, the biggest thing I would say why, of course, Snowflake has a similar product, but one of the big trade-offs is dealing with unstructured data. And again, I know Snowflake can deal with unstructured data, but not as easily as Databricks did because Databricks, the database itself, the files in object store can be exposed directly to the customer. So I guess that's an additional way you can think of classic ML help. Like before, when you had the classic ML, you want to classify images, you could do it in Databricks because you just copy unstructured images into an S3 bucket or ADLS, and you can expose that to a notebook and train the data. So now, similarly, now you could put textual data, multimodal data into an object store and use it to fine-tune the model or use it to do a batch load of unstructured data to let it annotate it with LLM. So some of this functionality around how easily Databricks integrated unstructured data, both from a governance, accessibility, and treating as a first-party object just like a structured data, made that a bit of an advantage over Snowflake, which treated most data prior to AI as a structured data only. And again, I know that they can deal with unstructured data as well, but not as natively as Databricks did. And then Databricks adopted vector database pretty quickly, and they built their own vector database product and made syncing between unstructured, again, object store and your vector
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 Snowflake 可以读 Iceberg，但要最佳性能仍需转成专有格式——这让计算留在 Snowflake 内
PM: I think the Iceberg is the common thing, as you said, but I think the history of it and that adoption is a bit different. From my understanding with Snowflake, you can bring the data as Iceberg, but if you want to get best performance, then you convert it to Snowflake's proprietary format. So there is, at least up to 6 months ago, friction between directly working with the Iceberg tables versus moving it to the warehousing system. With Databricks, everything is open source file formats, Delta and Iceberg. Mostly Delta now. Iceberg is part of that family as well. Databricks has 1 unified governance tool, the Unity Catalog, that only deals with these open file formats. And from my understanding, Snowflake has the Iceberg catalog, Horizon I think they call it, and then they have the other catalog, which again I'm sure they made it.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 Snowflake 开箱更易用，SQL 优先用户上手更快
DG: Yeah, I see your point. And even the ease of use is a bit better in Databricks than it used to be. I think in the past you could argue, even now to some extent, if you were SQL first and you want to have much quicker onboarding, historically and even now, maybe Snowflake is a bit easier to turn the key on and just get your data in and out. With Databricks onboarding, it was a bit slower, but they addressed the ease of use. I think mostly they addressed it, and onboarding now with serverless is there. But overall, if you step back and say as a product, if you're just coming in, which one is easier? I would say Snowflake. It exposes less of tuning and engineering features to the end user right off the box, but Databricks, you can get into more details and you have more options. And again, even on the deployment side, you have more optionality. Okay, that could be good or it could be bad. That's why I think still people who are more technical, and if they look at the products in detail, they may like Databricks more. It gives them a bit more control, more low-level access compared to Snowflake that has most of the business features you like, but then it doesn't tell you how it does it or the MLflow piece, the governance piece that Databricks has. Even when Microsoft uses it, they kind of hide the open source part of it and the guts of it, and in Databricks, if you want to, you can see more of what's going on. So I think if you're engineer-minded and you want to build more things on your own and you have to use one of these products, you would like Databricks a bit more. Fine-tuning a model or building an LLM from scratch, getting access to GPUs and doing something, optimizing your queries.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 中性 工程型公司多选 Databricks，BI/报表/仓库需求和非技术初创更偏 Snowflake
PM: So a lot of these things that you want to tweak, Databricks makes it possible. And I think Snowflake is a bit more of a black box. And again, as a business user, you completely don't care. And maybe it's even better that it doesn't expose it to you. Maybe you control the governance and cost more on Snowflake than Databricks. But if you're a digital native and you have a bunch of engineers, data engineers, I think let's say you're in Netflix and you have to pick, I know it's not people like Netflix that have their own platform, but if they have Uber or Netflix, if they have to pick one of these 2 platforms for their engineers. Most of the time I would think they would pick Databricks. But if you're a business intelligence guy and you need reporting and warehousing, yeah, it's obvious Snowflake comes across as a much easier tool to use from the scratch. DG: Hmm. PM: And you see the thing that's puzzling to me sometimes, a lot of startups still like Snowflake more. And I think part of that is a lot of startups are not techie startups. They're startups that for a business concept and they want the tech to be easier. They don't want to pay expensive data engineers. They like Snowflake. Shops that have a big... Starting on the engineering side, think of like Nike, people, companies that are not necessarily tech companies, but they have tech focus into it. They like mostly they get attracted to Databricks because they can do more things with it and more exposure to even like let's say app development on Snowflake is easy, Databricks is easy, but Databricks can also make it so that you can make it hard for yourself and gives you more flexibility on that front.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 中性 Snowflake 在成本和治理上可能更可控 ('黑盒')
PM: So a lot of these things that you want to tweak, Databricks makes it possible. And I think Snowflake is a bit more of a black box. And again, as a business user, you completely don't care. And maybe it's even better that it doesn't expose it to you. Maybe you control the governance and cost more on Snowflake than Databricks. But if you're a digital native and you have a bunch of engineers, data engineers, I think let's say you're in Netflix and you have to pick, I know it's not people like Netflix that have their own platform, but if they have Uber or Netflix, if they have to pick one of these 2 platforms for their engineers. Most of the time I would think they would pick Databricks. But if you're a business intelligence guy and you need reporting and warehousing, yeah, it's obvious Snowflake comes across as a much easier tool to use from the scratch.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 反对 BigQuery 是最原生的 lakehouse，Databricks 与 Snowflake 在 GCP 上足迹都小
PM: Yeah, I think BigQuery is a bit different story with BigQuery. GCP in general is not that, generally speaking, partner-friendly with vendors like Databricks and Snowflake and for both of them have a much smaller footprint on GCP because GCP is more of an app-first platform on its own. It's not an infrastructure play as much. So I think competing with BigQuery is harder and BigQuery is probably the most native lakehouse architecture on its own and pretty advanced and lots of features. So I think the way Databricks competes in GCP and BigQuery is more on the ETL side. Cost is a bit more if you spark from Databricks versus other things that GCP has on the ETL side. But as soon as it comes to AI and data querying, BigQuery is pretty good. But again, sometimes the cost is too high and Databricks can win, but yeah, it is a challenge.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 Fabric/Synapse 在易用和集成度上不如 Snowflake，在仓库端仍难以抗衡 Snowflake
Because Fabric... I think a lot of what's inside Fabric started from Databricks like the Delta file formats and a bunch of other stuff. But again, I don't think Fabric or Synapse, everything they build, they don't come at the same level as ease of use and highly integrated like Snowflake. So they still compete and struggle against Snowflake on the warehousing side. But ideal positioning for Microsoft is to deal with Databricks as an ETL tool, as maybe machine learning in some areas, but not for model for data serving. The data serving, they're very adamant about using Fabric.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 AWS 与 Snowflake 的 GTM 配合可能更好，AWS 不太投入 Redshift
PM: Yeah, with Microsoft and Databricks, across the board, except for the GPUs, they get very good deals and they get a pre-commitment and a better relationship with Microsoft. When it comes to AWS, both parties use Graviton extensively and that's been great for AWS. But again, when it comes to GPU stuff, I don't think that either one has a special relationship. Basically, Bedrock competes directly against these products and they use their access to their own chips and all that as a leg up and as an advantage to compete against Databricks and Snowflake. BigQuery is a competitor across the board. I don't think BigQuery plays well with either party. But yeah, that's how I see it. And again, in terms of partnership, go-to-market, Microsoft and Databricks work pretty well, synchronized. They sometimes compete as well. I think on the AWS side, probably Snowflake has a better messaging with them in terms of go-to-market because my guess is AWS makes more money on the EMR side, which overlaps more with Databricks versus on the Redshift side. They don't invest as much into Redshift, so they're kind of reverse of Azure in that sense, and they don't mind Snowflake having a bigger footprint versus their own product.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 反对 微软对 Databricks 更友好，Azure 把 Snowflake 更多视为竞争者
PM: On the, I think on the AWS side, both parties have the easiest time because AWS loves to sell the infrastructure and that's their... they don't have too much focus on competing against these products with Redshift or EMR and others. And they compete there, this story is simple. You say these are disparate products that you have to attach, stitch together to make them work. Let's say EMR with Glue, with Redshift to make a product, and you need a lot of cloud services people to help you with Snowflake or Databricks; you can bypass them. When it comes to Azure, I think Databricks has a bit of advantage because there's a core development of the Azure Databricks with them. And I think Azure sees Snowflake a bit more of a competitor to Databricks, obviously because they have profit sharing with Databricks and with Snowflake, they have deals. And I know Snowflake is pretty big on Azure too, but maybe a bit of a better partnership with Databricks. Because Fabric... I think a lot of what's inside Fabric started from Databricks like the Delta file formats and a bunch of other stuff. But again, I don't think Fabric or Synapse, everything they build, they don't come at the same level as ease of use and highly integrated like Snowflake. So they still compete and struggle
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 中性 锁定主要来自数据存放位置 (内部表、serverless) 而非治理本身
PM: Makes it more decomposable, more separated from the cloud itself. Moving to Unity Catalog is a bit of a headache, the migration part, but they make it easier. But moving out of it, if you fully adopt Databricks, then it becomes harder. That ease of using the cloud is pretty unique, especially if you go to a multi-cloud system where now Databricks has a disaster recovery and HA across different clouds, which I think is coming more and more over time. And if all of that is managed through Unity, then it is pretty appealing and sticky. Especially for compliance and all that stuff, but again, doesn't mean every customer would be locked in with Databricks, especially with the Spark and data processing stuff. You know, it is some work, but it's not that difficult to, especially if the data, the control plane and the data plane is on the customer side, then migration off of Databricks is easier. But if you go with the serverless offering, just like Snowflake, then I think getting your data out becomes harder. But how the data is residing and is it external tables, internal table, that kind of drives the lock-in more. And I think the data being on the Databricks side and the serverless makes a huge difference in terms of lock-in, maybe more even than the governance itself. But all of those definitely help with a bit of a stickiness.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 支持 Snowflake 增长更多来自份额获取，新产品甚至促使部分 Databricks 客户迁到 Snowflake
Snowflake的增长更多是由于其获取了新的市场份额，而非仅仅受益于行业整体需求的增长。自2024年新任CEO上任以来，Snowflake在产品战略上进行了一系列革命性变革，高度聚焦于AI。例如，其Hybrid Table产品在最近一个季度内增加了20多个新功能，Postgres Page也因支持二级索引而大幅提升了点查速度。此外，公司还推出了Cortex Evolve等新产品。这些新产品的功能和性能提升吸引了新客户，甚至促使部分原Databricks的客户迁移至Snowflake。公司还推出了一款用于AI迁移的工具，这表明其迁移策略已在GCR、ACR、美国和欧洲市场找到了应用场景并取得了成效。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 支持 Snowflake Hybrid Table/Postgres 推进快于 Databricks；8 月支持的 inline stored procedure Databricks 尚不支持；Databricks Lakehouse 交易型功能不成熟
Snowflake在AI领域发展迅速，其Hybrid Table和PostgreSQL的推进速度超过了Databricks，这得益于其丰富的功能特性。例如，Snowflake在2026年8月最新支持的inline stored procedure，允许在SQL语句中嵌入存储过程，从而可以调用外部API，极大地扩展了使用场景。Databricks目前尚不支持此功能，且在其现有技术路线下实现该功能的复杂度较高，工作量也很大。这是Snowflake当前的一个优势和增长驱动力。相比之下，Databricks选择的Lakehouse架构旨在将两种数据库合二为一，这是一个愿景更高的技术路线，但现阶段产品还非常不成熟。尽管宣称可以一体化使用，但在交易型数据库的许多功能上仍有缺失。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 中性 Databricks Genie 仍在促销期，尚未正式收费 (潜在价格竞争)
客户迁移至Snowflake主要有两个原因。第一，Snowflake与AWS的深度绑定关系能帮助大客户获取稀缺的GPU算力资源。例如，印度的一家大型客户Alliance，为了进行大规模数据训练和对K3、K2.7及GLM-5.2等模型进行微调（fine-tune），需要大量GPU资源，因此选择通过Snowflake来确保从AWS获得所需的GPU capacity。第二，Snowflake推出的新产品，如Hybrid Table和Postgres Page，在功能和性能上表现出色，获得了客户的认可。尽管Databricks的AI产品Genie目前仍在促销期，尚未正式收费，但上述因素依然构成了客户选择迁移的充分理由。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 支持 Snowflake 封闭体系使客户迁移成本高、粘性好，尤其在北美大公司；但难以吸引倾向开源的新兴公司
Snowflake的业务重心主要在北美市场，其创始团队及早期员工（约200人规模时，其中150人来自GCP）背景决定了其客户基础。创始人源自Google BigQuery团队，因此Snowflake在创立之初就带来了许多原先使用BigQuery的企业客户。Snowflake采用的是非开源的封闭技术体系，这带来了双重影响：一方面，一旦客户采用其平台，由于迁移成本高，客户粘性非常好，这在数据平台已经稳定且数据量庞大的大型北美公司中尤为明显；另一方面，这种封闭体系对于吸引新兴公司，尤其是那些倾向于采用开源、开放体系的公司而言，难度较大。相较之下，Databricks在GCR等新兴市场表现更佳，而Snowflake则不太关注这些区域的客户，其售前团队也不会像Databricks那样主动与客户共同探讨使用场景。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 中性 客户用 CoCo Desktop 替代 Codex/Claude Desktop，理由是一个合同访问所有模型
目前，许多客户也在用相关工具替代 Codex，并通过 Coco Desktop 替代 Cloud Desktop。Cloud Desktop 只能提供对第3方模型的访问，OpenAI 也只能提供对 OpenAI 模型的访问，而在 Snowflake 中，客户可以在同一平台访问所有基础模型公司的模型。客户无须分别购买订阅，只需签订 Snowflake 合同，即可向整个公司开放，并在此基础上进行部署。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 反对 迁入容易，迁出也一样容易
We have Iceberg tables where we're encouraging you to input your data from any anywhere anyhow or whatever, it's our -- we have to create great products like CoCo to incentivize you to use us. One of the things around that we talk a lot about internally is it's customers first. If you don't make a great experience for your customer, we talk about it's really easier to migrate into Snowflake with all these tools, it's just as easy to migrate out of Snowflake. And so, let's say the reverse of it, too, right? And so, we have to create those great customer experiences to keep them within Snowflake and to give them the ROI of doing that. And since we don't run -- we are in HR application, we are in CRM, we don't run all these various components like a lot of people want to put all their data in Snowflake. So, then they can use tools to actually be able to understand more about their business. And so, I'm very encouraged with how things are going and what we're seeing from a business momentum perspective with people put data in Snowflake. Yeah.
- 反对 Databricks CEO 的反 Snowflake 打法：开放格式、AI、宣称 3 倍 TCO 优势；最近一轮募资 $50 亿 (需求约 $200 亿)
1. **Databricks CEO Ali Ghodsi on Never Wanting the Job, Beating Snowflake and Why Databricks Stays Private** *Long Strange Trip: CEO to CEO with Brian Halligan, and Ali Ghodsi, co-founder and CEO of Databricks* The last round raised $5B against roughly $20B of demand, the anti-Snowflake playbook (open format, AI, a claimed 3x TCO edge), why he says Databricks would be worth more public, and the one-bottleneck operating model. 2. **Dan Niles: This Isn’t the Top, It’s a Speed Bump** *RiskReversal with Dan Nathan, and Dan Niles, founder of Niles Investment Management* The AI trade’s leading indicators: Texas’s data-center moratorium, the labs’ own slow-down talk, token prices down by half since May with share shifting to open-weight models, a 5% ten-year, and where he still wants to be (Meta, Google, Microsoft, Apple). 3. **Palantir CEO Alex Karp on AI Safety, Liability and Nationalizing the Frontier Labs** *CNBC’s Squawk on the Street with Carl Quintanilla, David Faber and Sara Eisen, and Alex Karp, co-founder and CEO of
来源背景：TMTB Podcast Intel (由 Ethmos 研究 agent 生成的播客摘要，明示)。相关内容为 Databricks CEO Ali Ghodsi 在 Brian Halligan 播客 'Long Strange Trip' 中的说法，以及 Dan Niles (Niles Investment Management 创始人) 的观点。视角局限：只有一句话摘要，正文是图片未读到；Ghodsi 是直接竞争对手，有利益偏向。 - 支持 credit 价格通常保持稳定，涨价会导致客户流失；近期折扣没有大的调整
关于客户折扣，目前没有听说有大的调整。Snowflake希望留住大客户，尤其是长期使用或使用稳定的客户，不会通过变相增加费用来影响客户。Snowflake收客户的钱不仅仅是通过credit折扣或均价变化，更像AWS的消费模式。Snowflake希望客户多用不同的服务和新功能，让资源节点和AI token消耗更多，通过这种方式让客户付更多credits，而不是单纯通过折扣调整。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 反对 未见大量客户从 MongoDB 迁到 Snowflake，只是零星个案；两者 OLTP/OLAP 定位不同，更多是配合使用
以我的观察，并没有出现很多客户从 MongoDB 迁移到 Snowflake 的情况。有一些零星的小部分客户这样做，但并没有大批量用户迁移。因为本质上两者还是有区别。Snowflake 最主要的功能是 OLAP，而 MongoDB 在 OLTP 上表现不错。MongoDB 更可能作为上游的数据源，用户更多用 MongoDB 完成在线事务处理，然后将数据归档或存储到 Snowflake 上，为后续 BI 服务做准备。从这个角度看，并不是谁取代谁，也不是从一个产品迁移到另一个产品的概念。更应该从业务场景来看，MongoDB 和 Snowflake 的赛道和使用方式不同，可以将它们配成一个 workflow。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 口径不齐 从 Oracle/Teradata 迁到 Snowflake 的案例有但不多；主要竞争对手是 Databricks，Databricks 客户迁来也只是偶尔
从Oracle或Teradata迁移到Snowflake的案例是有的，但并不算多。需要说明的是，Oracle、Teradata、Cloudera等并不是Snowflake的核心竞争对手。Snowflake目前的主要竞争对手还是Databri cks。客户选择Oracle、Teradata或Cloudera，往往是基于自身业务需求和运营模式，因此不会轻易迁移到Snowflake。迁移本身是一个庞大的项目，只有在有非常强的业务驱动或功能需求时，客户才会考虑迁移。如果没有明确的业务理 由，客户一般不会轻易更换平台。即使在与Databricks等竞争对手的竞争中，也只是偶尔有客户迁移过来，案例并不多。迁移的原因也不完全是产品功能，有时还涉及销售团队的人脉关系等多方面因素。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 中性 Anthropic 给 Snowflake 和 Databricks 的模型基本一致；客户通常用自己平台的 agent，不交叉使用
如果单纯从大模型层面来看，Anthropic为Snowflake和Databricks提供的模型功能基本一致。Anthropic不会针对不同客户单独调整模型，都是统一标准和接口。差异主要体现在与第三方厂商合作时的定位和用户业务场景。例如，Anthropic与Databricks Agent合作时，主要定位在构建生产级的Agent和AI应用；而与Snowflake的CoCo集成后，则更侧重于Snowflake的数据工程和开发任务，更聚焦于Snowflake数据库本身。不同厂商的合作对象和目标用户也有所不同。Databricks主要面向AI工程师、机器学习工程师和应用开发者，而CoCo则更多面向数据工程师、数据分析师或Snowflake的开发者，用户群体有所区别。此外，数据上下文也不同。Databricks侧重于自身的Unity
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 反对 Iceberg 推动多云/混合架构，数据目录成为争夺焦点——计算和控制权可脱离单一平台
FH: So it's a great question. And when we start thinking about where we are in this hybrid approach, what we're seeing is the market is shifting to adopting multi-cloud hybrid architectures that are on-prem and cloud, leveraging Iceberg capability. And so for those in the audience, I'll try to simplify it. Essentially, that is shifting the importance of the data catalog, where historically the data catalog was more of a pure governance play. Like, let me understand lineage of my data. Let me understand risk. Let me understand who owns my data, where is it at, etc. Now all of a sudden the fight in the landscape is who can win the data catalog because that is now a core component of not only interoperabilities from a strategy standpoint to allow hybrid, on-prem, multi-cloud architectures, but enabling and enriching AI from a context perspective.
来源背景：Frank Hernandez，前 Qlik 美洲区专业服务 Practice 与解决方案负责人 (约 2018 年至 2025 年底，明示/)；主持人确认其 2026 年起加入 Snowflake 并任职至今 ；25 年以上技术经验，约 20 年数据相关 。视角局限：现任 Snowflake 员工，对 Snowflake 的正面评价有明显利益偏向 (如 'every day I wake up and I'm super excited')；访谈主题是 Qlik，Snowflake 内容是侧面。 - 反对 Snowflake 客户数据量超过 2PB、慢查询约占 15% 时开销急剧增长，是部分客户转向 Databricks 的核心原因；集群升级价格翻倍
为维持并提升客户粘性，关键在于识别客户流失的风险点。观察发现，当Snowflake客户的数据量超过2PB，且慢查询占系统总量的15%左右时，其开销会急剧增长（这是一些客户转移至Databricks核心原因）。这主要是因为Snowflake的技术栈在数据量增大时，其按capacity unit计费的模式会导致成本阶梯式跃升，例如从中等集群升级到大型集群，价格会翻倍。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。 - 反对 Databricks 专门接触数据积累量大的 Snowflake 长期客户，展示成本优化空间以挖走客户
Unit）计费，其底层的Spark和MPP机制在数据量增加时能更好地控制成本，避免了成本的急剧跳升。因此，在进行POC时，尤其是在数据量巨大（如超过1TB）且查询和并发量持续增加的场景下，Databricks的成本优势会非常明显，而Snowflake则会出现成本陡增的情况。利用这一点，Databricks倾向于接触那些数据积累量巨大的长期客户，向他们展示潜在的成本优化空间，从而可能导致Snowflake的客户流失。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。 - 中性 客户在 Databricks 和 Snowflake 之间的双向迁移一直存在
客户在Databricks与Snowflake之间的双向迁移情况一直存在。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。
E05 新产品收到钱了吗
判断：CoCo 是真的爆款，采用速度是 Snowflake 历史上最快的，但它约占收入 2–2.5%，而且本身不收费；CoWork (面向业务人员) 增速已经明显放缓；Observe 和 Postgres 还在早期。
CoCo (Cortex Code)：数据工程师用的编程 agent，2026-02-05 GA，有 CLI、Desktop、VS Code 和 Snowsight 网页四种入口。账户数 7,100+ (Q1) → 9,100+ (Q2，单季净增 2,000+)。CFO 说 CoCo 本身不收费，客户付的是它驱动的消费。采用深度：Funda 的渠道说实质使用的客户约 12% (前两季约 6–7%)，广义使用从 25% 到 45%；一位专家说试用客户中只有 20–30% 形成显著消耗，一个 500–600 人注册的账户里只有 30–40% 会用，天花板受开发者人数限制；CEO 自己说 7,000 个用 CoCo 的客户里"多数只有一个用户在用"。收入：欧洲伙伴估约占收入 2–2.5%、环比接近翻倍，长期上限约 10% 且"10% 已经很激进"；Funda 模型里 CoCo 五个季度从 $3 万到 $5,940 万。
CoWork (原 Snowflake Intelligence)：给业务人员用自然语言问数据的 agent。Q1 时账户数环比翻倍，Q2 到 5,800，环比只有 +11%。CEO 在 Investor Day 说 CoWork "还要努力"，要证明能部署到 1 万人、证明 AI 的单位成本低于订阅软件；一位专家说客户对这类"随手问"的查询成本很防御，token 用量难预测导致决策停滞。
Observe、Postgres、Openflow：Observe (2026 年初收购的可观测性公司) 渠道说进展不多、需要面向传统 IT 买家的新销售动作，对用量贡献约 2–3%；Snowflake Postgres 讨论在增加、Q4 才有重大决策，主要是绿地 AI 应用，不是从 MongoDB 大规模迁移；Openflow (数据接入) 被一位专家称为"被高度低估"，攻击独立 ETL 产品的支出。管理层也承认不是所有押注都成功：SPCS (第三方容器负载) 没达预期，Native Apps 收缩为小团队。
Sayari, whose risk intelligence supports Fortune 100 enterprises and national security agencies, chose Snowflake to rebuild its global data infrastructure and cut costs by more than half. Its engineers are now using CoCo to accelerate the migration of 12 billion records into an AI-ready foundation. And as more customers see what's possible with this technology, adoption continues to build. CoWork expanded to 5,800 accounts, up nearly 11% quarter-over-quarter. Meanwhile, CoCo continues to see rapid adoption surpassing 9,100 accounts and adding more than 2,000 net new accounts in this quarter alone. We have customers like 1Password, the security company trusted by more than 200,000 businesses, which choose Snowflake for our CoCo capabilities. CoCo enabled their team to move key data pipelines into Snowflake, quickly laying the foundation for their data and AI work. And the world's number one job site, Indeed, has rolled out CoWork and CoCo across its data teams and integrated Snowflake into its core data architecture, citing lower cost and greater efficiency, which compounds at the scale that they operate at over 60 countries and 28 languages.
And so, it may be more personas. And so, you're reaching more people within a business. So typically, pre the CoCo and all the AI functions, you would basically interact with Snowflake through your data team who would actually write this SQL queries. Now anybody within the business can get access and we don't charge like CoCo is not -- we don't charge for it. You just get it being a Snowflake customer, and then the charge is the consumption that you actually draw in using the product. And so, we see more personas, and then we're seeing migrations go faster. And so, we're seeing more data come in, relative migrations. So, there's a whole bunch of different things that AI is doing that's driving strength and acceleration in the quarter.
1. Approximately 25% of customers were using Cortex Code in 1Q26, and that figure has now risen to 45%. 2. Usage stages vary from trials to actual applications. They broadly fall into three categories: trials, pilots, or proofs of concept; general usage rather than active production environments; and actual production use through workflows, integrations, or implementations after large-scale Snowpark deployment. Together, these three categories account for approximately 45%. 3. Currently, approximately 12% of customers use Cortex Code in a substantive way, meaning they have active workflows or applications governed, processed, or supported by it. This was approximately 6-7% in the prior two quarters or 1H of this year. 3. CoCo Creates a Flywheel for Other Products
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。
Still there's a lot of sweat, still there's a lot of work. And while we have – clearly, we have a lot of work to do in terms of driving adoption by each and every one of our customers. It's still pretty early. When we talk about the 7,000 customers, it will be more the case that there's like one user that's gone and done. It's not that they have switched over to this agentic way of operating and it's not like we have made migrations easy enough that anyone can migrate from anything into Snowflake so there's work to be done. I would say in the near-term, there's just – that's where there's enormous potential for us but it's sort of playing our game of be a good data platform. CoCo accelerates everything that you can do with the data platform, including creating products like CoWork that people can get even more value from. And so, in that sense, CoWork is an expansion play from where we are and it's early.
在所有已经进行小规模投入并产生消耗的客户中，大约 20%-30% 已经形成了显著消耗，而且这一比例呈上升趋势。我们预计其他一些客户的使用量也会增长，但并非所有 Snowflake 客户都在使用 Coco，未来也不一定会全部使用。Coco 主要适用于拥有多个数据源等需求的客户，并且通常只由部分开发人员使用。如果有 500-600 名注册用户使用 Snowflake，其中可能只有 30%-40% 会使用 Coco。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。
1. Cortex Code, also known as CoCo, is Snowflake’s enterprise coding agent. It is growing quickly because adoption is relatively easy. I estimate it represents about 2% to 2.5% of revenue, which is already meaningful. It and Snowflake CoWork, formerly Snowflake Intelligence, are Snowflake’s two fastest-growing areas. The core platform remains the largest in absolute and incremental revenue. Coding agents are easier to adopt than agentic workflow automation, which requires workflow redesign and substantial implementation work. 2. Quarterly sequential growth is approaching 100%. Its share of consumption was well below 1% in 4Q2026 and has now risen to two or even three times that level. 3. CoCo could eventually reach 10% of consumption, but getting above 10% would be difficult. If it expands beyond Snowflake, the growth opportunity would be larger, but the core platform and Snowflake CoWork are also growing rapidly. I would not raise my long-term estimate for CoCo above 10%. Even reaching 10% would mean roughly quadrupling its share, and because total revenue is also growing, the absolute incremental revenue would be much larger. Using 10% as the ceiling is already quite aggressive.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。
It placed us firmly on the AI map. It generated a lot of momentum and relevance for Snowflake as a product, but CoWork is like a giant step ahead in terms of the things that it can do, but we very much have to prove ourselves in terms of getting the product deployed by large departments like we have with Snowflake. But the fact that I can get Brian to talk to every CFO and look them in the eye and say, this is how I transform how we operate is a huge asset and it's the same for JB and team. What we have to translate that into the logos, into the 10,000 user deployments, we have to get customers happy with the cost of spending money on AI. We have to convince them that the per unit cost of using CoWork is a lot less than paying for some number of – some amount of subscription software. So, that's the potential but I'm the first person to say like that is early and we have to prove the scaled use cases to you. And between the two, it's like CoCo sells itself. Why? Because we have, whatever, 14,000 customers that love Snowflake. I just go to them and say, everything you do with Snowflake is going to be 10 times faster. They'd be foolish to not like go try it out. CoWork, we have the work – we have work to do to sell it and prove ourselves. We got one, third row right there.
BJ: I think, great question. I think we've seen early stage traction on CoCo. I think it converts faster than CoWork. And I think that's a scope difference rather than product quality difference. CoCo is just a coding agent. A lot of people are already used to coding agents for local files, and it's leveraging Snowflake data, which was there. The reason for that is because when the model answers wrong, you immediately go back to the data, right, and the architecture that powers it. And that is when that developer notices immediately, hence CoCo just becomes a comfortable, familiar interface. CoWork, I think, is where I'm personally interested because it's aimed at business users making decisions about output, but that ICP comes in afterwards, right? It directs the whole question. I think business users are really worried about the insight that is to be trusted or not. I think what we're also seeing is early stage defensive posture on costing, right? It's more conservative, and the problem is that ad hoc queries don't behave like scheduled dashboard refresh, right? It is hard to forecast; a finance function that can forecast, it's a warehouse spend, which was easy, right? We could have predicted a few % every quarter, every time period. Now, what we don't know is if someone asks a question, what is the predictable token users associated with it? I think that does stall some decision-making within the organizations. So that's where some of the decisions get stalled because there's more analysis and paralysis that comes with it. At the same time, I think a lot of that is necessary to really model the governance structure so that we can at least from finance to functional technical aspects, be more reliable in the ability to build business cases and the contracts that are associated with those cases accordingly. And then the movement happens. I think right now, fundamentally, every client of mine is worried about compute costs. And there's DG: Right.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。
The reason they entered that business is because, um, many customers are modernizing their, um, data environments, and for observability, even though the log data is stored within their environment, they have to ship it to data dog or dynatrace and pay for storage costs there also. So right? So there was double storage costs and also, you know, uh, open telemetry is gaining traction rather than agent- based solution from, uh, data dog and others. But in my view, we are not seeing a lot of traction with Observe, uh, from Snowflake at this time. Um, I think that's a complex business to get into, um, not something that Snowflake is, uh, is known for. So I would say the impact is less than 5% for consumption, right? two to 3%, not not more. Uh, but the PostgreSQL um, products are are gaining traction and um, you know, the the messaging around consolidating operational data store and analytical data store is is gaining some traction and I think will continue to do so. Uh, MongoDB is much more expensive compared to these offerings. So, uh, Mongo has some, you know, challenges to deal with.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。
But when you make it, it's still painful. So, I think we're being pretty flexible about where we are allocating, where we need to pull away from and we will continue to do that. Actually, so 100% – I have more examples. For example, SPCS, as a third party customer bring your your workload, yeah, it has not gone the way we envisioned it but I just shared the cloud agent sandbox that is enabling a lot of power to CoCo. That is the same theme and borrowing technology from what was built there. The momentum with notebooks, the momentum with Streamlit came from, hey, instead of pushing SPCS so hard as something external, make sure that you have a great momentum environment. So, some of that reallocation is happening, for sure.
展开本节全部 117 条原文 (支持 58 / 反对 22 / 其他 37，按时间)
- 支持 每周 7,300+ 账户在用 AI 功能，1,200 客户用 agentic 能力 (Snowflake Intelligence)
Sridhar Ramaswamy: In Q3, more than 7,300 accounts are using our AI capabilities every week. Sridhar Ramaswamy: we are seeing the fastest ramp in product adoption in our company history. Sridhar Ramaswamy: Already 1,200 customers are harnessing next-generation agentic AI capabilities to drive real business impact at scale. Sridhar Ramaswamy: AI influencing 50% of the bookings signed this quarter. Sridhar Ramaswamy: 28% of all use cases deployed during the quarter incorporated AI.
- 中性 Postgres (Crunchy Data) 还需几个月 GA
Robins: We're also happy with the developments in AI, and also the data engineering workloads. Sridhar Ramaswamy: products like Openflow are taking off Matthew Hedberg (RBC): how customers thinking about that long-term balance of OLTP and OLAP within Snowflake? Sridhar Ramaswamy: AI revenue is predominantly driven by the Cortex product suite, including Snowflake Intelligence Sridhar Ramaswamy: With respect to Crunchy data, it will take us a couple of more months to get the product into GA. Sridhar Ramaswamy: customers are very welcoming of Postgres support within Snowflake. Sridhar Ramaswamy: bringing Postgres to market will be an important step forward for us Sridhar Ramaswamy: we will continue to have both of these Tyler Radke (Citi): are you expected to see significant growth in those deals? Sridhar Ramaswamy: Product revenue is the best indicator of how our collective customers are going to be spending Brian G. Robins: I can assure you that there will be no change to the guidance philosophy. Brian G. Robins: Tyler, as you mentioned, we'll guide to FY 2027 on our next call.
- 支持 收购 Datometry 迁移技术，用于加速传统数仓迁入
Sridhar Ramaswamy: AI is a key driver of the strength that we see in our core business. Sridhar Ramaswamy: today, we've announced an expanded partnership with Anthropic. Sridhar Ramaswamy: This brings native model availability into Snowflake, Sridhar Ramaswamy: enable our customers to move from legacy data warehouses to Snowflake Sridhar Ramaswamy: We recently acquired the technology behind Datometry's software migration solution Sridhar Ramaswamy: Key capabilities like Snowflake Openflow are making it easier than ever to bring in structured, Sridhar Ramaswamy: bringing our total GA product capabilities to 370 year-to-date, Sridhar Ramaswamy: over 40,000 customers, partners and prospects joined us across 23 events Sridhar Ramaswamy: we announced new partnerships with Workday, Splunk, Palantir, UiPath and more Sridhar Ramaswamy: to make the latest Gemini models available to our 12,600-plus customers Sridhar Ramaswamy: we introduced Cortex AI for financial services Sridhar Ramaswamy: We've also agreed to acquire Select Star to enhance our Horizon Catalog Sridhar Ramaswamy: Snowflake has already surpassed $2 billion in sales through AWS Marketplace in a single calendar year,
- 口径不齐 专家称 CoCo Q3 收入 <100 万美元、Q4 可达 1.14 亿美元，占收入 2-3%——与 CoCo 刚 GA 的事实矛盾，疑似混淆 AI 收入 run-rate
我们预计仅Cortex Code的采用就将达到约2500家或更多。从收入角度看，第三季度的估计低于100万美元。在第四季度，这一项预计可达1.14亿美元，暗示该特定组成部分环比增长非常强劲。这只是一个产品——Snowflake有多种产品。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 支持 CoCo 采用预计达约 2,500 家或更多
我们预计仅Cortex Code的采用就将达到约2500家或更多。从收入角度看，第三季度的估计低于100万美元。在第四季度，这一项预计可达1.14亿美元，暗示该特定组成部分环比增长非常强劲。这只是一个产品——Snowflake有多种产品。
来源背景：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 - 反对 Snowflake Intelligence (含 Cortex) 只占其样本消费 <5%，采用本身不驱动消费
它对增量消费有贡献，但从消费量来看它不到 5%。单独采用 Snowflake Intelligence 并不会驱动消费；消费是由你在平台上构建的用例以及它在本地环境中提供的功能驱动的。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 到年底 AI 层消费占比也只到 6-7%，不会到 10%，因为试验→生产有滞后
我预计会增加，但不会像渗透率那样大。可能会超过 5%，但我认为不会达到 10%。我仍会估计到今年年底为 5+%，也许在 6 到 7% 之间。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 支持 Snowflake Intelligence 客户渗透率由 >10% 升到 15-17%，年底约 35%
A:到今年年底，我估计渗透率为 30% 到 40%——可能接近中间范围，所以大约在 2026 年第四季度为 35%。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 反对 CoCo 当时数字很小，几乎看不见
A:有的。数字很小——几乎看不见——但从概念上很有趣。这是位于云数据仓库之上的 AI 栈。他们还有 Cortex Intelligence 的模型托管，增长很快。关键在于试验。我们有许多客户在做预测性用例的试验并运行分析。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 中性 Observe 是增量收购，不会重塑市场
从 Snowflake 的角度看，这一层——位于数据仓库、存储和计算之上——是战略性的。他们也热衷于保护它。我把这看作一次增量收购，而不是一次重塑市场的大型收购。
来源背景：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 - 支持 Snowflake Intelligence 三个月内到 2,500+ 账户，环比近翻倍
Sridhar Ramaswamy: We added 2,332 net new customers this year, and we are seeing more and more businesses move over to Snowflake. Sridhar Ramaswamy: This quarter, we delivered the largest sequential increase in accounts using AI, Sridhar Ramaswamy: bringing the total to more than 9,100 accounts. Sridhar Ramaswamy: And in just three months, Snowflake Intelligence has scaled from a nascent offering to an essential capability Sridhar Ramaswamy: Snowflake Intelligence has scaled from a nascent offering to an essential capability for over 2,500 accounts, Sridhar Ramaswamy: almost doubling quarter-over-quarter. Sridhar Ramaswamy: It's a truly transformational coding agent that's already helping over 4,400 customers Sridhar Ramaswamy: helping over 4,400 customers build and scale AI-powered applications Sridhar Ramaswamy: This year we launched over 430 product capabilities, underscoring the strength of our product velocity. Sridhar Ramaswamy: Snowflake Openflow, now generally available, makes it easier than ever to bring in structured, unstructured Sridhar Ramaswamy: Now generally available, Snowflake Postgres is a world-class operational database Sridhar Ramaswamy: This transforms Snowflake from a system you analyze with into a platform that you build on.
- 支持 Cortex Code GA，已服务 4,400+ 客户
Sridhar Ramaswamy: This quarter, we delivered the largest sequential increase in accounts using AI, Sridhar Ramaswamy: bringing the total to more than 9,100 accounts. Sridhar Ramaswamy: And in just three months, Snowflake Intelligence has scaled from a nascent offering to an essential capability Sridhar Ramaswamy: Snowflake Intelligence has scaled from a nascent offering to an essential capability for over 2,500 accounts, Sridhar Ramaswamy: almost doubling quarter-over-quarter. Sridhar Ramaswamy: It's a truly transformational coding agent that's already helping over 4,400 customers Sridhar Ramaswamy: helping over 4,400 customers build and scale AI-powered applications Sridhar Ramaswamy: This year we launched over 430 product capabilities, underscoring the strength of our product velocity. Sridhar Ramaswamy: Snowflake Openflow, now generally available, makes it easier than ever to bring in structured, unstructured Sridhar Ramaswamy: Now generally available, Snowflake Postgres is a world-class operational database Sridhar Ramaswamy: This transforms Snowflake from a system you analyze with into a platform that you build on. Sridhar Ramaswamy: And our recent acquisition of Observe, a market-leading observability platform, Sridhar Ramaswamy: Our deepened partnership with Anthropic is already helping customers like Intercom see significant impact.
- 中性 Snowflake Postgres GA，定位为可在上面构建应用的平台
Sridhar Ramaswamy: And in just three months, Snowflake Intelligence has scaled from a nascent offering to an essential capability Sridhar Ramaswamy: Snowflake Intelligence has scaled from a nascent offering to an essential capability for over 2,500 accounts, Sridhar Ramaswamy: almost doubling quarter-over-quarter. Sridhar Ramaswamy: It's a truly transformational coding agent that's already helping over 4,400 customers Sridhar Ramaswamy: helping over 4,400 customers build and scale AI-powered applications Sridhar Ramaswamy: This year we launched over 430 product capabilities, underscoring the strength of our product velocity. Sridhar Ramaswamy: Snowflake Openflow, now generally available, makes it easier than ever to bring in structured, unstructured Sridhar Ramaswamy: Now generally available, Snowflake Postgres is a world-class operational database Sridhar Ramaswamy: This transforms Snowflake from a system you analyze with into a platform that you build on. Sridhar Ramaswamy: And our recent acquisition of Observe, a market-leading observability platform, Sridhar Ramaswamy: Our deepened partnership with Anthropic is already helping customers like Intercom see significant impact. Sridhar Ramaswamy: We also recently announced a $200 million expanded partnership with OpenAI.
- 支持 Snowflake Intelligence 三个月内超过 2,500 个账户、环比几乎翻倍；使用 AI 的账户超过 9,100 个
* “但Cortex Code才是真正改变游戏规则的产品，因为它是整个数据生命周期的强大加速器……构建DBT管道、运行数据工程、建立动态表，或调试性能问题，速度又提升了10倍以上。” * "本季度，我们实现了使用AI账户数量的最大环比增幅，总数超过9100个账户。仅仅三个月内，Snowflake Intelligence就从一个初期产品成长为超过2500个账户的核心能力，环比几乎翻倍。”
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 支持 Snowflake 客户不愿为第三方语义层'双重付费'，原生语义 view 只收算力——平台原生功能挤压独立工具
And I'm not sure where they are today on their pricing. When I left, they had changed that to more of a consumption-based pricing model, but the issue still is there. I even spoke with Snowflake customers who were like, well, why should I pay for AtScale? Because I have to double pay, because I have to pay you to have access to your software and I have to pay for compute costs with Snowflake because you're sitting inside it. So I get double cost or I could just bite the bullet. have this really manual semantic layer in Snowflake, but they're not charging me double.
来源背景：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 - 支持 Cortex Code 被称为数据与分析领域缺失的编码 agent，是 'game changer'
I've had a lot of lunch and learns within our team and self-trainings around using Cortex Code as a coding agent. And I will say right off the bat that Cortex Code from a Snowflake perspective really is a game changer because I know coding agents have been around forever with Cursor and VS Code and Windsurf, and there's so many coding agents out there for software engineers. In this space, we're not software engineers. We need a coding agent for data and analytics, and that's exactly what Cortex Code brings to the table. And so we've really, in order to meet that gap that I was talking about last quarter, we've had to find these efficiencies using Cortex Code. Since, yes, we hired 5 more engineers, but of those 5, we needed probably 10 to 15.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 Claude/Cursor 等通用编码 agent 只能覆盖其列举用例的约 20%，因为缺少 Snowflake 内部数据层
FC: With that said, you know, you can have Claude go through your entire repository and understand from a repository perspective what code exists, what's in there, and how is it operating? Like you can see all my code, no problem. But what you're not seeing is the other half of the world, which is what's the interplay of these data layers that we've set up over the last 10 years. And that's the internals that only Snowflake has. So no, I think Claude does... trust me, I've gone down that road and used a lot of the initial coding agents that had come out, but from a data and analytics perspective, they don't work, unfortunately. They don't work well. How about that? They can give you some of the... like all those use cases I gave earlier, I would say they work for 20% of those use cases.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 客户端 Cortex Code 采用度高，熟练工程师产出 '至少 5 倍'
FC: I will say we are training them to use Cortex. So the rollout of it is pretty high. Of course, in some ways we don't want to reveal our secret sauce as far as how we're generating more code than their own internal teams because obviously we can't make them look bad. But with that said, they catch on pretty quickly and want to start using Cortex Code as well and gain those efficiencies. So yes. Once they see it live, I think it doesn't take much to impress them because they quickly garner, especially with all the buzz that's out there, that I can become more efficient. So the smart ones figure it out very quickly. They want to impress their bosses. They start doing it. The slower adopters are the ones that have been around the block a lot longer and might be tried and true, and they'll grumble about how you can make mistakes easily through that. But at some point, they have to recognize that I have engineers on my team that are at least 5x-ing the output that I'm doing, and therefore I need to start thinking about using Cortex Code as well. So the adoption's there, but you have to demonstrate that value as well as the safety of it. So again, there's a deployment rulebook that we have to follow around evaluating all code, doing exactly what I was talking about, which is have Cortex criticize your code and create more efficiencies. Have another human reviewer, which is one of your coworkers, review your code, ensure that it looks good. We're obviously going through full test cycles and then we are putting it through the deployment cycle. So that is how we have garnered a lot of interest and success using Cortex Code.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 Cortex Code 也被作为把其他 RDBMS 用户拉来 Snowflake 的卖点
FC: From a model perspective, so more of the AI layer in a standard fashion so that other tools could use it. I know Anthropic, which has created MCP, that's exactly why they want that information, right? From a Claude perspective, if you could pass all the information through MCP, then it would be usable. But with that said, you know, Snowflake is going to... that's their moat, right? If you think about it, anyone that's a Snowflake customer is not going to want to share their Snowflake setup, their semantic models, and their schemas, et cetera, and broadcast that across into Anthropic, right? That is their own internal workings and what they want to keep proprietary as, "Hey, you're a Snowflake customer, we're going to keep your data safe," as well as, "If you want to run it, just use Cortex Code." It's the most powerful coding agent for a data and analytics layer, especially for Snowflake users. And that's a great selling point for why someone who's on another RDBMS would say, "Hey, I want Snowflake as my enterprise data platform." BB: Yeah. And just last question on Cortex. What have you seen on a quarter by quarter basis? How fast is the ramp then?
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 反对 KeyBanc 称 CoCo 采用数据弱于管理层说法
KeyBanc says recent SNOW surveys were mixed ahead of earnings, with April consumption trends remaining healthy/in-line while forward spending intentions deteriorated meaningfully. The firm flagged weaker NTM partner growth expectations, potential F2Q guide risk (~24% vs Street 26%), and softer-than-expected Cortex Code adoption data versus management commentary. However, KeyB also notes Snowflake continues gaining strategic relevance versus Databricks among many enterprise customers, while core consumption trends into F1Q remain constructive, maintaining OW and $200 PT.
来源背景：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。 - 支持 CoCo 时间线：2026 年 2 月左右 GA，3 月开始试点，多数账户 4 月 1 日开始计费；约一半账户以某种形式在用
+ GA: ~February 2026 + Pilot/showcasing began in March + Most accounts started incurring billing on April 1 * **Adoption:** Nearly 50% of accounts are using Cortex Code in some form (pilot, production, or evaluation) * **Consumption mix:** ~5% as of end of June; expected to reach 10–12% by year-end
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 伙伴 C 称公司宣称的 CoCo 采用率和生产力提升被夸大，实际约为公司说法的 50-60%
+ Snowflake claims 52% of customers are actively using; the expert’s own number is closer to 40% + Supported models include Claude Opus 4.6, OpenAI GPT-5.2, etc. * **Performance — Snowflake’s claims vs. expert’s observations**
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 专家观察到的实际效果约为公司宣称的一半多
+ Overall, the expert’s observed real-world impact is roughly 50–60% of Snowflake’s claims
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 欧洲伙伴 (Cognizant) 估计 CoCo 占其 Snowflake 业务 5-10%，未来最多约 10%，不会到 30-40%，更多是间接驱动
* **Expert’s estimate of revenue contribution:** 5–10% (hard to break out, since it’s bundled inside Snowflake engagements) * **Future growth:** could rise to 7–8% or perhaps 10%, but will not reach 30–40%. The expert sees it more as an indirect driver of broader Snowflake consumption. * **Productivity uplift (expert’s observation):**
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 中性 Observe 仍处实验阶段，少数客户试点，尚未产生收入
* **Current state:** Still experimental. Discussions began in March–April; a small number of customers are in the pilot stage.
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 支持 CoCo 已有 7,100+ 账户；Snowflake Intelligence 账户数环比翻倍以上 (未给绝对数)
Snowflake Intelligence is the business user surface of that control plane; CoCo is the builder interface. Together, they help customers move from insight to action and from prompt to production, all within Snowflake's trusted governance model. In fact, accounts using Snowflake Intelligence more than doubled quarter-over-quarter as more organizations embrace a governed conversational way for business users to ask questions, get answers, and act on enterprise data. And CoCo is already in use with more than 7,100 accounts, giving builders a natural language way to create applications, pipelines, agents, and workflows directly in Snowflake. Just recently, our partner, Infinite Lambda, was preparing for a major customer pitch. One of their engineers used CoCo to build a true customer 360 application in just five hours, bringing together customer data, chart insights, recommended actions, and live dashboards into a single experience, and they showed it to the customer.
- 支持 管理层称 CoCo 贡献'有意义的' AI 收入，但未给美元金额
By leveraging CoCo to build and deploy intelligent applications directly within Snowflake, its teams can turn complex regulatory data into actionable insights in seconds while accelerating product development. This approach maintains the fiduciary-grade governance and reliability required for high-stakes professional use. CoCo is contributing meaningful AI revenue while also driving increased engagement across the broader platform. This tangible momentum, together with continued strength in our core platform, is reflected in our increased FY 2027 outlook. Today, with the announcement of our intended acquisition of Natoma, we are extending the Snowflake agentic control plane beyond data and development workflows into the everyday applications where work happens.
- 中性 宣布收购 Natoma (约 20 人)，把 agent 动作扩展到 Slack/邮件/日历/Jira
CoCo is contributing meaningful AI revenue while also driving increased engagement across the broader platform. This tangible momentum, together with continued strength in our core platform, is reflected in our increased FY 2027 outlook. Today, with the announcement of our intended acquisition of Natoma, we are extending the Snowflake agentic control plane beyond data and development workflows into the everyday applications where work happens. With Natoma, users can do things like send emails, summarize Slack conversations, check calendars, and open Jira tickets without ever leaving Snowflake Intelligence or CoCo. The important point is not just convenience, it is control.
- 支持 Snowflake Intelligence 账户数环比翻倍以上；CoCo 已在超过 7,100 个账户使用
“Accounts using Snowflake Intelligence more than doubled QoQ.” “CoCo is already in use across more than 7,100 accounts.”
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 Funda 估计 Snowpark 一年才到收入 1%，CoCo 两个月就到 1%，单季可贡献 2-3%
* This is the most intriguing Snowflake print in the past two years, on par with the 3Q24 quarter. * The 3Q24 Snowflake story was that the new CEO was driving company-wide reform, and that Iceberg was incremental rather than a drag on the business. * The 1Q26 Snowflake story is that AI revenue has officially become a new growth driver—and unlike the prior driver Snowpark, this one is not ramping slowly. Snowpark took a year to reach 1% of revenue; Cortex Code, in our view, took just 2 months to reach 1%. * Imagine a product that contributes 2–3% of revenue in a single quarter. If it can sustain a similar growth trajectory, this year is bound to be one of acceleration—and from where we stand today, the most compelling part is that you don’t know how high the growth rate could go. * Historically, every new product launch from Snowflake has had a long commercialization cycle. Snowpark reaching 1% in a year was already considered strong. Iceberg and Cortex AI ramped even more slowly. But now with Cortex Code and Snowflake Intelligence, these products don’t behave like database products at all—they look much more like API products, with explosive growth potential.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 中性 CoWork 纯消费计价，不收 per-user 费，支持按用户/账户设上限；有客户为 3,000 名销售部署并要求 per-user limit
And even though people optimize more with CoCo, we actually get more consumption anyway because they just do a whole lot more. And that's the benefit of a general purpose Swiss Army knife like tool that CoCo is and we have already worked on things like per user limits, per account limits for how much some tool should be used. In fact, I'm having a conversation with a company about deploying CoWork for 3,000 odd sales folks within that company and part of the guarantee that they want is a per user limit. And our model, which is pure consumption, we don't charge a per user fee for CoWork is actually very beneficial here because customers end up getting the best of both worlds. They can both place a limit on how much one user can consume. But if users don't consume anything at all, they spend zero. So you see little innovations like that actually be helpful especially in a consumption model that starts at zero that does not have a seed-based license. Many of the coding agent providers do like a blended seat plus token pricing and I'm actually pretty happy that we stayed away from those.
- 反对 CoWork 还在早期，需要证明大规模部署、证明 AI 单位成本低于订阅软件；CoCo 自己卖得动，CoWork 还要努力
It generated a lot of momentum and relevance for Snowflake as a product, but CoWork is like a giant step ahead in terms of the things that it can do, but we very much have to prove ourselves in terms of getting the product deployed by large departments like we have with Snowflake. But the fact that I can get Brian to talk to every CFO and look them in the eye and say, this is how I transform how we operate is a huge asset and it's the same for JB and team. What we have to translate that into the logos, into the 10,000 user deployments, we have to get customers happy with the cost of spending money on AI. We have to convince them that the per unit cost of using CoWork is a lot less than paying for some number of – some amount of subscription software. So, that's the potential but I'm the first person to say like that is early and we have to prove the scaled use cases to you. And between the two, it's like CoCo sells itself. Why? Because we have, whatever, 14,000 customers that love Snowflake. I just go to them and say, everything you do with Snowflake is going to be 10 times faster. They'd be foolish to not like go try it out. CoWork, we have the work – we have work to do to sell it and prove ourselves.
- 中性 CoWork 拓展到数据团队以外更难，客户内部用户渗透率很低
Are we actually seeing an acceleration in the number of seats for those organizations? I do think that we're seeing a broadening of the reach of Snowflake. But as Sridhar said when he was talking about CoWork, it is harder to go beyond our core audience. But those examples that we're quoting on, we have customers saying, hey, I'm going to put this in front of 500 users, 600 users, 3,000 users. It is happening and it reaches different functions and disciplines. So, I still think that it's on in our journey with CoWork, convincing customers to deploy something to every employee organization. It takes time, it takes effort and we're very early on. So from a number of people that an organization that leverages Snowflake, we're very under-penetrated. But to be clear we just started on that journey. And this is something we track pretty extensively internally in the context of Snowflake Intelligence/core, which is the number of unique users for each of the customers that we have, how do we drive that up, how do we the get entire departments to adopt it.
- 反对 产品押注并非都成功：SPCS (第三方容器工作负载) 没达预期，Native Apps 收缩为小团队
So, I think we're being pretty flexible about where we are allocating, where we need to pull away from and we will continue to do that. Actually, so 100% – I have more examples. For example, SPCS, as a third party customer bring your your workload, yeah, it has not gone the way we envisioned it but I just shared the cloud agent sandbox that is enabling a lot of power to CoCo. That is the same theme and borrowing technology from what was built there. The momentum with notebooks, the momentum with Streamlit came from, hey, instead of pushing SPCS so hard as something external, make sure that you have a great momentum environment.
- 支持 CoCo 在上季度中旬 GA，1-2 个月内采用很快，全球 8,000+ 客户在用
回 答 者：首先，Snowflake 通常不按产品线划分并公布收入，因此公司不会单独披露 Cortex Code 的收入数据。Cortex Code 在上个季度中旬才正式发布（GA），在过去的 1 到 2 个月里，其采用率增长非常迅速。目前全球已有超过 8,000 家客户正在使用 Cortex Code。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 中性 CoCo 的全部费用计入 Snowflake credits，没有单独收费
其次，部署过程毫无阻力。Cortex Code 是 Snowflake 的原生内置功能，客户无需安装新软件、更改 UI 或进行任何额外操作。当他们在 Snowsight UI 中编写 SQL 查询时，Cortex Code 窗口会自动弹出。此外，与 GitHub Copilot 等工具不同，Cortex Code 支持从代码创建、调试、迭代、部署到命令行（CLI）操作的端到端开发全流程。最后，其所有成本都包含在 Snowflake 的信用点数（credits）内，没有额外的费用。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 支持 CoCo 比 Claude Code 贵约 40% 是标价，企业客户有折扣，实际差距更小；企业客户对价差不敏感
回 答 者：首先，这 40% 的价差是基于官方定价，并未考虑针对企业级客户的折扣，实际差距会更小。其次，这个定价忽略了竞品的隐性成本。企业在引入新的第三方工具时，需要投入大量资源进行安全与合规审查，法务团队也需介入，这不仅延长了销售周期，也产生了高昂的审查和运维成本。此外，通过连接器接入 Snowflake 生态系统中的其他工具也会产生额外费用。综合来看，包含所有隐性成本后，竞品方案的总体拥有成本可能远高于 Cortex Code。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 支持 Snowflake 客户以财富 500 强为主，对 CoCo 价差不敏感；Claude Code 的主力用户是初创公司和个人开发者
通常情况下，使用 Snowflake 的主要是财富 500 强公司等大型企业客户，例如金融机构和零售商。这些客户对 Cortex Code 的价格差异并不十分敏感。而 Cloud Code 目前的主要用户是初创公司和个人开发者，对他们来说，40% 的成本差异是显著的。对于企业客户而言，尽管他们为 Cortex Code 多支付了 40% 的费用，但在集成、人力和劳动力成本方面节省的总拥有成本，远超这部分溢价。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 反对 Snowpark 增长已趋于稳定、没有爆发，原因是 DataFrame 学习曲线、调试难、性能问题 (与专家B“Snowpark 是 Q2 第三大贡献”矛盾)
Snowpark 的核心价值主张在于其 Python、Java 和 Scala 的开发能力，允许客户在无需移动数据的情况下，直接在 Snowflake 环境中编写代码。然而，Snowpark 的增长已经趋于稳定，并未出现爆发式增长。根本原因在于，客户需要学习 Snowpark 的数据帧（data frames），这与最常用的 Pandas 格式不同。此外，他们在调试方面遇到了很多问题，Snowpark 中的错误追踪起来非常困难，尤其是在处理大型分布式数据集时。有时，使用 Snowpark API 编写代码还会遇到性能问题。
来源背景：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 - 支持 CoCo 渗透势头强于 Genie，预计很快 50% 的客户会以某种形式使用 CoCo
在产品采用率方面，特别是比较 Genie 和 Cortex Code 时，我认为 Snowflake 更胜一筹。Cortex Code 的市场渗透势头更强劲，主要有三个原因：第一，它直接内置于一个非常庞大的现有客户群中；第二，它解决了开发者生产力这一迫切问题；第三，Snowflake 一直在积极地向其客户群推广 AI 应用。我预计很快将有 50% 的客户以某种形式开始使用 Cortex Code。
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 中性 CoCo 采用不是因为 Claude 等其他平台限流溢出，限流最多是次要因素
回 答 者：我认为需要将这两个问题分开来看，它们之间没有必然的联系。我不认为 Cortex Code 采用率的增加是由其他平台的计算资源短缺或令牌限制所驱动的。恰恰相反，计算资源的限制实际上会减缓其采用速度，而不是加速。因此，即便存在一些边际效益，但这并非主要驱动因素。根据我的观察，其采用更多是由原生
来源背景：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 - 中性 Snowflake 4 月 (Q2 初) 推出 AI Credits，把 AI 定价和版本定价部分拆开
此外，Snowflake 在 4 月（即第二季度初）推出了 AI Credits，这在一定程度上将 AI 定价与版本定价进行了解耦。但即便 Snowflake 通过 AI Credits 机制在一定程度上降低了价格，整体情况也不会改变。Cortex Code 的核心价值在于，它能帮助团队更快地生成数据管道，从而推动整体计算资源的消耗。这才是 Cortex Code 的真正目的：作为一种工具，确保团队提高生产力，进而增加平台消耗。因此，我认为 Snowflake 更多地将其视为一个间接的收入驱动因素，而非直接的利润来源。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 40% 溢价从现有 credit 池里扣，不是单独的采购决策，客户心理上容易接受
回 答 者：是的，我认为客户在一定程度上会忽略价格差异。这里存在一个心理因素：您提到的 40% 溢价，对于现有 Snowflake 客户而言，并不像一笔独立的、需要重新决策的成本。它更像是在现有信用池中消耗的增量成本，这使得客户更容易接受。这与需要对比不同产品的独立采购决策不同，客户只是增加了信用额度的消耗。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 反对 CoCo 用户约 95% 是开发者，不是面向业务用户的低代码平台 (与专家A“业务分析师大量涌入”矛盾)
回 答 者：Codex Code 仍然是一个主要由开发人员使用的工具，我不会将其定义为面向非开发人员的低代码平台。基本上，可以认为其用户中 95% 是开发人员。这与 Claude Code 的情况类似，它也是由我们的开发人员使用，而非业务用户。Codex Code 并非像 Lovable 或 Replit 那样的平台，它本质上仍是一个专业的编程平台。我所接触到的场景都将其定位为提升开发人员生产力的工具。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 中性 Snowflake Intelligence 吸引非技术用户，但目前对业务增长影响不大，更多是为未来做的布局
然而，目前它并未对业务增长产生巨大影响。我更倾向于将其定位为平台未来发展的一种演进，是为了顺应企业级智能代理（Enterprise Agent）和低代码开发趋势而做的布局，旨在保障平台的未来竞争力，而不是一个现阶段的主要增长驱动力。
来源背景：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 - 支持 Snowflake 通过 Cortex、语义 view 和 Snowflake Intelligence 基本形成端到端方案
**Expert:** If I was still in presales, I could probably make up some stories, but very, very limited, unfortunately. As Databricks especially have got the AI BI Genie and their Unity Catalog and their semantic layer, they've pretty much got everything you need, especially with the acquisition of Mozaic Data, where if you don't like the answer you're getting, you can always go in and start developing your own agentic solution. The same very much with Snowflake. Snowflake had a bit of a soft start with Snowsight. That's been deprecated. They're not too heavy on the analytics side. With the Cortex piece, with the semantic Views, which is a bit technical, with Snowflake Intelligence, I think it's called, they pretty much have an end-to-end. I know for a fact, Qlik do partner with both of those folks. I do know for a fact that there are conversations where the Databricks rep is thinking, "Well, hang on a minute, we can do all of this. We can then maybe discount the front end because they're all about using the compute." Unfortunately, I think the Qlik value proposition diminishes particularly strongly, not only in Snowflake and Databricks, but also within AWS as well with Glue, with Bedrock, with QuickSight and QuickSuite. They're pretty strong tools around analytics and the AI story. There's been a few customers, I know have churned to those tools.
来源背景：Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 之后重复一遍，以前半部分行号为准。 - 支持 Snowflake 销售更愿意推 CoCo 或自家可观测套件，因为能直接拉动客户环境内的计算
**TL** [00:40:06]: No. The reason is largely because Monte Carlo by design doesn't and shouldn't be driving a ton of compute, let's just say, for Snowflake. If you're on the go-to-market team at Snowflake, you're going to have potentially 15 or 20, maybe more, vendors that are ahead of Monte Carlo on the list of what you're going to recommend because it doesn't drive that much compute. If I'm a Snowflake account exec, much more likely to push people towards CoCo or push people towards the Snowflake observability suite because I know that's going to just drive compute within my customer environment. I have no real incentive to bring in a third party for these specific use cases.
来源背景：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 - 中性 客户在 Snowflake 的承诺额通常只能有 10-20% 用于第三方 marketplace 供应商
Now, when a customer is already down the buying cycle and they just want to utilise some of their remaining commit, it can grease the wheels of a customer engagement or a deal, but it then comes down to how their contract looks like with Snowflake. If they have a USD 10m commit with Snowflake, they can typically only use 10%, maybe 20% of that on third-party vendors. Would Monte Carlo get 100% of that 20% or not? Most likely, not because there are a lot of other tools that people will use their marketplace spend for first. Especially if you want to run LLMs on Bedrock, for example, you might use your commit for that over Monte Carlo. It's never been a real catalyst for the business. I don't see where it would become a catalyst for the business unless it's really customer-driven or very rarely driven by the cloud partner themselves.
来源背景：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 - 支持 Snowflake 的 CoCo 和 Databricks 自有基础设施让独立可观测性厂商难以差异化
**TL** [00:04:01]: Personally, I don't see that happening at Monte Carlo necessarily. I think a lot of things are going to have to go their way for a standalone player to be able to come in and do something that's really differentiated from what you're getting with what's called CoCo now in Snowflake and then with Databricks's own infrastructure, where they're giving you the ability to do this kind of thing now. Unless you're a really large enterprise, and you have multi-cloud strategy, you've got multiple vendors in for warehousing, data lake and LLMs, I think at that moment in time, at that scale, it makes sense to have a centralised trust layer. However, for the majority of companies who might be all-in on Azure, for example, and they're doing everything with Azure DevOps and with Copilot, I would struggle to see how they would differentiate themselves in a meaningful enough way to command a high ARR for a standalone service in that space.
来源背景：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 - 支持 使用 Snowflake AI 的账户从 5,200 增至 6,200 (同为 FY26 口径)
回 答 者：AI已成为用例的主要驱动力。截至本季度，使用Snowflake AI的客户或账户已超过6,200个，高于上一季度的5,200个，与去年相比也有所增长。具体来看，Q1为5,200个账户，Q2约为6,200个账户，这是主要原因之一。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 支持 Cortex Code 是上季度最热产品；客户端免费、不按用户收费，客户只需买 credits
回 答 者：3个月前，Snowflake 发布了名为 Cortex Code 的桌面客户端，类似于 Cloud Desktop。除数据集成外，客户还可以利用 Cortex Code 处理其他应用场景。此前，客户主要使用 Snowflake 进行数据提取、编写 SQL 查询等工作。现在，客户可以通过 Cortex Code 生成文档、制作 PPT，并开展深度研究和外部研究。由于该客户端免费且不按用户收取许可证费用，许多客户正通过使用 Cortex Code 增加消费量。因此，Snowflake 的策略之一是，客户无需按用户购买 Anthropic 或 OpenAI 的许可证，只需购买 credits，即可让公司内的每位用户和员工使用 Cortex Code 或 Cortex Code Desktop。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 收购 Natoma 获得 MCP 服务器能力，外部应用可调用 Snowflake API
Snowflake也在收购更多公司。Snowflake约在6至8个月前收购了一家名为Natoma的公司。Natoma为Snowflake带来了MCP服务器能力，而Snowflake此前并不具备MCP能力。完成收购后，客户可以从其他不同应用程序调用Snowflake API，包括与Salesforce、SAP、Slack等进行集成。通过增加MCP能力，Snowflake已经成为类似Salesforce的互联平台。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 Cortex Code 采用率约 50% (无确切数字)，很多客户同时用 Claude Code、GitHub Copilot
回 答 者：我没有确切数字，但可能在 50% 左右。许多客户也在使用 Claude Code、GitHub Copilot 等产品。Cortex Code 的优势在于，它预先了解客户的 Snowflake 资产、代码库和环境等情况。如果客户需要开展较大规模的 Snowflake 开发工作，使用 Cortex Code 确实合理。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 中性 Cortex AI Gateway (7 月底推出的模型路由) 不会立即带来收入，是必要的基础模块；欧盟 AI 法案限制跨公司动态换模型
回 答 者：这款产品不会立即推动 Snowflake 的收入增长或业务发展，但我认为这是必要的一步。首先，随着 AI 应用规模扩大，尤其是 AI 智能体工作规模扩大，这款产品的重要性会进一步提升，而目前我们仍处于相对早期的阶段。其次，从成本角度看，这款产品也很有价值。正如我此前提到的，行业一直在讨论如何为不同任务选择合适的模型。目前已有越来越多的工具能够帮助用户为相应任务选择合适的模型，我在其他平台上也看到了这种趋势。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 支持 超过 50% 采用 AI 的客户使用 Cortex Code，替代 Claude Desktop、OpenAI Codex 等桌面工具 (仅见于摘要稿)
是的，Cortex Code 确实推动了大量使用。超过 50% 采用 AI 的客户在使用 Cortex Code。Cortex Code 在很多客户中正在替代像 Cloud Desktop、OpenAI Codex 之类的桌面工具。
来源背景：Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。 - 口径不齐 摘要称 Cortex Code 使用率约 10% (采用率约 50%)，转写原文无此句
A:Cortex Code 应被视为对整体 Snowflake 消耗的加速器。它将 Snowflake 重新定位为一个主动的开发环境，可以使用 AI 代理来压缩开发和流水线构建时间，从而增加整体平台使用率。从系统集成商的角度来看，我们并不喜欢这种情况，因为我们更偏好软件开发复杂且耗时，那样可以增加我们的可计费工作量。所以在平台上有 Cortex Code 或 AI 编码代理会压缩我们计费的数据工程工作。我们在自己的交付团队中向 CloudCode 迁移也看到了同样的动态，即劳动力压缩效应。Cortex Code 不应被看作是最终产品；它更像是在与系统集成商生态争夺利润的手段。至于占比，很难说，因为它提高了 Snowflake 开发的速度，从而增加消耗。我们看到 AI 编码代理通常能将开发速度提高 30% 到 40%，这可能意味着采用率大致线性增长。这是我能给出的最接近的估计。目前 Cortex Code 的使用率大约是 10%。
来源背景：MeritForum 标为'北美软件渠道商专家'，但内容与 260806 Merit Snowflake季度调研III 转写一致 (Q4 29%/Q1 30%/Q2 31%/Q3 33-34%)，专家自述在北欧地区、是 Snowflake 合作伙伴兼经销商，标签应以转写为准。局限同该转写。 - 反对 Cortex Code 尚未真正使用，产品较新需观察成熟度
我们可能会继续在这方面开展开发和探索，但目前尚未真正使用，因为该产品比较新，还需要继续观察其成熟度，以及是否符合我们的要求。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 反对 Cortex Code 目前尚未带来加速，可能要到明年
回 答 者：目前还没有。我认为可能要到明年才会开始加速。许多客户尚未真正采用相关方案，因此预计明年会开始看到加速趋势。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 口径不齐 Cortex Code 约占 Snowflake 消耗 15-20% (与'尚未加速'自相矛盾，可能指 Cortex 整体或增量)
回 答 者：大量开发人员正在高频使用并采用Cortex Code，因此它确实提升了Snowflake的整体消耗，占比约为15%至20%。现阶段很难进行准确分析，因为虽然已有大量开发人员使用，并产生了大量数据，但目前还无法获得确切的统计结果。当前估计占比约为15%至20%，未来这一比例会高得多。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 反对 Snowflake 可观测性很难与 Datadog/Dynatrace 竞争，今年未见客户转向
Snowflake很难与Datadog和Dynatrace竞争。这2家供应商在各自领域的表现都非常出色，尤其是Dynatrace主要服务大型企业，而Datadog更多服务中端市场和小型企业。它们的竞争力很强，产品颗粒度也更细。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 反对 未见大量客户从 Dynatrace/Datadog 转向 Snowflake 可观测性
目前，我们尚未看到大量客户从Dynatrace或Datadog转向Snowflake。明年或许会有所变化，但今年Snowflake在可观测性平台方面尚未有太多以获取新业务为明确目的的行动。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 超过 50% 客户以某种形式用 Cortex Code，积极使用占 25-30%；受限于治理审批、已有 Cursor/GitHub 承诺和推理成本
A:我从来没说是30%。我说的是超过50%在以某种形式使用，但“正在积极使用”和“仅在试用”的人是有区别的。积极使用的占比是25%到30%。所以我不会说30%在定期使用——我认为超过50%在使用。我也不指望会有100%的普及率，即便采用速度很快，因为Cortex
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 可观测性：Snowflake 有可信机会但不会取代 Datadog/Dynatrace；跨模型 AI 监控后两者更强
A:就可观测性而言，我认为Snowflake有一个可信的机会，但我不会把它定位为完全取代Datadog或Dynatrace。它最强的机会在于Snowflake原生的数据和AI代理可观测性。他们今年动作很快，现在提供了带有细粒度token和信 用消耗记录的Cortex代理使用历史、AI函数成本监控、代理评估和Cortex代理监控——因此你可以获得执行和延迟的经济学信息。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 支持 Openflow 被低估：攻击独立 ETL 产品的支出，并便于把数据带回 Snowflake
在我看来，OpenFlow 被高度低估了，因为他们正在构建更强的摄取和集成层，包括与 Oracle 的相关活动以及与 Iceberg 的互操作性。它攻击了通常会分配给独立 ETL 产品的那部分支出。更重要的是，它使得将额外的客户数据带回 Snowflake 变得更容易。第三个变化是 Snowpark Container Services，它增加了自动伸缩和大幅更广的 GPU 基础设施，让他们看起来更像通用的数据与 AI 执行平台，而不仅仅是云数据仓库。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 Snowflake Postgres 未来 24 个月份额增长主要来自绿地 AI 应用，而非从 MongoDB 大规模迁移
在接下来的 24 个月里，我预计 Snowflake 会在运营数据库份额上获得增长，但初期将来自绿地项目和以 Snowflake 为中心的 AI 应用，而不是从 MongoDB 那里大规模迁移。战略性变化意味着 Snowflake 不再满足于仅作为数据的分析终点。所有这些——PostgreSQL、Unistore、Cortex agents、OpenFlow 以及相关的应用能力——都表明他们希望掌控更大部分的数据生命周期。这增加了与 MongoDB、Databricks 以及超大云提供商数据库的竞争。不过在短期内，PostgreSQL 最大的机会在于绿地的 AI 与 agent 式应用，而不是替代成熟的 MongoDB 库。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 Databricks 与 Snowflake 一样在推进 Postgres (LakeBase vs Snowflake Postgres)
目前，任何需要事务性状态的代理或应用都无法由 Lakehouse 支持。这就是 Databricks 通过 LakeBase 向 Postgres 数据库方向进军的原因。现在与 MongoDB 存在一些重合，但并不覆盖整个 MongoDB 市场。Databricks 正在构建一个生态，不仅把 LakeBase 作为一个 Postgres 产品，还通过近期收购的 PG Lite 来支持迁移，这使得转换变得容易并支持朝这个方向发展。Databricks 认为 OLTP 本身是一个数千亿美元的市场。我觉得他们会非常迅速进入该计算场景，尤其是在针对 Amazon Aurora、AlloyDB、Azure PostgreSQL 数据库、Supabase 以及部分 MongoDB 客户的计算上。我同意，和 Snowflake 一样，Databricks 也在寻求朝这个方向拓展。
来源背景：MeritForum 标题明示“databrick离职专家”，作者 Lantian ；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering,)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。 - 支持 Genie 主要拉动存量客户，CoCo 更开放、能爆发式带来新客户
Genie用户增长很快，目前 yoy增长大约八成，但没有 CoCo那么明显的加速。Snow是上市公司，有财务报表等，加速更强。Genie主要拉动的是存量客户，数据已经上来，Databricks建立了语义层。已有用户因为上了 Genie 用量会增加，但不像 CoCo那样因为 Genie 爆发式带来新客户。Databricks新客户一般还是从数据开始，先把数据放上来才能很好地用 Genie， CoCo那边更开放，用起来方便，可以问任何问题。不过CoCo在治理方面缺乏治理。有些普通员工可以查询公司内部数据，只要能访问的数据都能查到。Snowflake系统有两个 catalog 做治理Snowflake在 2024 年换了新 CEO，新 CEO 主推 AI，之前 Snowflake的 AI 推进较慢。
来源背景：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 - 支持 企业客户中 50-60% 已开始试用 CoCo
回 答 者：在企业客户中，大约有 50%-60% 的客户已经开始试用 Coco。大多数大型企业客户都已开始进行评估，因为他们的数据较为复杂且类型多样，需要使用 Coco。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 反对 试用客户中只有 20-30% 形成显著消耗；一个 500-600 注册用户的账户里只有 30-40% 会用 CoCo，天花板受开发者人数限制
在所有已经进行小规模投入并产生消耗的客户中，大约 20%-30% 已经形成了显著消耗，而且这一比例呈上升趋势。我们预计其他一些客户的使用量也会增长，但并非所有 Snowflake 客户都在使用 Coco，未来也不一定会全部使用。Coco 主要适用于拥有多个数据源等需求的客户，并且通常只由部分开发人员使用。如果有 500-600 名注册用户使用 Snowflake，其中可能只有 30%-40% 会使用 Coco。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 反对 Observe 采用率低，对用量的贡献仅约 2-3% (低于 5%)
So right? So there was double storage costs and also, you know, uh, open telemetry is gaining traction rather than agent- based solution from, uh, data dog and others. But in my view, we are not seeing a lot of traction with Observe, uh, from Snowflake at this time. Um, I think that's a complex business to get into, um, not something that Snowflake is, uh, is known for. So I would say the impact is less than 5% for consumption, right? two to 3%, not not more. Uh, but the PostgreSQL um, products are are gaining traction and um, you know, the the messaging around consolidating operational data store and analytical data store is is gaining some traction and I think will continue to do so. Uh, MongoDB is much more expensive compared to these offerings.
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 Snowflake Postgres 获得关注，Q2 讨论增加、Q3 形成项目储备、Q4 才有重大决策；潜力大于可观测性
我认为，至少还需要几个季度，才能判断是否会产生负面影响，最早也要到今年年底，即Q4。我们看到相关讨论在Q2开始增加，Q3已经形成项目储备，客户将在Q4作出一些重大决策。因此，我认为相关变化将体现在数据中。Snowflake和Databricks的Q4业绩至少会反映出这一趋势，MongoDB受到的负面影响也会在Q4显现，这可能就是我们看到影响的时间点。
来源背景：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 - 支持 模型从前沿闭源走向多样开源，增加数据集成复杂度，长期利好 CoCo/CoWork 需求
从仅依赖 frontier 模型转向多种开源或开放权重模型进入市场，会增加数据集成和训练的复杂性，客户将在其环境中拥有更多多样化的模型。总体而言，我认为客户对 Cortex Code 和 co-work 的需求在长期内会增加。
来源背景：标题'北美软件渠道商离职专家'；与 2026-08-19 Merit 音频转写为同一场访谈，身份局限见该条 - 支持 CoCo 早期有牵引且转化快于 CoWork (范围差异而非产品质量)
BJ: I think, great question. I think we've seen early stage traction on CoCo. I think it converts faster than CoWork. And I think that's a scope difference rather than product quality difference. CoCo is just a coding agent. A lot of people are already used to coding agents for local files, and it's leveraging Snowflake data, which was there.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 客户对 CoWork 类 ad hoc 查询成本防御性强，token 用量难预测导致决策停滞
It directs the whole question. I think business users are really worried about the insight that is to be trusted or not. I think what we're also seeing is early stage defensive posture on costing, right? It's more conservative, and the problem is that ad hoc queries don't behave like scheduled dashboard refresh, right? It is hard to forecast; a finance function that can forecast, it's a warehouse spend, which was easy, right? We could have predicted a few % every quarter, every time period. Now, what we don't know is if someone asks a question, what is the predictable token users associated with it? I think that does stall some decision-making within the organizations. So that's where some of the decisions get stalled because there's more analysis and paralysis that comes with it. At the same time, I think a lot of that is necessary to really model the governance structure so that we can at least from finance to functional technical aspects, be more reliable in the ability to build business cases and the contracts that are associated with those cases accordingly.
来源背景：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 - 反对 Observe 收购后进展不多，需要面向传统 IT 买家的新销售动作
1. We have not seen much progress since the acquisition. Traditional observability products such as Datadog and Dynatrace require customers to send logs to the vendor even when those logs are already stored in the customer’s own environment, leading to complaints about log storage costs. Snowflake’s approach through Observe is to store this data in Snowflake at a more attractive price and add intelligence through Observe. 2. For many Snowflake salespeople, Observe requires a new sales motion targeting traditional IT buyers rather than the data teams Snowflake typically serves. For these reasons, progress remains limited. 6. Competition Between Databricks and Snowflake
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 支持 CoCo 实质使用客户约 12%，前两季约 6-7%；广义使用率从 25% 升到 45%；CoCo 约占 consumption 3%
1. Approximately 25% of customers were using Cortex Code in 1Q26, and that figure has now risen to 45%. 2. Usage stages vary from trials to actual applications. They broadly fall into three categories: trials, pilots, or proofs of concept; general usage rather than active production environments; and actual production use through workflows, integrations, or implementations after large-scale Snowpark deployment. Together, these three categories account for approximately 45%. 3. Currently, approximately 12% of customers use Cortex Code in a substantive way, meaning they have active workflows or applications governed, processed, or supported by it. This was approximately 6-7% in the prior two quarters or 1H of this year. 3. CoCo Creates a Flywheel for Other Products
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 欧洲伙伴估计 CoCo 约占收入 2-2.5%，季环比接近 100%，长期上限约 10% consumption
1. Cortex Code, also known as CoCo, is Snowflake’s enterprise coding agent. It is growing quickly because adoption is relatively easy. I estimate it represents about 2% to 2.5% of revenue, which is already meaningful. It and Snowflake CoWork, formerly Snowflake Intelligence, are Snowflake’s two fastest-growing areas. The core platform remains the largest in absolute and incremental revenue. Coding agents are easier to adopt than agentic workflow automation, which requires workflow redesign and substantial implementation work. 2. Quarterly sequential growth is approaching 100%. Its share of consumption was well below 1% in 4Q2026 and has now risen to two or even three times that level. 3. CoCo could eventually reach 10% of consumption, but getting above 10% would be difficult. If it expands beyond Snowflake, the growth opportunity would be larger, but the core platform and Snowflake CoWork are also growing rapidly. I would not raise my long-term estimate for CoCo above 10%. Even reaching 10% would mean roughly quadrupling its share, and because total revenue is also growing, the absolute incremental revenue would be much larger. Using 10% as the ceiling is already quite aggressive.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 欧洲伙伴把 CoCo 长期上限定在 10%，并认为 10% 已很激进
1. Cortex Code, also known as CoCo, is Snowflake’s enterprise coding agent. It is growing quickly because adoption is relatively easy. I estimate it represents about 2% to 2.5% of revenue, which is already meaningful. It and Snowflake CoWork, formerly Snowflake Intelligence, are Snowflake’s two fastest-growing areas. The core platform remains the largest in absolute and incremental revenue. Coding agents are easier to adopt than agentic workflow automation, which requires workflow redesign and substantial implementation work. 2. Quarterly sequential growth is approaching 100%. Its share of consumption was well below 1% in 4Q2026 and has now risen to two or even three times that level. 3. CoCo could eventually reach 10% of consumption, but getting above 10% would be difficult. If it expands beyond Snowflake, the growth opportunity would be larger, but the core platform and Snowflake CoWork are also growing rapidly. I would not raise my long-term estimate for CoCo above 10%. Even reaching 10% would mean roughly quadrupling its share, and because total revenue is also growing, the absolute incremental revenue would be much larger. Using 10% as the ceiling is already quite aggressive.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 CoWork 年底占消费低于 8%，此前估计偏高
1. Similar to Cortex Code, my previous estimate may have been too high. I would view 8% as the ceiling, perhaps even an absolute ceiling, despite the product’s very rapid growth. 2. If Snowflake CoWork becomes an enterprise-grade or end-to-end process-automation platform, its long-term ceiling could be higher than CoCo’s, although its near-term contribution should remain below 8%. 4. Snowflake CoWork Can Serve as a Team Collaboration Tool
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 口径不齐 CoCo 账户 9,100+ (单季净增 2,000+)；CoWork 5,800 账户，环比仅约 +11%——Q1 时 SI 账户环比翻倍，CoWork 增速明显放缓
Its engineers are now using CoCo to accelerate the migration of 12 billion records into an AI-ready foundation. And as more customers see what's possible with this technology, adoption continues to build. CoWork expanded to 5,800 accounts, up nearly 11% quarter-over-quarter. Meanwhile, CoCo continues to see rapid adoption surpassing 9,100 accounts and adding more than 2,000 net new accounts in this quarter alone. We have customers like 1Password, the security company trusted by more than 200,000 businesses, which choose Snowflake for our CoCo capabilities. CoCo enabled their team to move key data pipelines into Snowflake, quickly laying the foundation for their data and AI work.
- 支持 CoCo/CoWork 带来新用户，采用账户内用户数阶跃增长 (未量化)
But the opportunity goes beyond adoption. By making it possible to build, collaborate and interact with enterprise data through conversational language, CoWork and CoCo are bringing entirely new users to Snowflake. Within accounts adopting these products, we see a step change in user growth as Snowflake reaches new lines of business and expands its footprint within existing teams. As we continue to develop CoWork and CoCo as agentic control planes, we are also building out the broader platform enterprises need to put AI to work at scale. Model choice gives customers the flexibility to select from leading frontier and open models and evolve their approach as the market changes, whilst training lets them adapt models to their specific data and business context and agent observability and analytics give customers full visibility into what their AI is doing, how it's performing, and what it costs.
- 中性 推出 Cortex AI Gateway，按客户策略动态路由到合适模型、内置成本与治理控制；整合 Natoma
As we continue to develop CoWork and CoCo as agentic control planes, we are also building out the broader platform enterprises need to put AI to work at scale. Model choice gives customers the flexibility to select from leading frontier and open models and evolve their approach as the market changes, whilst training lets them adapt models to their specific data and business context and agent observability and analytics give customers full visibility into what their AI is doing, how it's performing, and what it costs. And to help our customers optimize cost, performance and speed, we have introduced Cortex AI Gateway, which dynamically routes each task to the right model based on customer-defined policies and real-world performance data with cost and governance controls built in. As those economics improve, customers can deploy AI more broadly and with greater confidence, creating another catalyst for adoption and consumption on Snowflake. Cortex AI Gateway also extends AI from insight to action through its integration of Natoma.
- 中性 Frontier engineering (FDE) 项目按结果收费，只有交付结果才拿钱
We are, for example, helping a big financial institution effectively overhaul their digital and data strategy and bring it to the modern world in a way that is very, very sustainable for them. And the confidence that we have going into these kinds of engagements is not just that we commit to delivering the outcome. Obviously, we get paid only when we deliver outcomes in situations like this, but it's also in the fact that Snowflake is an open, well-understood platform. And compared to some pretty proprietary folks out there where you have to go back to them after you get the first outcome, we can confidently tell them that their data team is very, very capable of driving further engagement with the projects that they have done and building on top of it. It's the combination of these things: our ability to truly talk about business outcomes, commit to delivering them, but deliver it on a clean, open, well understood architecture that makes the customer looks good and stay good that I'm most excited by.
- 中性 会更多谈 hybrid tables 和 Postgres，作为新一代 agentic 应用 (含 system of record) 的基础
Yeah, and our investments in which Christian has pioneered and spearheaded with the team for a very long time around being able to host applications in Snowflake, small and big, also positions us exceptionally well for many applications, not just analytic ones, but also systems of record, operational ones that can be built right on top of Snowflake. And so internally, we have many projects, some of which Christian and I like don't even know of people that are building interesting applications on top of the analytic data and operational stores that they're setting up within Snowflake. You can definitely expect to hear a lot more about things like hybrid tables and Postgres, because they are the foundation, we think, for a new generation of agentic applications, some of which will have UI and some of which won't on top of Snowflake. Thank you. Thank you.
- 支持 CoCo 账户超 9,100，本季净增 2,000 以上
> “CoCo continues to gain adoption rapidly. We now have more than 9,100 accounts, with more than 2,000 net new accounts added this quarter alone.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 口径不齐 CoWork 账户 5,800，环比仅增约 11% (远慢于 CoCo)
> “CoWork has expanded to 5,800 accounts, up nearly 11% sequentially.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 推出 Cortex AI Gateway，按客户策略和实时性能把任务路由到合适模型
> “We introduced Cortex AI Gateway. It dynamically routes each task to the right model based on customer-defined policy and live performance data, with built-in cost and governance controls.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 OLTP/OLAP 融合上 Snowflake 发展更快更成熟 (hybrid table 支持行级锁、主键、二级索引，宣称 benchmark 提升 8 倍)，Databricks Lakebase 7 月才 GA
融合确实非常关键。Databricks看到客户已经被市场教育，不再需要分别建设两套系统，应该可以用一套方案来满足需求。从目前来看，Snowflake在这方面发展得更快，也更成熟。今年（2026年）他们的hybrid table迭代非常快，最近又推出了新的功能，比如支持row level locking、primary key和二级索引。今年还宣称benchmark提升了8倍，8月份又支持了OLTP的inline存储过程。相比之下，Databricks是在7月份才宣布GA他们的LTAP产品，目前还处于非常早期阶段。虽然他们可以用一个副本完成两种类型的任务，但真正的transactional database不仅要支持查询，还需要像Postgres那样支持存储过程、trigger、二级索引等功能，这些目前Databricks都还不支持，只具备最基本的transactional功能。如果客户已经有很多Delta Table（OLAP体系），想要创建Lake Base，初始时在Databricks的UI上其实是创建了一个新的副本，并不是简单地复制原有数据，而是对columnar storage上的packet文件和JSON文件做了格式转换，以适应未来OLAP和OLTP的同时使用。所以总体来看，Snowflake在这方面比Databricks走得更快，产品也更成熟一些。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Databricks 现任员工，负责 GCR/APAC 区域 (给出 GCR/APAC 约八成项目仍在 PoC，了解内部 Omnigent '零号用户' 使用情况与销售策略)。视角局限：竞争对手内部人员；对 Snowflake 是外部观察 ('不太确定 Snowflake 的情况')，尤其是中国市场的 Snowflake；样本为亚太。 - 反对 Snowflake 有 hybrid table 和 Snowflake Postgres 两套体系，不止一个客户反映定位模糊、互相竞争，增加维护和人员成本
但Snowflake的问题在于它采用了两套体系：hybrid table和Snowflake PostgreSQL。在构建agent时，不止一个客户反馈，觉得Snowflake的产品定位比较模糊，比如他们要建agent时会问到底该用hybrid table还是Snowflake
来源背景：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Databricks 现任员工，负责 GCR/APAC 区域 (给出 GCR/APAC 约八成项目仍在 PoC，了解内部 Omnigent '零号用户' 使用情况与销售策略)。视角局限：竞争对手内部人员；对 Snowflake 是外部观察 ('不太确定 Snowflake 的情况')，尤其是中国市场的 Snowflake；样本为亚太。 - 支持 设立 Activation Engineer 新岗位，专门让新客户更快上线
And so there's this big shift. We've even created basically new job functions. One of them is called an Activation Engineer, an Activation Solution Engineer that are expressly charged with what Brian said, which is, how do you get a new logo to go live on Snowflake much faster than what they would otherwise? And so, that is a trend that we want to keep leaning into and pushing. I would also say selling into or actually describing what we do to all these personas.
- 支持 CoWork (Snowflake Intelligence 的 CoWork 形态) 定位为面向知识工作者的统一 AI 工作空间，可经 MCP 连 Slack/邮件/Google Calendar、生成 Slides 并交付——目标用户从数据团队扩到业务人员 (潜在新增查询来源)
基于这一逻辑，今天想向大家介绍Snowflake Co-work。它需要与Copilot加以区分。Co-work是面向知识工作者推出的统一AI工作空间，并不只是一个Chatbot。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 支持 CoWork 已推出 Automation，可按日/周/月/季定时执行周报、QBR——定时任务意味着可重复、持续的后台查询负载
回 答 者：不需要。我们现在已经推出 Cowork 的 Automation。只要预先设置好相应的工作流，就可以规定执行频率，例如每个月或每个季度，主动将汇报信息发送给我。这样，我不再需要担心遗漏信息。
来源背景：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 - 支持 Snowflake 安全负责人基于 100+ 次客户交流称，AI 治理是推动 AI 部署的催化剂而不是阻碍
Um, my name is Sebastian Weitzner, I'm the security practice lead at Snowflake. And it's an absolute pleasure to be with you here today to talk about what is probably becoming my absolute favorite topic, um, and that is the security and governance of AI. The reason it's my favorite topic isn't just because it's super interesting, which it is, but because after more than 100 customer conversations I've come to realize that AI governance isn't a tick box on the way, standing in the way of products being sold. It is actually a catalyst that really lets you go at speed and gives you the confidence and trust you need to move forward, deploy your great innovative ideas. I was supposed to have my intro slide, but that's... anyways.
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 支持 Snowflake 在开发 AI 治理仪表板，统一展示所有 agent、MCP 服务器等 AI 资产 (称 Q2 推出，先预览)——把自己定位为 agent 的治理控制面
对于对 AI 持有其他观点的人，我还想说明一点。此前我提到过数据治理仪表板。我们目前正在开发面向 AI 的数据治理仪表板，计划于 Q2 推出。该仪表板能够提供整体的端到端视图，涵盖生态系统中的所有智能体、MCP 服务器，以及未来可能来自平台的任何内容。该功能很快将开放预览，敬请关注。
来源背景：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 - 支持 CoCo 有四种入口：CLI、Desktop、VS Code、Snowsight Web UI；企业 IT 不允许装第三方软件时可直接在 Snowsight 用——降低企业采用门槛
回 答 者：使用 Coco 时，可以选择 Coco CLI，也就是命令行界面；也可以选择 Coco Desktop，开发者可通过该桌面工具构建应用和数据工程管道；还可以通过 Web UI 使用 Snowflake 中的 Coco，无需在笔记本电脑上安装任何软件，直接在 Web 界面中使用。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 支持 CoCo 设计理念是融入开发者已有工具 (与 Claude Code/Cursor 同类的 agentic 编码形态)，'开发者不必学习新工具'
专家：Visual Studio Code 是 Microsoft 的 IDE，CoCo 也可以在 Visual Studio Code 中使用。其界面完全相同，开发者可以继续使用自己熟悉的操作界面。核心理念是，开发者和工程师不必学习新工具，Snowflake 和 CoCo 会融入现有工作环境。无论是命令行界面、Snowflake 桌面应用程序，还是 Microsoft Visual Studio Code，都可以使用 CoCo。例如，我可以输入“构建一个管道”，CoCo 就会为我构建管道。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 支持 CoCo 可直接调 Marketplace 数据集 (含免费数据集) 建应用和训练模型——把 Marketplace 数据纳入 CoCo 工作流
专家：可以直接提出“显示市场中所有可用的数据集”这样的请求。市场中有很多不同的数据集，可直接用于今天的项目，无需担心如何将数据导入 Snowflake。这也是 Snowflake 的默认界面。如果无法使用命令行界面或 CoCo 桌面应用程序，也不必担心，只需登录账户，即可立即使用 CoCo，无需任何外部依赖。
来源背景：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 - 支持 Snowflake 把模型、企业数据与上下文、应用集成汇聚到'agentic control plane'，整合为 AI 数据云——公司自己在讲'控制面'叙事
所有这些要素最终汇聚到智能体控制平面。我们将展示Snowflake如何提供此类技术并满足相关需求。
来源背景：Snowflake 官方 Data for Breakfast 上海站录音。讲者：(1) Christian，Snowflake 高管 keynote ；(2) 冯金鹏，Snowflake 中国区解决方案顾问 。视角局限：高管愿景演讲 + 售前演示，无任何财务、用量或计费数据；演示企业为虚构/示例场景，转写质量很差 (中文演示部分大量缺字)。 - 支持 CoCo Q1 覆盖约 1,700 客户，co-work 超过 2,500 客户
第一方 agent 的采用也非常可观。许多客户将 Snowflake Coco 用作编码和数据 agent，而不是使用诸如 Acrobat 或 Codex 等工具。Snowflake Coco 在第一季度覆盖了约 1,700 个客户，较去年增加了近 1,000 个账户。在所述具体季度中，超过 2,500 个客户采用了 co-work，体现了 11% 的同比增长。Snowflake 的营收指引也显示，纯粹归因于 AI 的收入有了显著增长，同时与 AI 工作负载相关的毛利率为 34%。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 部署顺序先 CoCo 后 co-work：CoCo 做语义建模建数据基础，再在其上建 co-work、Cortex Analyst/Search agent
客户也使用 Coco 来构建许多 co-work agent。每个 co-work agent 都需要数据基础，客户将 Coco 作为语义建模环境来创建该基础。然后他们在数据基础之上构建 co-work agent、Cortex Analyst 和 Cortex Search agent。因此部署顺序通常是先 Coco、后 co-work。客户在 Coco 中进行开发，Coco 主要作为编码 agent，可用于编写 SQL、Python、Java 和 Scala。这些是当前客户部署中 Coco 的主要用途。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 支持 客户用 Coco Desktop 取代 Codex 和 Claude Desktop，公司签一份 Snowflake 合同即可
客户也在用 Coco Desktop 取代 Codex 和 Claude Desktop 等工具。Claude Desktop 仅能访问 Anthropic 的模型，而 OpenAI 仅能访问 OpenAI 的模型。Snowflake 让客户在一个地方访问来自多家基础模型公司的可用模型。客户无需购买单独的订阅，可以为整个公司签一份 Snowflake 合同并从那里部署。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 中性 客户通常以年度合同购买固定美元额度，无容量上限，按消费计价
Snowflake 包含摄取、转换、ML 平台、生成式 AI 平台、LLM 或 ML 推理端点、MCP 连接性以及与 Tableau 和 Power BI 等工具的连接。因此，客户无需离开 Snowflake。他们通常以年度合同购买规定美元金额的服务，没有容量上限。定价基于消费，允许客户从小规模开始，随着用例扩大逐步增加消费。
来源背景：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 - 反对 SQL 中的 AI 函数/批量推理两家几乎一样，没有差异化
database quite fast and governed pretty easily. And I think that's another advantage that Databricks had. In terms of... And early on, I think Databricks built their own model that I think under the purchase of MosaicML definitely helped. But then Mosaic, I think, focused more on the training and for inference, I think the partnership mattered. And I think in that front, they're pretty similar, I would say. How much Databricks or Snowflake go buy their own GPUs and serve open source models, I don't know the contrast there, but they both obviously work with third-party Anthropic and OpenAI. In that regard, I don't think there's any difference. They just use their endpoints. When it comes to fine-tuning and some of the training stuff, Databricks has some innovation around reinforcement learning, but that's not much of a revenue driver, but it's definitely a driver for credibility and the Mosaic AI team has built lots of capabilities around applying reinforcement learning and supervised fine-tuning post-training stuff to your data. I don't think there's much of an adoption there, but I think even that front for some advanced digital native companies, Databricks' AI is a bit more attractive and easier to access. The areas that are pretty common is like batch inference, like AI functions to SQL; they're pretty identical, maybe governed a bit differently, but both of them have it using AI functions and exposing it in LLM and using SQL to run batch inference on massive amount of data. Yeah, go ahead.
来源背景：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 - 反对 Cortex 类问数产品留存不确定、粘性弱、换平台门槛低，获客速度在下降
Cortex这类问数产品的客户留存存在较大不确定性，用户流失的可能性较高，因为其产品粘性不强，用户更换平台的门槛很低。虽然Cortex的获客速度在下降，但其能否留住新增用户仍需进一步观察。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 反对 Genie 100 个 space 中 90% 一段时间后不活跃，Cortex 早期增长可能部分是尝鲜效应
Databricks的对应产品是Genie。为了解决单纯问数场景粘性不足的问题，Databricks正引导客户从Genie转向Agent Bricks。此举旨在推动客户解决更复杂的业务问题，应用场景从仅对结构化数据提问，扩展到将非结构化数据也纳入分析范围。通过构建更高级、更复杂的应用场景，来提升用户粘性。用户粘性不佳的具体表现包括：用户在不同平台间（如Databricks、Snowflake、BigQuery）频繁切换，或者在平台内创建了大量工作空间（例如100个Genie space），但一段时间后，其中90%都变得不活跃，这表明初始使用可能更多是出于新鲜感。Cortex的上线时间晚于Genie，因此其早期用户增长可能部分得益于市场的尝鲜效应。
来源背景：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 - 支持 CoCo 约 7,100 客户 (同比增近 1,000 账户)，CoWork 单季 2,500+ 客户；注意未给收入
此外，我们也看到第一方智能体得到大量采用。目前，许多客户将Snowflake Coco用作编程和数据智能体。与Anthropic和Codex相比，我们看到许多客户正在使用Snowflake Coco。第1季度，Snowflake Coco的客户数量已达到约7100家，较去年增加近1000个账户。仅在本季度，就有超过2500家客户采用了Co-work，同比增长约11%。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 中性 部署顺序：先用 CoCo 做语义建模/数据基础，再基于 Cortex Analyst/Search 建 CoWork 智能体
对于每个Co-work智能体而言，数据基础都非常重要。客户使用Coco作为语义建模环境来构建数据基础，随后基于该数据基础，或基于Cortex Analyst、Cortex Search智能体，构建Co-work智能体，相关能力也在逐步提升。因此，第1阶段是Coco，第2阶段是Co-work，这就是部署顺序。客户先在Coco中开展开发工作，所以有大量用户使用Coco进行开发。
来源背景：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 - 中性 CoCo 本身不单独收费，客户为它驱动的消费付费
And so, you're reaching more people within a business. So typically, pre the CoCo and all the AI functions, you would basically interact with Snowflake through your data team who would actually write this SQL queries. Now anybody within the business can get access and we don't charge like CoCo is not -- we don't charge for it. You just get it being a Snowflake customer, and then the charge is the consumption that you actually draw in using the product. And so, we see more personas, and then we're seeing migrations go faster. And so, we're seeing more data come in, relative migrations.
- 支持 CoWork 已有 5,000 多个账户
And on those new products, I mean, what are kind of the top two or three new products that are now out there that you're most excited about or customers are seeing kind of initial traction with? Absolutely. CoCo's sibling is called CoWork. And so CoWork has over 5,000 accounts. CoWork is what I'll call data with rails around. It's like certified by the data scientists.
- 支持 约 9,100 个账户采用 CoCo，5,000 多个采用 CoWork
Now we have two quarters' worth of data, so we're getting closer to what that number is. But we're really happy with how we've had adoption across a number of different accounts. And so on the call, we talked about roughly 9,100 accounts have adopted CoCo. A little over 5,000 have adopted CoWork. And what we're trying to focus on now is get our sales team to go in and talk about the use cases to drive deeper penetration in those accounts so we can get more users. And so that's what we're in the process of doing.
- 中性 下一步是让销售推动账户内更深的用户渗透 (说明目前渗透浅)
And so on the call, we talked about roughly 9,100 accounts have adopted CoCo. A little over 5,000 have adopted CoWork. And what we're trying to focus on now is get our sales team to go in and talk about the use cases to drive deeper penetration in those accounts so we can get more users. And so that's what we're in the process of doing. But it's really a – the 5% comes down to the core business is doing well and is accelerating.
- 支持 Snowflake 自己取消了所有 BI 和分析软件合同，用 CoCo + Streamlit 替代
But we're currently working as much as possible to bring down the migration timeframe. For myself, I use it for, call it, like, analytics and business intelligence. And so, we've actually canceled all of our BI and analytics contracts. And so, I use CoCo to Streamlit and can spin up dashboards in a heartbeat with whatever I want around the business, whatever time period. They aren't static. They're interactive.
- 反对 Snowflake Postgres 与 MongoDB 差距仍在，很多客户仍把 Snowflake 视为 OLAP 报表工具
我认为差距还是存在，不一定会变得非常小，因为两者的核心和赛道不同。Snowflake 的 Postgres 最大的价值是让 OLTP 的数据更靠近 Snowflake 后端的 OLAP，减少 MongoDB 到 Snowflake 的 CDC 或 ETL。公司也希望往 OLTP 方向发展，但还要看客户的业务情况。客户一个月的使用行为和习惯，很多客户还是把 MongoDB 定义为在线数据库，即使 Snowflake 想做 OLTP，很多客户还是认为它是 OLAP，用于报表分析。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 中性 CoCoEvolve 尚未 GA，目标是把人工反复调 prompt 的过程变成可并行、可验证的自动优化闭环，且可以不依赖 Snowflake 数据库运行
CoCoEvolve 目前还没有 GA，它是在现有 Agent 能力的基础上，比现在的 CoCo更灵活，也能更深入地解决一些问题。现在的 CoCo主要是让客户自主管理流程，甚至可以帮助客户写程序，并不一定要和 Snowflake 数据库集成。当然，大部分用户用 CoCo 或未来用 CoCoEvolve，还是会和 Snowflake 集成，否则没必要用 Claude Code 这些 Agent。但 CoCoEvolve 和 Cortex Code 相比，更加灵活，即使不用 Snowflake 数据库，也可以单纯用 Agent 和大模型来完成一些任务。
来源背景：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 - 支持 Snowflake 环境内就能获得被商品化的 ETL/目录能力，按用量计费、不受产品许可限制，挤压点工具
FH: Well, the needs are there, but... When the capabilities are commoditized, right? Like I can get those within my Snowflake environment, right? My platform provides that. And so the hyperscalers are, you know, more of a... we'll give you access to everything. We're just going to charge you for what you use. I'm not being limited to, hey, I'm only licensed for the BI product or the ETL product or the data catalog. So that shift is an aspect of it.
来源背景：Frank Hernandez，前 Qlik 美洲区专业服务 Practice 与解决方案负责人 (约 2018 年至 2025 年底，明示/)；主持人确认其 2026 年起加入 Snowflake 并任职至今 ；25 年以上技术经验，约 20 年数据相关 。视角局限：现任 Snowflake 员工，对 Snowflake 的正面评价有明显利益偏向 (如 'every day I wake up and I'm super excited')；访谈主题是 Qlik，Snowflake 内容是侧面。 - 支持 Snowflake 用 AI 覆盖管道、目录、数据质量到洞察的全链条，使点工具难以竞争
And there's a battle between custom ad hoc stuff that can be governed within a platform; that's still an issue. So, I think of the innovation to really accelerate each of their platforms leveraging AI is definitely an opportunity that would have helped create more stickiness or longer-term stickiness. But when you weren't there and you didn't have that, and all of a sudden you get an organization like Snowflake saying, I can help you build out your pipelines, I can help you catalog your data, I can help you incorporate data quality into trust factors, and then I can help you present the insights and you have the ability of doing that with AI, they really accelerate it; it becomes really challenging at that point.
来源背景：Frank Hernandez，前 Qlik 美洲区专业服务 Practice 与解决方案负责人 (约 2018 年至 2025 年底，明示/)；主持人确认其 2026 年起加入 Snowflake 并任职至今 ；25 年以上技术经验，约 20 年数据相关 。视角局限：现任 Snowflake 员工，对 Snowflake 的正面评价有明显利益偏向 (如 'every day I wake up and I'm super excited')；访谈主题是 Qlik，Snowflake 内容是侧面。 - 支持 新 CEO 后 Snowflake 在 AI 领域步伐加快，SPCS 和 Hybrid Table 使其在 AI 投入和产品进化上有所超越
Databricks与Snowflake在多个层面存在显著差异。Snowflake的核心优势在于其产品的简易性，客户通过图形化界面即可轻松上手。自2024年6月新任CEO上任后，Snowflake显著加快了在AI领域的步伐，其Cortex产品发布虽晚，但后续推出的Snowpark Container Services（Snowpark容器服务）以及今年（2026年）发布的Hybrid Table等产品，使其在AI投入和产品进化上有所超越，例如Cortex已拥有7,000个用户。
来源背景：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。
E06 预期和仓位
判断：预期很高。市场已经在按"每季 beat 3–5% + 上调全年"建模，Q2 只超出买方预期约 1%。
指引方法：CFO 说核心业务每天凌晨按客户、产品线预测，精度很高；新产品基本不计入 (CoCo 首季没纳入，第二季开始纳入但打了折扣)；"3% 的 beat 就是很强的 beat"。历史上四个财年的首次指引都低估了 5–10 个点。FY27 的轨迹：2 月 +27% → 5 月 +31% → 9 月 +36% (一个季度上调 500bp，公司历史上最大的上调之一)。Q1 上调的最大驱动是 CoCo，Q2 在核心和 AI 之间更均衡。
现在的指引隐含什么：Q3 指引 $1.588–1.593B (+37–38%)，加惯常 3% beat 约 +41–42%。全年 $6.07B 减去前三季，隐含 Q4 约 $1.65B，环比只加约 $6,000 万 (Q2 加了 $1.58 亿)，这是典型的保守 Q4。
卖方和买方：TMTB：Q2 产品收入 $1.49B (+37%) vs 卖方 $1.42B (+31%)、买方 bogey 36%；Q3 指引 +37.5% vs 卖方 +29%、买方 32–33%；盘后 +20%。Funda：超卖方 5.2%，但只超买方约 1%；9 月初说软件的空头回补基本完成、跑赢已显拉伸。5 月时 KeyBanc 还在警告前瞻支出意愿恶化、CoCo 采用数据弱于管理层说法 (事后看错了)。Merit 的路演也提醒：渠道看到的增长仍温和，超预期主要在前瞻指标，股价更多是估值修复。
That same team as well as another team builds detailed product launch models for us. And so, we have all the products over the course of Snowflake, we have all the curves and the trajectory of what those products have done. And then we apply the pattern recognition behaviors to those based on what we're observing. And so, with CoCo only being out two quarters, the first quarter that we did it, we didn't include it because we didn't have any of their behavior. Now we're able to, the next where we're able to but we don't want to put in the initial sort of small numbers, but really, really high growth percentages. And so, we took a haircut to that. But now that we're two quarters in, we have more data. We feel like we have a better accuracy and we're guessing it. Then we have a whole bunch of new products that obviously are new.
But congrats from me as well. Amazing quarter. Thank you. Let me emphasize that there's been no change in guidance philosophy, and we view a 3% beat as a very solid beat. The difference that happened this quarter was CoCo was launched in the quarter, and we base guidance on observed behavior, and so we didn't have any observed behavior for guidance for CoCo. And so we had a unique opportunity now, since we've been able to watch it for a quarter, to layer that in now for the full year, and that's what we've done. And then we also saw the acceleration in the core, and we've included that for the full year based on what we've seen.
* **F2Q product revenue:** $1.49B, +37% y/y vs. $1.42B, +31% and bogeys at 36% * **F3Q product-revenue guide:** $1.591B, +37.5% vs. $1.50B, +29% and bogeys of 32-33% * **FY27 product-revenue growth:** 36% vs. 31% prior
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。
* **F2Q product revenue:** $1.49B, +37% y/y vs. $1.42B, +31% and bogeys at 36% * **F3Q product-revenue guide:** $1.591B, +37.5% vs. $1.50B, +29% and bogeys of 32-33% * **FY27 product-revenue growth:** 36% vs. 31% prior
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。
* FY27Q2 product revenue grew 37% year over year, beating sellside consensus by 5.2%. That matched the prior quarter’s beat, and revenue also came in 1% above buyside expectations. * FY27Q3 product revenue guidance implies 37% to 38% year-over-year growth. Applying a typical 3% beat would imply roughly 42% growth. Even after adjusting for the tough FY27Q2 comparison, FY27Q3 still represents an acceleration. * Full-year product revenue growth guidance was raised from 31% to 36%, one of the largest annual guidance increases in Snowflake’s history. * FY27Q3 non-GAAP operating margin guidance is 15.5%, also above expectations. Product gross margin is modestly lower because of CoCo, but companywide margin expansion remains clear, supported by faster growth and AI-driven organizational efficiency.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。
The software vs semi split that defined last week finally started to rebalance. Software that printed what the channel had been pointing to kept working through most of the week, while the semi tape spent the first half getting tagged by rates — the 10-year pushed to 4.8% on Tuesday, the highest since January 2025, after Warsh’s hawkish Jackson Hole turn had traders pricing hikes rather than cuts. By Friday the setup flipped: semis clearly outperformed, and memory names broke out. Near-term software short cover looks largely done, the outperformance looks stretched, and the momentum unwind in semis has already gone to an extreme — which leaves less of that particular overhang in the names that got hit hardest.
来源背景：作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。
KeyBanc says recent SNOW surveys were mixed ahead of earnings, with April consumption trends remaining healthy/in-line while forward spending intentions deteriorated meaningfully. The firm flagged weaker NTM partner growth expectations, potential F2Q guide risk (~24% vs Street 26%), and softer-than-expected Cortex Code adoption data versus management commentary. However, KeyB also notes Snowflake continues gaining strategic relevance versus Databricks among many enterprise customers, while core consumption trends into F1Q remain constructive, maintaining OW and $200 PT.
来源背景：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。
不过，仔细分析这些公司的业绩，其超预期部分主要体现在前瞻性指标上，例如净增长的AR及其他净增长前瞻指标，实际收入增长并没有那么快。因此，我认为这个季度股价的表现更多反映了估值修复。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。
Sanjit K. Singh (Morgan Stanley): congrats on reasserting 30% product revenue growth in Q4. Brian G. Robins: it's really based on the high stable growth that we see in our core business. Brian G. Robins: It's also the growing contribution from AI workloads. Brian G. Robins: The margin profile for those right now aren't as high as the core business, Brian G. Robins: but we're offsetting that by finding more efficiencies in the core business. Sridhar Ramaswamy: There are a lot of wins to be had with AI both above the gross margin line Mark R. Murphy (JPMorgan): can you describe the $400 million deal in terms of the customer type. Sridhar Ramaswamy: We didn't do anything particularly special in the quarters. Sridhar Ramaswamy: it represents a maturity of Snowflake as a durable provider, Brian G. Robins: when the big contract, over $400 million, it was an existing customer, so it's already built into the run rate. Brad Zelnick (Deutsche Bank): what are you going to do differently in the field to win and drive upside in fiscal 2027? Sridhar Ramaswamy: what drives momentum for the whole company Sridhar Ramaswamy: I have never seen more excitement from our sales force about the products that we create. Brian G. Robins: It's no longer tied to head count. Chirag Ved (Evercore): How does Observe fit into that topography and what were you seeing in the market Sridhar Ramaswamy: Observability, especially in the world of AI, is a big deal. Sridhar Ramaswamy: There is a huge overlap between potential customers of Observe and customers of Snowflake. Brian G. Robins: we guide based on the Observe customer behavior up until really the point of earnings. Sheldon McMeans (Barclays): how are you working with customers to help reduce the risk of cycles of strong growth and optimization? Sridhar Ramaswamy: absolutely, we don't want our customers to have sticker shock, we want to be predictable Sridhar Ramaswamy: At Snowflake, we charge based on consumption, and we therefore offer a very predictable model. Sridhar Ramaswamy: So, we will be launching features like a per-user cap on top of Snowflake Intelligence Koji Ikeda (BofA): any air pockets where you were surprised that they didn't contribute more Brian G. Robins: Overall, we had a good sales execution quarter Brian G. Robins: just really points to the business outcomes that we're driving for our customers Matthew Hedberg (RBC): is AI-related products, is that some or all of the kind of the upside Sridhar Ramaswamy: we see a lot of upside to where the business can go. Sridhar Ramaswamy: The level of innovation that we are seeing across the company is pretty inspiring Christian Kleinerman: We definitely see in the telemetry activity on the platform being increased based on the ease of use Brent Thill (Jefferies): first, do you think your customers understand that yet? Sridhar Ramaswamy: No AI model is going to help you if there are four sources of the truth. Sridhar Ramaswamy: our secret sauce, which has existed since the beginning of the company, Sridhar Ramaswamy: We support Iceberg as a first-class construct within Snowflake. Sridhar Ramaswamy: They get the manageability that they get with Snowflake while feeling confident that another engine can read that data. Sridhar Ramaswamy: building dbt pipelines to run data engineering on that data, or to build dynamic tables, Sridhar Ramaswamy: factors away; not like 10%, 20%; factors more efficient. Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
展开本节全部 63 条原文 (支持 30 / 反对 5 / 其他 28，按时间)
- 中性 管理层自认 Q3 beat 仅约 2.5%，低于其'3% 是好 beat'的标准
Sridhar Ramaswamy: partnerships with Anthropic, which include a buy, but also a broader go-to-market motion Brad Reback (Stifel): The op margin obviously down-ticked on the first half sales and marketing investment. Sridhar Ramaswamy: We clearly invested pretty heavily in our sales and marketing teams in the first two quarters Sridhar Ramaswamy: I don't think of this as an either/or. Mike Cikos (Needham): Is it fair to think that 3Q here was just a strong execution quarter Sridhar Ramaswamy: We've consistently told all of you that we view a 3% beat, as a very good beat Sridhar Ramaswamy: I view the Q3 beat as actually still a very solid beat at some 2.5%. Sridhar Ramaswamy: large migrations are lumpy and not all that easy Sridhar Ramaswamy: there is also a natural variability in a consumption business Brian G. Robins: the quarterly variability is not the right way to evaluate the consumption model.
- 中性 唯一意外是 hyperscaler 宕机 (影响约 $1-2M)
Sanjit K. Singh (Morgan Stanley): Brian, first for you, when we look at the growth rates on product revenue this quarter, really attractive at 29% Brian G. Robins: There was really only one surprise in the quarter, and that was a hyperscaler outage Sridhar Ramaswamy: Snowflake Intelligence amplifies the investments that our customers have made Kirk Materne (Evercore): I was particularly interested in the 600 new customer wins. Sridhar Ramaswamy: Previously, we used to be just the analytics provider, but we can be there from soup to nuts Sridhar Ramaswamy: the complete data offering will end up playing a larger and larger role Brent Thill (Jefferies): And when do you expect this batch of go-lives to go up Sridhar Ramaswamy: We gave guidance for Q4. It's a pretty hefty beat and raise.
- 中性 新 CFO 承诺指引哲学不变 (保守、基于历史消费)
Sridhar Ramaswamy: bringing Postgres to market will be an important step forward for us Sridhar Ramaswamy: we will continue to have both of these Tyler Radke (Citi): are you expected to see significant growth in those deals? Sridhar Ramaswamy: Product revenue is the best indicator of how our collective customers are going to be spending Brian G. Robins: I can assure you that there will be no change to the guidance philosophy. Brian G. Robins: Tyler, as you mentioned, we'll guide to FY 2027 on our next call.
- 中性 FY27 产品收入指引 $56.6 亿 (+27%)
Brian G. Robins: We're guiding Q1 non-GAAP operating margin of 9% Brian G. Robins: For FY 2027, we expect product revenue of approximately $5.66 billion, representing 27% year-over-year growth. Brian G. Robins: We expect Observe to contribute approximately 1 percentage point of product revenue growth in FY 2027.
- 中性 指引沿用现有消费模式，保守
Robins: Our hiring this year will be weighted to the first quarter, reflecting the addition of 178 employees from Observe. Brian G. Robins: As always, our forecast is built on using existing patterns of consumption. Brian G. Robins: We collect the majority of our cash in the fourth quarter.
- 反对 分析师问 AI ARR 更新，CFO 只答 FCF——回避
Ryan MacWilliams (Wells Fargo): what do you think are some of the advantages that Snowflake has to be the best place Sridhar Ramaswamy: our motto always has been easy, connected, and trusted, Sridhar Ramaswamy: we are stewards of their most important data Alex Zukin (Wolfe Research): Yeah, Brian, I was just going to ask if any update on the Snowflake AI ARR and then the free cash flow margin guide. Brian G. Robins: Yeah. Just on free cash flow overall, the seasonality will follow prior years. Alex Zukin (Wolfe Research): Maybe just talk specifically about kind of what you saw post-holiday in January Sridhar Ramaswamy: the guidance process is a pretty strict one that focuses on historical information Sridhar Ramaswamy: it is very much about using every bit of data that we have until this moment.
- 支持 RPO 增速由 38% 加速到 42%，可能是收入加速前置指标
RPO从38%加速增长到42%，可能是2027年收入加速的前置指标
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 支持 研究员：SNOW 是少数 AI 带来加速且无 margin 负面的股票，但估值较贵；模型 2026 年 NonGAAP 净利 8.7 亿、2027 年 13 亿美元
* Cortex Code + Iceberg + AI Labs/AI Native Customer增加消耗仍然会带动收入加速逻辑 * 更多还是估值问题，我的模型仍然估算2026年870mn NonGAAP NP，2027年1.3Bn NonGAAP NP
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 支持 Benchmark 目标价 $190→$200，认为 FY27 能维持约 30% 产品收入增速 (事后看远低于实际 36% 指引)
Benchmark reiterates Buy and raises PT to $200 from $190, arguing Snowflake enters F1Q27 with healthy momentum across cloud, data, and AI workloads. The firm expects Snowflake to comfortably exceed consensus on product revenue and margins, supported by strong hyperscaler/cloud consumption trends and accelerating adoption of Cortex Code, Cortex Analyst, and Observe. Benchmark also highlights constructive reads from DDOG/JFrog/Nebius/DOCN checks and believes Snowflake remains well positioned to sustain ~30% product revenue growth through FY27 as AI workload adoption broadens.
来源背景：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。 - 反对 KeyBanc 调查显示前瞻支出意愿明显恶化，F2Q 指引风险约 24% vs 街 26%
KeyBanc says recent SNOW surveys were mixed ahead of earnings, with April consumption trends remaining healthy/in-line while forward spending intentions deteriorated meaningfully. The firm flagged weaker NTM partner growth expectations, potential F2Q guide risk (~24% vs Street 26%), and softer-than-expected Cortex Code adoption data versus management commentary. However, KeyB also notes Snowflake continues gaining strategic relevance versus Databricks among many enterprise customers, while core consumption trends into F1Q remain constructive, maintaining OW and $200 PT.
来源背景：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。 - 支持 Funda 预测本季增速约 31% (超指引 3%)，CoCo 2Q26 带来约 2% 增量；2Q26 基数高 (2Q25 有一次性迁移)
**2.** Given 2Q25 had a one-off migration tailwind, 2Q26 faces a tougher base; however, with Cortex Code contribution offsetting part of the high-base headwind, we still expect ~30% growth.
来源背景：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 - 口径不齐 Q1 上调全年指引的最大驱动是 CoCo (而非核心)，因 CoCo 2 月 5 日才 GA，本季才首次把观察到的 CoCo 行为纳入模型——意味着全年上调中 AI 产品占比高
Thanks, Sridhar. I'll unpack that a little on terms of impact. And so, CoCo had the largest driver to the increase in our forecast. As a reminder, when we forecast we only forecast observed behavior. And as Sridhar mentioned, that just happened in the quarter. And so this quarter we had a very unique opportunity to layer CoCo in the model, and that's reflected throughout the remainder of the year. We also saw acceleration in our core business, and that informs our outlook as well. And so there's no change to our guidance philosophy where 3% we view as a really strong beat. Thanks, Brian, for the color.
- 中性 FY27 产品收入指引从 +27% 上调到 $5.84B/+31%；Q2 指引 $1.415–1.42B (+30%)，即 Q2 指引增速低于 Q1 实际 34%
There are no changes to our forecast methodology or guidance philosophy. Given the strength we've observed in both our core data platform business and AI business, we are raising our guidance for the year. For FY 2027, we now expect product revenue of $5.84 billion, representing 31% year-over-year growth. In Q2, we expect product revenue between $1.415 billion and $1.42 billion, representing 30% year-over-year growth. Our Observe acquisition is progressing well, consistent with our initial expectations. Observe contributed less than 1 percentage point of product revenue growth in Q1, and we continue to expect the acquisition to add approximately 1 percentage point of revenue growth – product revenue growth for the full year.
- 中性 指引方法不变，3% beat 被视为'很强的 beat'，只预测已观察到的消费行为
Amazing quarter. Thank you. Let me emphasize that there's been no change in guidance philosophy, and we view a 3% beat as a very solid beat. The difference that happened this quarter was CoCo was launched in the quarter, and we base guidance on observed behavior, and so we didn't have any observed behavior for guidance for CoCo. And so we had a unique opportunity now, since we've been able to watch it for a quarter, to layer that in now for the full year, and that's what we've done.
- 中性 RPO 同比 +38% (去年同期 +34%)，续约越来越集中在 Q4
We now have 79 customers spending more than $1 million on a trailing 12-month basis. 46 customers crossed the $1 million threshold in Q1 compared to 26 a year ago period. Remaining performance obligations grew 38% year-over-year, compared to 34% in Q1 of last year. We continue to see customers favor Q4 renewals. As a result, we expect bookings to be increasingly weighted towards the fourth quarter. We remain committed to delivering both growth and margin expansion.
- 支持 2Q 指引隐含 30% 增速，按典型 3% beat 推算约 34%；全年指引从 27% 上调到 31%
* 1Q26 product revenue grew 34% YoY, beating by 5.3%—the largest beat in recent history. Even excluding the Observe acquisition, the beat was still 4%. * 2Q26 revenue guidance implies 30% YoY growth. Applying the typical 3% beat, this implies 34% YoY growth—meaning growth is still not decelerating. * FY26 revenue guidance was raised from 27% to 31%, one of the largest full-year revenue raises in Snowflake’s history. Based on this guide, Snowflake still believes growth will not decelerate for the full year. * 2Q26 OPM guidance of 12.5% also came in 1% above expectations. Margins were not impacted by Cortex Code.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 Funda 认为 DDOG 和 SNOW 是当下最好的 AI SaaS 标的
* Datadog and Snowflake are clearly the best plays right now, in our view. * In our Snowflake Preview, we specifically wrote that Snowflake is now the company that most resembles a Datadog-style print. * Datadog’s AI story is that all of its revenue is linearly correlated with the growth of the AI TAM, with the non-AI business benefiting from AI customers. * Snowflake’s AI story is that Snowflake now has the most explosive coding product in AI, and is in the middle of a rapid ramp from a small base. (And 2–3% is no longer that small a base.)
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 Funda 在模型中假设 CoCo 单独贡献约 2% 增量收入，管理层几乎整场电话会都在讲 CoCo
This week of earnings belonged to data and SaaS. The standout was Snowflake — the Cortex Code ramp we had been flagging for three straight quarters finally showed up in the numbers. We had dedicated almost our entire preview to it, built into our model that Cortex Code alone could add roughly 2% of incremental revenue, and called it the fastest-ramping product in Snowflake’s history; management then spent nearly the whole call on it. On the other hand, in our MDB preview, we highlighted that investors may be too optimistic about the consumption tailwind from Anthropic, as well as about Enterprise Advanced becoming increasingly important, which the market might not appreciate. Both were validated in the earnings and caused significant volatility post-earnings.
来源背景：作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。 - 支持 产品收入 +34%、全年指引 27%→31%，SNOW 与 DDOG 是 Funda 最喜欢的两个 AI SaaS
**SNOW 1Q26 — the Cortex Code call lands.** Product revenue grew 34% YoY, the largest beat in recent memory, and the full-year guide was lifted from 27% to 31%; the print validated the Cortex Code thesis we had been building for three quarters. We continue to see Snowflake, alongside Datadog, as one of our two favorite AI SaaS names.
来源背景：作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。 - 中性 TAM 今天约 $225B，5 年后 >$460B (公司口径，不是收入目标)
There are some of you that are relatively new. So I'll walk in and talk about the market, some of the revenue drivers, get into GAAP profitability, capital allocation and so forth. So the market today is roughly $225 billion. We expect that market over the next five years to more than 2x over $460 billion. AI is expanding our market opportunity. Last year, we went through this.
- 口径不齐 被追问为何只有 30% 出头增速 (而主要竞争对手快一倍)，CEO 拒绝给更高目标，只说按观察到的行为指引
So, when you think about ultimately what you think these new solutions can do to help accelerate growth, I know you're not giving guidance here but it doesn't feel like you're at cruising altitude from where the rest of the industry is out right now. I don't know what to say. Absolutely, we aspire for more but showing is the new doing. Right. Our guidance is based on rooted observed behavior and we did considerably guide up this last quarter. As we see things happen, that's when we'll update our guidance. And having said that, I can't resist the ask for a GAAP profitability guidance from said person you mentioned. I'll go right over here.
- 支持 Benchmark 目标价 $270→$290，认为 SNOW 正成为企业 agentic AI 控制面
Benchmark raised its PT on SNOW to $290 from $270 and reiterated Buy following Snowflake Summit and Investor Day, arguing the company is increasingly positioning itself as the agentic AI control plane for the enterprise. The firm highlighted strong traction for CoCo and CoWork, noting AI products are accelerating migrations, increasing consumption, and helping customers move from experimentation into production faster. Benchmark also pointed to improving operating metrics, including faster sales cycles, expanding use cases, stronger customer acquisition, and a reduction in time-to-value from months to weeks. The analyst believes Snowflake is successfully shifting from a data platform to a broader AI workflow and orchestration platform, supporting sustained growth and a path to GAAP profitability by FY28.
来源背景：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 - 支持 Jefferies 称 SNOW 相对 DDOG 估值仍有明显折价
Jefferies came away from Snowflake Summit and Investor Day with increased confidence in Snowflake’s AI monetization story, highlighting rapid adoption of CoCo and improving GTM execution. The firm’s checks suggest CoCo has become materially more reliable since April and is outperforming third-party coding agents on Snowflake-native workflows, particularly text-to-SQL and governed enterprise data use cases. Jefferies argues context and governance are becoming key differentiators in enterprise AI, positioning Snowflake favorably as more agents move into production. The firm also sees improving sales productivity, faster migrations, and a clearer path to GAAP profitability by FY28, while noting Snowflake still trades at a meaningful valuation discount to Datadog despite growing AI momentum.
来源背景：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 - 口径不齐 专家称'上次分析师电话会给出未来 2 季度 30-35% 指引'——公司并无此表述，疑似误记
回 答 者：未来趋势无法预测，我们也没有从 Snowflake 获得相关信息。目前其环比增长指引为30%、31%至32%，这也是过去2个季度的趋势。由于我们仅讨论公开信息，Snowflake并未提供更远期季度的指引。但根据上一次分析师电话会，未来2个季度的指引区间为30%至35%。
来源背景：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 - 中性 Q3 预期 33-34%，是加速但不是曲棍球杆式
回 答 者：是的。不过，Q2仍然实现了小幅加速，而且是在更高基数的基础上实现的，因此我们仍然非常满意。对于Q3，我们不会进行这种季度预测，但预计增速至少会与此前持平或略高，可能在33%至35%之间，更具体而言，约为33%至34%。总体而言，增速确实在加快，但并非曲棍球杆式的大幅加速。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 支持 明年增速 30-45%，理由是数据已在 Snowflake 上且有治理，便于跑 agent
回 答 者：我认为会。Snowflake目前增长较快，而且规模已经相当大，因此判断其能否延续当前增长趋势并不容易。不过，我认为它明年可能会实现30%至45%的增长，这是一个概括性的区间。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 - 反对 下半年/明年增速会加快，但只是 1-2 个百分点
回 答 者：公司在Snowflake上的投入肯定会继续增长，这一点毫无疑问。增速方面，目前整体意向是继续加快，但具体数据还没有出来。根据个人观察，增速也会有所加快，至于是加快1个百分点还是2个百分点，目前很难判断。
来源背景：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 - 中性 年底前维持约 25%，明年希望翻倍到 40-45% 甚至 50%
我们希望明年的消费增速能够翻倍，达到约40%至45%，甚至可能达到50%。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 混合模型模式下预计明年客户用量增长 50-60%
回 答 者：我们的增长预期主要面向明年。预计明年客户用量将增长约50%至60%，这将显著推动整体用量增长。
来源背景：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 - 中性 Q3 消费增速可达 33-36%，下半年约 35%
A:第三季度会很强劲，我们可能会看到相较于我们对第二季度的指引一些潜在的再加速。我的预期是消费增长可以达到 33% 到 36%，而对于半年而言，我认为我们会维持在大约 35% 左右。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 Q1 超预期部分来自年初指引保守
A:这是一个合理的问题。根据我的经验，关键在于我们把来自产品的收入与底层消费势头混在了一起。第一季度异常强劲是因为产品收入。我们在第一季度末设定的指引相较于年初设定的要好一些，你可以说年初的指引有点保守。这可能是导致超出我们预期的一个原因。
来源背景：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 - 中性 企业 AI 试验成本已到董事会层面，有公司 1 个月用完全年预算，预算转向可控项目
目前情况正在发生变化。企业开始认识到，它们在测试阶段投入了相当多的资金。此前的方式是允许所有人自由使用人工智能，探索能够实现哪些成果，之后再考虑如何治理。如今，这些成本已经引起董事会层面的关注。例如，有些公司在1个月内就用完了全年的预算。
来源背景：Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 - 中性 RPO 加速发生在消费显现之前，是信心的领先指标，但若 AI 部署慢于计划，消费会滞后于承诺
关键的细微差别是，RPO 的加速发生在消费完全显现之前。客户实际上是在为预期的 AI 工作负载预留容量，因此承诺的规模可以比近期收入增长得更快。这通常是信心的有用领先指标，但也会带来执行风险。如果 AI 采用或部署比计划更慢，已确认的消费可能会滞后于合同承诺。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 口径不齐 专家称 Snowflake RPO 增速从 Q4 的 40%+ 降到 Q1 的 30%+，Q2 中高 30%
当时可得到的数据是 Q1 的数据，因为 Q2 的报告流程还在进行中。与 Q4 相比，RPO 的增长率已从 Q4 的超过 40% 下降到 Q1 的超过 30%。预计在本财政年度将保持在 30% 以上。这些数字指的是 Snowflake。MongoDB 的 RPO 增幅则要高得多。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 支持 2027 年会出现生产端变革的第二波
二阶效应对 Snowflake 和 MongoDB 非常重要，因为生产级 AI 的渗透速度可能远快于 AI 项目数量的增长。一个 POC 可能只访问一份数据集、供 100 名用户使用；其生产版本可能会连接到 ERP、CRM、文档和运营系统，服务 10,000 名用户、持续运行，并引入即便无人触发也会产生机器对机器活动的代理。我认为有意义的第二波将于 2027 年开始显现：第一波是试验阶段，而第二波将是生产端的变革。
来源背景：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 - 中性 渠道观察到的增长仍温和，与财报后股价积极反应不完全一致；超预期主要在前瞻指标
从本季度的调研结果来看，包括对Snowflake和MangoDV的渠道调研，渠道观察到的增长仍然较为温和，这与财报发布后股价的积极反应似乎并不完全一致。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 反对 超预期主要体现在 ARR 等前瞻指标，实际收入增长没那么快，股价更多是估值修复
不过，仔细分析这些公司的业绩，其超预期部分主要体现在前瞻性指标上，例如净增长的AR及其他净增长前瞻指标，实际收入增长并没有那么快。因此，我认为这个季度股价的表现更多反映了估值修复。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 中性 Merit 预计 Q2 产品收入 +35-36%，超指引 5-6 点；买方部分更乐观
从整体调研反馈和加速趋势来看，我们对Snowflake 2季度的趋势仍较为乐观。我们认为其同比可能实现35到4、35到36的增长。公司的guidance基本上略高于30%，我们认为其表现可能会beat公司指引约5%、5%到6%。
来源背景：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 - 支持 Funda 基准预测 FY27Q2 增速 35%，超指引 4%；3Q 加速到 36%；落在或略低于 bogey 也视为正面
1. Our base case is 35% growth, a 4% beat versus guidance, modestly above a typical beat, and 1 percentage point faster than 1Q26. 2. Given a one-off migration benefit in 2Q25, 2Q26 faces a difficult comparison. We believe continued acceleration in the core data warehouse business is challenging in 2Q26, but incremental contributions from Cortex Code, Snowpark, and other products can still drive acceleration in total revenue growth. 3. 3Q26 will no longer face a similar high-base headwind, so we expect further acceleration to 36%.
来源背景：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 - 反对 RPO 同比增速从 Q1 的 38% 放缓到 30% ($9B)；不过 54% 在未来 12 个月确认，cRPO 估算同比 +42%
In Q2, 48 net new customers surpassed $1 million in trailing 12-month spend. We now have 828 customers spending above the $1 million threshold. Remaining performance obligations grew 30% year-over-year, totaling $9 billion. As a reminder, we continue to see customers favor Q4 renewals. As a result, we expect bookings to be increasingly weighted towards the fourth quarter. Of the $9 billion RPO, we expect approximately 54% to be recognized as revenue in the next 12 months. This represents approximately 42% year-over-year growth compared to our estimate in the same quarter last year. Our Q2 results reinforce our commitment to delivering both growth and margin expansion. In Q2, non-GAAP operating margin expanded over 400 basis points year-over-year to reach 15%.
- 支持 FY27 指引上调到 $6.07B/+36% (含约 1pt Observe)；Q3 指引 $1.588–1.593B，+37–38%，即指引继续加速
There are no changes to our forecast methodology or our guidance philosophy. Given the strength we've observed, both in our core data platform business and AI business, we're raising our product revenue guidance for the year. For FY 2027, we now expect product revenue of $6.07 billion, representing 36% year-over-year growth. This includes approximately 1 percentage point of growth from Observe consistent with our previous outlook. In Q3, we expect product revenue between $1.588 billion and $1.593 billion, representing 37% to 38% year-over-year growth. Turning to margins. For FY 2027, we now expect 74% non-GAAP product gross margin.
- 中性 本季全年上调在核心与 AI 之间更均衡 (Q1 主要是 CoCo)，CFO 说 AI 产品采用者在核心上也增加消费
Thanks for the question. We base our guidance based on observed behavior up until the call that we have. And what we saw is that we talked about sort of CoCo, CoWork and all the AI functions driving additional business, but as well as the people who adopt them, they're also increasing business within the core. So, it's a reflection of the strength that we're seeing in our AI products as well as the underlying strength that we're seeing in the core. Thank you. Thank you.
- 支持 SNOW 盘后 +20%，在软件里最拥挤的名字上清除了高门槛
**SNOW +20%: Another 5% product-revenue beat plus an accelerating guide clears high bar for the most crowded name in sw**
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 - 支持 按 Q3 指引，Q3 很可能达到 40%+，比多头预期早一个季度；管理层把远超当季 beat 的幅度加入全年指引
Product revenue grew 37%, ahead of both Street and buyside expectations, while the Q3 guide calls for further acceleration and FY27 growth was raised five points to 36%. With this q3 guide, you now very likely get 40%+ growth in q3 a quarter earlier than even bulls expected. This is an unusually clean print: core demand remains strong, AI revenue stepped up meaningfully, and management passed substantially more than the quarterly upside into the annual outlook. Bulls in full control here. This explaisn the CoCO + core magnitude of the beat well (from our [weekly](https://substack.com/redirect/1aa17942-4a71-4deb-beb9-3d8d65ad436f?j=eyJ1IjoiNDR3MndrIn0.NBlbpuapVvjXEQ-22Fcu951Q9Y3oMbVG8TDoGPfg2x8) on Sunday):
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 - 支持 Q2 产品收入 $1.49B (+37%) vs 街 $1.42B (+31%)，买方 bogey 36%
* **F2Q product revenue:** $1.49B, +37% y/y vs. $1.42B, +31% and bogeys at 36% * **F3Q product-revenue guide:** $1.591B, +37.5% vs. $1.50B, +29% and bogeys of 32-33% * **FY27 product-revenue growth:** 36% vs. 31% prior
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 - 支持 Q3 指引 $1.591B (+37.5%) vs 街 $1.50B (+29%)，买方 bogey 32-33%
* **F2Q product revenue:** $1.49B, +37% y/y vs. $1.42B, +31% and bogeys at 36% * **F3Q product-revenue guide:** $1.591B, +37.5% vs. $1.50B, +29% and bogeys of 32-33% * **FY27 product-revenue growth:** 36% vs. 31% prior
来源背景：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 - 支持 FY27Q2 产品收入 +37%，超卖方一致预期 5.2%，但只超买方预期 1%
* FY27Q2 product revenue grew 37% year over year, beating sellside consensus by 5.2%. That matched the prior quarter’s beat, and revenue also came in 1% above buyside expectations. * FY27Q3 product revenue guidance implies 37% to 38% year-over-year growth. Applying a typical 3% beat would imply roughly 42% growth. Even after adjusting for the tough FY27Q2 comparison, FY27Q3 still represents an acceleration. * Full-year product revenue growth guidance was raised from 31% to 36%, one of the largest annual guidance increases in Snowflake’s history. * FY27Q3 non-GAAP operating margin guidance is 15.5%, also above expectations. Product gross margin is modestly lower because of CoCo, but companywide margin expansion remains clear, supported by faster growth and AI-driven organizational efficiency.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 Q3 指引隐含 37-38%，加典型 3% beat 约 42%；全年指引从 31% 上调至 36%
* FY27Q2 product revenue grew 37% year over year, beating sellside consensus by 5.2%. That matched the prior quarter’s beat, and revenue also came in 1% above buyside expectations. * FY27Q3 product revenue guidance implies 37% to 38% year-over-year growth. Applying a typical 3% beat would imply roughly 42% growth. Even after adjusting for the tough FY27Q2 comparison, FY27Q3 still represents an acceleration. * Full-year product revenue growth guidance was raised from 31% to 36%, one of the largest annual guidance increases in Snowflake’s history. * FY27Q3 non-GAAP operating margin guidance is 15.5%, also above expectations. Product gross margin is modestly lower because of CoCo, but companywide margin expansion remains clear, supported by faster growth and AI-driven organizational efficiency.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 全年产品收入增速指引从 31% 上调到 36%
* FY27Q2 product revenue grew 37% year over year, beating sellside consensus by 5.2%. That matched the prior quarter’s beat, and revenue also came in 1% above buyside expectations. * FY27Q3 product revenue guidance implies 37% to 38% year-over-year growth. Applying a typical 3% beat would imply roughly 42% growth. Even after adjusting for the tough FY27Q2 comparison, FY27Q3 still represents an acceleration. * Full-year product revenue growth guidance was raised from 31% to 36%, one of the largest annual guidance increases in Snowflake’s history. * FY27Q3 non-GAAP operating margin guidance is 15.5%, also above expectations. Product gross margin is modestly lower because of CoCo, but companywide margin expansion remains clear, supported by faster growth and AI-driven organizational efficiency.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 Funda 模型：CoCo 五个季度从 $30k 到 $59.4m，数仓占产品收入从 96% 降到 83%，产品增速从 26% 低点回升到 42%；净收入 $1.21b 到 $2.39b，OPM 10.8% 到 22.4%
[](https://substack.com/redirect/09acd712-9eb4-46e0-b895-ea4a746cc369?j=eyJ1IjoiNDR3MndrIn0.NBlbpuapVvjXEQ-22Fcu951Q9Y3oMbVG8TDoGPfg2x8)
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 Funda 认为 DDOG 面临大客户逆风、MDB AI 收入基数小，SNOW 叙事最清晰
* After this quarter, the AI infrastructure landscape is clearer. DDOG faces large-customer headwinds, while MDB’s AI revenue base remains small. * By contrast, SNOW’s investment narrative is clearer after the quarter. It is no longer a story about revenue contribution from a single product, and investors can now take a longer-term view of the company.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 口径不齐 Funda 认为软件近期空头回补基本完成，跑赢已显拉伸 (仓位提示)
The software vs semi split that defined last week finally started to rebalance. Software that printed what the channel had been pointing to kept working through most of the week, while the semi tape spent the first half getting tagged by rates — the 10-year pushed to 4.8% on Tuesday, the highest since January 2025, after Warsh’s hawkish Jackson Hole turn had traders pricing hikes rather than cuts. By Friday the setup flipped: semis clearly outperformed, and memory names broke out. Near-term software short cover looks largely done, the outperformance looks stretched, and the momentum unwind in semis has already gone to an extreme — which leaves less of that particular overhang in the names that got hit hardest.
来源背景：作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。 - 支持 Funda 认为估值应基于数据飞轮而非单一 SKU；客户在把交易型数据归集到平台
On the software prints, SNOW was the cleanest expression of the channel read. We went into the print cautious on 2Q because of tough comps and more constructive on 3Q, and the flywheel was stronger than even that. Product revenue grew 37% year over year, a 5.2% beat that matched last quarter, and the full-year product-revenue guide jumped from 31% to 36%, one of the largest annual raises in the company’s history. CoCo is no longer a product call. Accounts passed 9,100 with more than 2,000 net adds in the quarter, CoWork is at 5,800, and customers are consolidating transactional data onto the platform. Our preview had CoCo adding about two points this quarter; the print confirmed the direction and then some. The data flywheel is what we think the multiple should be underwriting, not a single SKU.
来源背景：作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。 - 支持 Funda 前瞻预计 35% 增速 (超指引 4 个点)，实际 37%
**Preview & Review | SNOW FY27Q2: Beyond a Single Product, a Data Flywheel That Merits a Longer Horizon.** Product revenue grew 37% YoY and beat sell-side by 5.2%, with the full-year product-revenue guide raised from 31% to 36%. Our preview had CoCo still very strong and called 35% growth, a 4-point beat versus guidance; the print confirmed the CoCo-to-flywheel read, with CoCo accounts above 9,100 and CoWork at 5,800.
来源背景：作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。 - 中性 新产品指引偏保守，不用一个月的数据外推；两个季度数据后更有信心
And so, when we do come up with guidance for the core platform like the migrations is based on observed behavior. And we have years and years of data with that and can get pretty close for the new products. We try to be a bit conservative. So we don't take a month worth of data and extract that out and say, this is what it's going to be like all for this year and next year. But now that we've had two quarters of data, we feel more confident in what we can extract from that observed behavior. When you and I first met, we started talking about the customer cohorts and how, look, it sometimes takes year one is the initial ramp and then year two is really when a customer gets into full swing Snowflake adoption. Has anything changed as you look at the speed at which these customers are ramping? I think from a cohort perspective, if you look at all the verticals across the company, we still have the same sort of vertical penetration, if you will, in financial services, manufacturing, government and so forth. As far as ramping, with the use of CoCo and AI, we're seeing customers ramp much quicker than what they've ramped historically. And so whether it's our partner network or what we're doing internally, what Sridhar talked about is outcome-based pricing, has really built the credibility with our customers. If someone comes to me and says, I can guarantee you X for this set price, and I knew historically that took a lot of time and a lot of materials and so forth, like I'm all in, and we're seeing that from a customer perspective.
- 支持 CoCo 首季没纳入指引，第二季开始纳入但打了折扣
And so, we have all the products over the course of Snowflake, we have all the curves and the trajectory of what those products have done. And then we apply the pattern recognition behaviors to those based on what we're observing. And so, with CoCo only being out two quarters, the first quarter that we did it, we didn't include it because we didn't have any of their behavior. Now we're able to, the next where we're able to but we don't want to put in the initial sort of small numbers, but really, really high growth percentages. And so, we took a haircut to that. But now that we're two quarters in, we have more data. We feel like we have a better accuracy and we're guessing it.
- 支持 核心业务每天夜里按客户、产品线预测，精度很高；新产品基本不计入
We have -- there's an awesome team back at Snowflake, the team's been there roughly about eight years. They are a bunch of data scientists and they forecast by customer and by product line literally everything at night. So, the pipelines run for 4 hours at 5:00 AM every morning. I get extremely detailed forecasts of what's happening within the business. And so, with the core, we have tons of data. We've done that for a while. Our degree of accuracy is very high there. And with the -- with AI and some of the new products, we don't have as much observed data. And so, we're going off of what we know sort of at the time. That same team as well as another team builds detailed product launch models for us. And so, we have all the products over the course of Snowflake, we have all the curves and the trajectory of what those products have done. And then we apply the pattern recognition behaviors to those based on what we're observing. And so, with CoCo only being out two quarters, the first quarter that we did it, we didn't include it because we didn't have any of their behavior. Now we're able to, the next where we're able to but we don't want to put in the initial sort of small numbers, but really, really high growth percentages. And so, we took a haircut to that. But now that we're two quarters in, we have more data. We feel like we have a better accuracy and we're guessing it. Then we have a whole bunch of new products that obviously are new. So, we don't have a lot of observed data and so we don't put a lot behind those right now. Yeah. And on those new products, I mean, what are kind of the top two or three new products that are now out there that you're most excited about or customers are seeing kind of initial traction with? Absolutely. CoCo's sibling is called CoWork. And so CoWork has over 5,000 accounts. CoWork is what I'll call data with rails around.
- 支持 FY27 产品收入增速指引从 31% 上调到 36%
And so, growth is extremely important. You've seen our guidance for the full year. We raised it from 31% year-over-year growth to 36% year-over-year growth. And so that's a pretty significant step-up off of a prior quarter where we had another big increase. But we've done that with increasing operating leverage in the business. And so, what we are very committed to is increasing operating leverage in the business.
- 支持 CoCo 是公司史上最成功的产品发布之一，增长远快于以往；短期高增速数据打折后纳入指引
So, we have years and years of launching new products, and we obviously save all that data and go – can go back and look at the pattern recognition of what's happened early on. And then we actually try to map that against what's happening. CoCo, being one of the most successful product launches in the company's history, has grown way faster than a number of those other products. And so, when you have a very short period of data, and so when we always talk about sort of our guidance, one, we say that we haven't changed our guidance philosophy or our forecast methodology. You go back and you look at the observed behavior. And with just a short amount of data with sort of high numbers, I think it would somewhat be reckless to extract that out and just apply that to guidance. And so we take that number and apply a haircut to that, and then that's sort of what comes up with the number. In first quarter, obviously, we had about a quarter's worth of data. Now we have two quarters' worth of data, so we're getting closer to what that number is.
- 中性 分析师：市场习惯 3% beat，最近是 5%+ beat
It's role-based access, has security, has governance. And so, it just makes it a different way of working than how I worked the previous 30 years. And relative to Snowflake and the performance, which has been very strong in the last couple quarters, the Street is all hung up on this 3% mark, which I think your model kind of... What's 3%? ...tunes to. You've probably never heard of before. Coming off to 5%-plus beat quarters. Right. What's really driving this underlying inflection in the business? And how durable is this growth rate? And how come, as you come in as CFO, your predictability has gotten worse because you're beating by more? Thank you, Rob. Yes.
- 支持 FY27 指引一个季度上调 500bp，从 31% 到 36%
But on this earnings call, I talked about we just – quarter or so when we reported the call, the number of heads that we've hired this year compared to the number of heads we hired last year really shows this. So, we're re-accelerating revenue. We increased our full year guide from 31% year-over-year to 36% year-over-year. So, increased it by 500 basis points in just one quarter. And at the same time, year-to-date this year, we've added roughly about 330 people, of which 170 to 180 of them came from Observe acquisition. So, 150 compared to the same period last year, there's roughly about 940 people.
- 中性 TAM 说法变成 5-6 年约 $300B (不含分析和 BI)
And then there's a lot of sort of solutions today that aren't AI-pilled. And so, AI is really going to help the ability to access those workflows, understand your data, get more people to access the data, and so forth. And so, one of the things I talked about in Investor Day, we're in a market that is roughly going to $300 billion in the next five or six years. And that doesn't include all the analytics and business intel, but does include all the other stuff around it that you can do with CoCo and CoWork and so forth. And so that's what's super exciting for us at Snowflake.
- 中性 多数合同在第四季度续约，CFO 称无担忧
It aligns with the budget season when they're budgeting for next year. And so, we've signed seasonality-wise. We've got the majority of our contracts in fourth quarter, and that's happening. It's – when you're a consumption business and you're a rep and that's your livelihood, you're talking to those customers all the time, right? Talking to them about how to use AI, how to use our AI gateway, our AI functions. CoWork, CoCo, all of our different AI stuff, but what migrations you can actually bring in, savings that you can get on migrating off legacy software, BI and analytical tools that you can replace. And so our reps are in with those customers literally on a weekly basis. And so, the timing of a renewal is just – it's not like a major event. It's just a normal event for us. And they're already working on that as we get up to that renewal. So, I don't have any concerns with that. Okay. So, we're going to have a breakout across the hall with Snowflake.
E07 股东经济学
判断：差，但这是 mapping 里所有赢家里改善最明确的一家：管理层第一次给了 GAAP 盈利的日期和 SBC 的下降斜率，而且用人头控制在兑现。
现状：过去 12 个月 (FY26 Q3–FY27 Q2)：总收入 $54.3 亿，FCF $11.95 亿 (22%)，SBC $16.42 亿 (30%)，FCF − SBC 约 -$4.5 亿 (-8%)。非 GAAP 稀释股数指引从 376M 上调到 380M，Q3 382M。可转债 $23 亿 (2027、2029 年各一半)，转股价约 $157.50，capped call 只对冲到 $225，按现价两期都是价内，名义转股约 1,460 万股。回购授权 $45 亿只剩约 $8 亿，Q2 没有回购。
改善：Investor Day 宣布 FY28 Q4 实现 GAAP 盈利，CFO 说这只依赖 opex 和 SBC，不依赖收入假设；SBC 占收入 41% (FY25) → 34% (FY26) → 27% (FY27E)，建议"按同样斜率建模"。非 GAAP 经营利润率 FY25 6.4% → FY27 指引 14.5%。人头：Q1 剔除 Observe 只净增 17 人；上半年净增 334 人 (173 来自 Observe)，去年同期 935 人；工程招聘改为 1–2 年经验的"AI 原生"年轻人，替代需要大额期权的资深人员。
问题：6 月股东大会上高管薪酬的咨询性投票 (say-on-pay) 没有通过，在大型科技公司里很少见。没有给长期收入、FCF 或稀释目标。按 SBC 斜率，FY28 SBC 约收入的 20%，FCF 率 23–25%，FCF − SBC 才转正到约 3–5%。分析师在 Piper 会议上问：主要对手 (Databricks) 在 FCF 盈亏平衡运行，为什么此时追求 GAAP 盈利？CFO 说没有约束投入。
All right. Let's talk about GAAP profitability. We're super excited to announce this today and the leverage that we're getting in the model while seeing the revenue growth that we have. So, we announced that will be GAAP profitable in 4Q FY 2028. There's three levers to do this, revenue, operating expense and SBC. We actually just played with the bottom two, operating expense and SBC. This is not a discussion about FY 2028 revenue. And so, we're seeing greater operating efficiency and operating expense. For modeling purposes to help you out, assume the same trend in SBC that you've seen for the following few years. And so, we are 41% of revenue, last year we are at 34% of revenue.
We actually just played with the bottom two, operating expense and SBC. This is not a discussion about FY 2028 revenue. And so, we're seeing greater operating efficiency and operating expense. For modeling purposes to help you out, assume the same trend in SBC that you've seen for the following few years. And so, we are 41% of revenue, last year we are at 34% of revenue. And this year, we said we'd be 27% of revenue. So that should help you from a model perspective on how we're going to reach GAAP profitability in 4Q 2028. With that as well is we don't expect to do any large M&A. Okay.
For FY 2027, we now expect 74% non-GAAP product gross margin. This revised outlook includes a higher revenue mix from fast growing AI workloads, which carry a lower contribution margin today. We're delivering continued operating margin expansion as we offset growing cloud costs with slowing head count expense. We are increasing our FY 2027 non-GAAP operating margin guidance from 13.5% to 14.5%. For Q3, we expect non-GAAP operating margin of 15.5%. We're reiterating our full year non-GAAP adjusted free cash flow margin guide of 23%. I'd like to close with my two key goals for the year. First, help the business to deliver growth and margin expansion. Second, support ongoing excellence in our go-to-market motion.
Our Q2 results reinforce our commitment to delivering both growth and margin expansion. In Q2, non-GAAP operating margin expanded over 400 basis points year-over-year to reach 15%. Our outperformance was driven by strong revenue growth and disciplined head count management. Year-to-date, we've added 334 employees, which includes 173 from our Observe acquisition. This compares to 935 added in the year-ago period. We ended the quarter with $4.3 billion in cash, cash equivalents, short-term and long-term investments. Moving to our outlook. As always, our forecast is based on observed consumption patterns.
And so, what we're doing is we're -- and one of the things that we're doing to get profitable was a big component of that with stock-based compensation. And what we're doing is we're hiring a different type of person that can actually be an engineer versus what we would historically hire. So, historically, you'd have to hire someone with eight to 10 years of experience, really senior, maybe had a management role in the past. And now you can hire someone one or two years out of college who really is AI build, that is vibe coded that use CoCo, and so forth. And so, we're seeing a dramatic difference in the two of them on the productivity. If you go into the finance organization, we have probably 150 different use cases. I say probably because they're developing them every day.
And so, we are 41% of revenue, last year we are at 34% of revenue. And this year, we said we'd be 27% of revenue. So that should help you from a model perspective on how we're going to reach GAAP profitability in 4Q 2028. With that as well is we don't expect to do any large M&A. Okay. We'll jump into capital allocation. Three primary areas. One is organic growth, R&D and sales and marketing. You heard Christian talk a lot about from a product perspective what we're releasing to the marketplace and the velocity that Sridhar talked about in his presentation. We'll continue to do that. From a sales and marketing perspective, it's really adding capacity to capacity model where needed. We have about $800 million left of authorization in our $4.5 billion that was announced earlier. And then from an M&A perspective, we have typically done small tuck-ins on a buy versus build more of an aqua hire perspective. So in conclusion, we have a very large and growing market. We have durable growth drivers with the land and expansion motion that we.
The proxies and ballots will be tabulated by the Inspector of the Election. Based upon preliminary information provided by the Inspector of Election, I can report that each of Teresa Briggs, Mark D. McLaughlin and Sridhar Ramaswamy is elected as a Class 3 director. The advisory proposal on the compensation of our named executive officers is not approved. The proposal to ratify the appointment of PwC as our independent registered public accounting firm for the fiscal year ending in January 31, 2027, is approved and the non-binding stockholder proposal requesting majority vote for the director elections is approved. We expect to report our final results within four business days after this meeting on a form 8-K. The Annual Meeting is now adjourned.
You've seen that acceleration. Why is this the right move at this time to push towards GAAP profitability? Sometimes... Because you've got a pretty big competitor out there that's running at free cash flow breakeven. Yeah. It's – I'll answer it at the reverse way in the sense, let's say, did I do anything that put constraints on the business to be able to make that comment? And the answer is, absolutely not. Like, we're spending as much as we want to spend in the business. We go in to these large enterprises, and our gross retention rate is really high. The core of our business really comes from customer six, seven, eight years ago because they got on Snowflake, they move more and more into Snowflake.
展开本节全部 33 条原文 (支持 21 / 反对 1 / 其他 11，按时间)
- 中性 Q3 回购 $2.33 亿/100 万股，均价 $223.35
Brian G. Robins: Our product revenue grew 29% Brian G. Robins: Together, these factors contributed to a stable net retention rate of 125%. Brian G. Robins: Financial services and technology verticals led growth in Q3. Brian G. Robins: Our Global 2000 customers now total 776 Brian G. Robins: each of these accounts spending, on average, $2.3 million on a trailing 12-month basis. Brian G. Robins: We achieved strong booking results, signing four nine-figure deals. Brian G. Robins: This represents a record number of large deals signed in a single quarter. Brian G. Robins: Q3 non-GAAP product gross margins was 75.9%. Brian G. Robins: Non-GAAP operating margin expanded more than 450 basis points year-over-year to 11% Brian G. Robins: Non-GAAP adjusted free cash flow margin was 11%. Brian G. Robins: In Q3, we used $233 million to repurchase 1 million shares at a weighted average price per share of $223.35. Brian G. Robins: We still have $1.3 billion remaining on our original authorization for $4.5 billion through March of 2027. Brian G. Robins: We ended the quarter with $4.4 billion in cash, cash equivalents Brian G. Robins: For Q4, we expect product revenue between $1.195 billion and $1.2 billion, representing a 27% year-over-year growth. Brian G. Robins: We expect non-GAAP operating margin of 7%. Brian G. Robins: We are raising our FY 2026 product revenue guidance. Brian G. Robins: We now expect product revenue of approximately $4.446 billion, representing 28% year-over-year growth. Brian G. Robins: We expect non-GAAP product gross margin of 75%, non-GAAP operating margin of 9%
- 支持 SBC 占收入 FY25 41% → FY26 34%
Sridhar Ramaswamy: As AI agents become central to how work gets done, those same capabilities become even more valuable Sridhar Ramaswamy: agents are only as powerful as the data they can access and the governance and security that surround it. Sridhar Ramaswamy: We've built the conditions that make agents safe, scalable, and enterprise-ready Sridhar Ramaswamy: Thanks to AI, we are both scaling revenue and becoming operationally more efficient. Sridhar Ramaswamy: Stock-based compensation declined meaningfully from 41% of revenue in fiscal 2025 to 34% in fiscal 2026, Sridhar Ramaswamy: and we expect it to further decrease to 27% of revenue in fiscal 2027. Sridhar Ramaswamy: Work is fundamentally changing and we are leading this transformation both within Snowflake and across the industry.
- 支持 FY27 SBC 预计降到收入的 27%
Sridhar Ramaswamy: As AI agents become central to how work gets done, those same capabilities become even more valuable Sridhar Ramaswamy: agents are only as powerful as the data they can access and the governance and security that surround it. Sridhar Ramaswamy: We've built the conditions that make agents safe, scalable, and enterprise-ready Sridhar Ramaswamy: Thanks to AI, we are both scaling revenue and becoming operationally more efficient. Sridhar Ramaswamy: Stock-based compensation declined meaningfully from 41% of revenue in fiscal 2025 to 34% in fiscal 2026, Sridhar Ramaswamy: and we expect it to further decrease to 27% of revenue in fiscal 2027. Sridhar Ramaswamy: Work is fundamentally changing and we are leading this transformation both within Snowflake and across the industry.
- 中性 FY27 调整后 FCF 率指引 23%，含约 150bp 收购逆风
Robins: We expect non-GAAP adjusted free cash flow margin of 23%. Brian G. Robins: This includes an approximate 150 basis point headwind related to our acquisition. Brian G. Robins: which we acquired for approximately $600 million in a combination of cash and stock.
- 中性 Q4 回购 $1.5 亿/约 66.8 万股，剩余授权 $11 亿
Brian G. Robins: In Q4, we used $150 million to repurchase approximately 668,000 shares Brian G. Robins: We have $1.1 billion remaining on our repurchase authorization Brian G. Robins: ended the quarter with $4.8 billion in cash, cash equivalents, short-term and long-term investments. Brian G.
- 支持 Q4 小规模裁员约 200 人，净增员工仅 37 人，增长与人数脱钩
* Margin触发了上季度的抛售，我们在Preview中特别提到我们不担心Margin的一个原因是观察到SNOW对Customer Success等部门进行了裁员，以及未观察到Memory涨价对Margin的影响 * “我想强调这一点。就在第四季度，我们借助AI获得了很多收益，我们进行了小规模的裁员，公司约200名员工受到了影响。因此，如果你看我们第四季度的净员工增加人数，我们只增加了37人。所以，AI真正改变了投资增长的框架，增长不再与员工人数挂钩。” * "我们的服务交付团队完成客户项目的速度提升了最多5倍，响应准确率提高了超过25%……我们构建了帮助销售人员优先处理客户、自动化研究和生成个性化触达的智能体能力，预计今年将节省相当于90名全职工程师的生产力。” * “我们推出新产品的目标首先是打造出色的产品，其次是让产品易于使用，再次是推动营收增长。达到这些目标之后，我们才会着眼于优化利润率。我们推出了大量新的AI产品，这些产品目前的利润率不如核心业务高，但我们正在通过提升核心业务的效率来抵消这一影响。我们会做出正确的决策来推动增长，并在营业利润率层面实现整体平衡。” * “利润率的改善将同时体现在毛利率层面和公司整体层面……因此，AI在毛利率线以上和营业利润率层面都有很多可以提升的空间。坦率地说，关键在于如何排列优先级，因为机遇实在太多了。”
来源背景：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 - 中性 专家唯一疑问是 Snowflake 是否在 AI 上 '过度投资'
FC: Right? They need to if they want to keep playing in this game. Snowflake, honestly, if anything, they're over- investing and over-indexing, whereas... Are they overspending would be my only question to Snowflake. But from my perspective as a partner with them, I love seeing it. I want to see them continue to make this spend because that allows us, since we're always on the cutting edge of Snowflake rollouts, we're implementing and oftentimes enterprises, our customers, they tend to fall behind and not utilize the latest tool sets and so this is where we come in with fresh blood and fresh eyes and retrain their team to leverage the latest from Snowflake which is exactly what Snowflake wants us to do for them because that infuses new revenue and new tools to their revenue stream. So yes, long answer, but yes. That's a key differentiator between Snowflake and MongoDB and how they're looking at their AI future.
来源背景：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 - 支持 Non-GAAP 营业利润率同比 +300bp 至 12%；全年指引 12.5%→13.5%，FCF 率 23% 维持 (含 Observe 约 150bp 拖累)
Turning to margins. We expect 75% non-GAAP product gross margin for FY 2027. We expect Q2 non-GAAP operating margin at 12.5%, and we are increasing our full-year non-GAAP operating margin guidance from 12.5% to 13.5%. We are reiterating our non-GAAP adjusted free cash flow margin guide of 23%. Our full-year outlook for both non-GAAP operating margin and non-GAAP adjusted free cash flow margin continues to include approximately 150 basis point headwind related to our Observe acquisition. This impact is unchanged from last quarter.
- 中性 Q1 用约 $300M 回购 1.7M 股，原 $4.5B 授权剩约 $800M；现金及投资 $4.4B。未讨论 SBC
Of these 190 employees, 173 joined Snowflake through the Observe acquisition. Excluding Observe, organic hiring was limited to 17 people in the quarter. In Q1, we used approximately $300 million to repurchase 1.7 million shares. We have approximately $800 million remaining of our original $4.5 billion repurchase authorization. We ended the quarter with $4.4 billion in cash, cash equivalents, short-term, and long-term investments. During the quarter, we entered into a five-year, $6 billion contract with AWS, more than doubling our prior contract signed in FY 2023.
- 支持 Q1 仅增 190 名员工，其中 173 名来自 Observe，有机新增仅 17 人——AI 提效压 headcount
Strong revenue growth and disciplined hiring both contributed to the outperformance in non-GAAP operating margin. We added 190 employees this quarter, compared to approximately 400 added in the year ago period. Of these 190 employees, 173 joined Snowflake through the Observe acquisition. Excluding Observe, organic hiring was limited to 17 people in the quarter. In Q1, we used approximately $300 million to repurchase 1.7 million shares. We have approximately $800 million remaining of our original $4.5 billion repurchase authorization.
- 支持 2Q OPM 指引 12.5%，高于预期约 1 个百分点，利润率未受 CoCo 影响
* 1Q26 product revenue grew 34% YoY, beating by 5.3%—the largest beat in recent history. Even excluding the Observe acquisition, the beat was still 4%. * 2Q26 revenue guidance implies 30% YoY growth. Applying the typical 3% beat, this implies 34% YoY growth—meaning growth is still not decelerating. * FY26 revenue guidance was raised from 27% to 31%, one of the largest full-year revenue raises in Snowflake’s history. Based on this guide, Snowflake still believes growth will not decelerate for the full year. * 2Q26 OPM guidance of 12.5% also came in 1% above expectations. Margins were not impacted by Cortex Code.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 剔除 Observe 本季只净增 17 名员工，效率提升
Excluding Observe, Snowflake hired only 17 people this quarter. This is consistent with last quarter—Snowflake is improving efficiency without needing to add more headcount. As we noted in our earliest Snowflake deep-dive in 3Q24, we believed the biggest impact of trusting the new CEO would be his serious commitment to driving developer productivity.
来源背景：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 - 支持 4Q FY28 实现 GAAP 盈利，只依赖 opex 和 SBC，不依赖收入假设
Let's talk about GAAP profitability. We're super excited to announce this today and the leverage that we're getting in the model while seeing the revenue growth that we have. So, we announced that will be GAAP profitable in 4Q FY 2028. There's three levers to do this, revenue, operating expense and SBC. We actually just played with the bottom two, operating expense and SBC. This is not a discussion about FY 2028 revenue. And so, we're seeing greater operating efficiency and operating expense. For modeling purposes to help you out, assume the same trend in SBC that you've seen for the following few years.
- 支持 SBC 占收入按 41%→34%→27% 的斜率继续下降，作为建模假设
This is not a discussion about FY 2028 revenue. And so, we're seeing greater operating efficiency and operating expense. For modeling purposes to help you out, assume the same trend in SBC that you've seen for the following few years. And so, we are 41% of revenue, last year we are at 34% of revenue. And this year, we said we'd be 27% of revenue. So that should help you from a model perspective on how we're going to reach GAAP profitability in 4Q 2028. With that as well is we don't expect to do any large M&A.
- 支持 CEO 把 GAAP 盈利归因于经营利润率提升和 SBC 受控
These are among the healthiest collaborations that we have had with a very motivated customer. They are doing similar things for large banks and we anticipate that we'll be leaning in to something like this as the impact of products like CoWork becomes obvious and people realize that their data teams like our own have to modernize themselves for them to be relevant in this age of AI. And the final comment that I want to make is that because we have invested so much in transforming ourselves, in being more effective as a company and because of our ability to increase non-GAAP operating margin but also bring SBC firmly under control, we feel confident enough to say that we'll be reaching GAAP profitability at the end of next year. And Brian will walk us through more of the details of that but I see this as the culmination of the work that we have done over the past three years to reinvent ourselves to be a more driven, more product focused, more quality obsessed, continuously self-improving company. With that, I will hand it off to Christian.
- 支持 Non-GAAP 经营利润率 FY25 6.4% → FY27 指引 13.5%，产品毛利率持平约 75%
Sridhar and I are 100% aligned. You can grow while getting operating leverage in the model. You can see in FY 2025, we delivered 6.4% non-GAAP operating margin, and on the last call we just guided to over 13 point – we guided to 13.5%, over 2x within the two years. And we're doing that by keeping non-GAAP product gross margin flat at roughly 75%. We talked about the AI products have a lower gross margin than our core. So how are we doing that? We're doing it really two ways. One is we're being extremely disciplined on head count. And so, last quarter, we reported absent of the observe acquisition, net head count increase only by 17, the quarter before that, only 37. And so, we're changing the way we work by using AI tools and necessarily and heads to get work done but transforming how we're working. On the other hand, cloud spend is going up a little.
- 支持 剔除 Observe 后最近两季净增员工仅 17 和 37 人，云成本小幅上升
So how are we doing that? We're doing it really two ways. One is we're being extremely disciplined on head count. And so, last quarter, we reported absent of the observe acquisition, net head count increase only by 17, the quarter before that, only 37. And so, we're changing the way we work by using AI tools and necessarily and heads to get work done but transforming how we're working. On the other hand, cloud spend is going up a little. The offset of these two is giving us operating leverage in the model. The billing payment terms for the company have been really consistent.
- 中性 不做大型并购；$4.5B 回购授权剩约 $800M；没有给 FCF 或 FY 收入长期目标
And this year, we said we'd be 27% of revenue. So that should help you from a model perspective on how we're going to reach GAAP profitability in 4Q 2028. With that as well is we don't expect to do any large M&A. Okay. We'll jump into capital allocation. Three primary areas. One is organic growth, R&D and sales and marketing. You heard Christian talk a lot about from a product perspective what we're releasing to the marketplace and the velocity that Sridhar talked about in his presentation. We'll continue to do that. From a sales and marketing perspective, it's really adding capacity to capacity model where needed. We have about $800 million left of authorization in our $4.5 billion that was announced earlier. And then from an M&A perspective, we have typically done small tuck-ins on a buy versus build more of an aqua hire perspective. So in conclusion, we have a very large and growing market.
- 中性 分析师质疑在增速加速期追求 GAAP 盈利；CEO 回答加人不等于多产出
Brian talked about the TAM here. Your business is accelerating. Why is GAAP profitability important at this juncture and why not pour more investment in now versus being governed by this promise? Because it isn't clear that simply throwing more humans at problems gets more things done. That's like my honest assessment. I think more of my energy should go into having each and every one of my engineers think and act like the person that like reliably got that feature out into ours.
- 支持 Benchmark 与 Jefferies 都提到 FY28 实现 GAAP 盈利的路径 (Investor Day 目标)
Benchmark raised its PT on SNOW to $290 from $270 and reiterated Buy following Snowflake Summit and Investor Day, arguing the company is increasingly positioning itself as the agentic AI control plane for the enterprise. The firm highlighted strong traction for CoCo and CoWork, noting AI products are accelerating migrations, increasing consumption, and helping customers move from experimentation into production faster. Benchmark also pointed to improving operating metrics, including faster sales cycles, expanding use cases, stronger customer acquisition, and a reduction in time-to-value from months to weeks. The analyst believes Snowflake is successfully shifting from a data platform to a broader AI workflow and orchestration platform, supporting sustained growth and a path to GAAP profitability by FY28.
来源背景：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 - 中性 登记日流通普通股 346,602,915 股
Laura has taken the customary oath of office and we will file this oath with the records of the meeting. Our bylaws provide that the presence in person or by proxy of a majority of the voting power of the outstanding shares of our common stock, entitled to vote at the meeting will constitute a quorum. There were 346,602,915 shares of our common stock outstanding and entitled to vote on the record to date, and Laura has informed me that proxies have been received for approximately 290,140,060 shares, or approximately 83.7% of the shares outstanding, which constitutes a quorum for today's meeting. Each share of common stock is entitled to one vote. There are four proposals to be considered by the stockholders at this meeting.
- 反对 高管薪酬咨询性投票 (say-on-pay) 未获通过——股东对薪酬 (含 SBC) 不满的信号
Based upon preliminary information provided by the Inspector of Election, I can report that each of Teresa Briggs, Mark D. McLaughlin and Sridhar Ramaswamy is elected as a Class 3 director. The advisory proposal on the compensation of our named executive officers is not approved. The proposal to ratify the appointment of PwC as our independent registered public accounting firm for the fiscal year ending in January 31, 2027, is approved and the non-binding stockholder proposal requesting majority vote for the director elections is approved. We expect to report our final results within four business days after this meeting on a form 8-K.
- 支持 Non-GAAP 营业利润率同比 +400bp 至 15%；全年指引 13.5%→14.5%，Q3 15.5%；FCF 率 23% 维持；GAAP 盈利目标 Q4 FY28 不变
This revised outlook includes a higher revenue mix from fast growing AI workloads, which carry a lower contribution margin today. We're delivering continued operating margin expansion as we offset growing cloud costs with slowing head count expense. We are increasing our FY 2027 non-GAAP operating margin guidance from 13.5% to 14.5%. For Q3, we expect non-GAAP operating margin of 15.5%. We're reiterating our full year non-GAAP adjusted free cash flow margin guide of 23%. I'd like to close with my two key goals for the year. First, help the business to deliver growth and margin expansion.
- 中性 仍按计划在 Q4 FY28 实现 GAAP 盈利
At the same time, use cases won per account executive increased 43% year-over-year, demonstrating both growing customer demand and strong sales productivity. And we're pairing this investment in growth with continued operational discipline. We remain on track for GAAP profitability in Q4 fiscal 2028, and the operating leverage we build along the way strengthens the durability of that outcome. Taken together, our rapid pace of innovation, tighter go-to-market execution and operational discipline positions us well to capture the huge opportunity ahead. This quarter demonstrated that the transition to the agentic enterprise is accelerating, and Snowflake is at the center of it.
- 中性 上半年新增 334 名员工 (含 Observe 173)，去年同期 935；用放缓的人头费用抵消上升的云成本。Q2 未提回购和 SBC
In Q2, non-GAAP operating margin expanded over 400 basis points year-over-year to reach 15%. Our outperformance was driven by strong revenue growth and disciplined head count management. Year-to-date, we've added 334 employees, which includes 173 from our Observe acquisition. This compares to 935 added in the year-ago period. We ended the quarter with $4.3 billion in cash, cash equivalents, short-term and long-term investments. Moving to our outlook.
- 支持 FY27Q2 非 GAAP OPM 15.3%，同比升 400bp 以上；FY27 OPM 指引从 13.5% 升到 14.5%；Q3 OPM 指引 15.5%
* Despite pressure on product gross margin, operating margin improved. FY27Q2 non-GAAP operating margin was 15.3%, up more than 400 basis points year over year, and FY27 guidance was raised from 13.5% to 14.5%.
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 年初至今净增 334 人 (含 Observe 173 人)，去年同期净增 935 人
> “Year to date, we have added 334 employees, including 173 employees who joined through the Observe acquisition. By comparison, we added 935 employees in the same period last year.”
来源背景：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 - 支持 上半年净增约 330 人 (约 180 来自 Observe)，去年同期约 950 人
And we guide 74% non-GAAP product gross margins and you can see the strong cash flow generation. And so, we're doing that with less people around the enterprise. One of the things I talked about on earnings, if you look at sort of year-to-date hiring for first and second quarter this year versus last year, we added roughly 330 people this year. Last year, we added about 950 people. Of the 330 this year that we added, 180-ish of those more or less came from the Observe acquisition. And so, the amount of people that we're adding with the acceleration of revenue, I think, is really a testament of how times are changing. Yeah.
- 支持 4Q FY28 GAAP 盈利承诺，SBC 有专门计划且显著下降
But we've done that with increasing operating leverage in the business. And so, what we are very committed to is increasing operating leverage in the business. For those of you who weren't at Investor Day, we committed to being GAAP profitable in fourth quarter of next year. And so, slightly over a year out. And so, we're seeing a really big -- we put a plan around stock-based compensation. You seeing a really big decline there. You're seeing us hold gross margins with the increased revenue that you see. And we guide 74% non-GAAP product gross margins and you can see the strong cash flow generation.
- 支持 工程招聘改为 1-2 年经验的 AI 原生年轻人，替代资深高薪人员 (降 SBC)
And what we're doing is we're hiring a different type of person that can actually be an engineer versus what we would historically hire. So, historically, you'd have to hire someone with eight to 10 years of experience, really senior, maybe had a management role in the past. And now you can hire someone one or two years out of college who really is AI build, that is vibe coded that use CoCo, and so forth. And so, we're seeing a dramatic difference in the two of them on the productivity. If you go into the finance organization, we have probably 150 different use cases.
- 中性 分析师指出主要竞争对手 FCF 盈亏平衡，质疑为何此时追求 GAAP 盈利；CFO 称没有约束投入
Why is this the right move at this time to push towards GAAP profitability? Sometimes... Because you've got a pretty big competitor out there that's running at free cash flow breakeven. Yeah. It's – I'll answer it at the reverse way in the sense, let's say, did I do anything that put constraints on the business to be able to make that comment? And the answer is, absolutely not. Like, we're spending as much as we want to spend in the business. We go in to these large enterprises, and our gross retention rate is really high.
- 支持 GAAP 盈利路径之一：工程不再招需要大额期权的资深人员，改招 1-2 年经验 AI 原生人员
And so we, as a company, will constantly get economies of scale and stuff that we do with the hyperscalers and all that. But we're using technology, primarily Snowflake, to actually get more out of the existing workforce. And then part of our plan to be GAAP profitable was, really, in engineering, you don't have to hire senior leaders that require large stock option packages. You can hire people with one to two years college experience or AI-pilled that can actually prompt and vibe-code and be more productive at a fraction of the cost. So, all around the business, like to me today, I got this thing I talk about called the Good Morning CFO skill. I run it.
- 支持 今年迄今净增约 330 人 (170-180 来自 Observe)，去年同期约 940 人
We increased our full year guide from 31% year-over-year to 36% year-over-year. So, increased it by 500 basis points in just one quarter. And at the same time, year-to-date this year, we've added roughly about 330 people, of which 170 to 180 of them came from Observe acquisition. So, 150 compared to the same period last year, there's roughly about 940 people. And so we're doing that with a lot less people as well. You were one of the first execs, at least that I interact with, that talked about the now disconnect between revenue growth and bodies, that you were driving massive efficiencies within your business.
产品理解
数据团队一天怎么用 Snowflake
这一节回答"数据团队每天实际怎么用 Snowflake"。按一天的时间顺序排列。
| 步骤 | 发生了什么 |
|---|---|
| 1. 数据接入 | 早上 Openflow/Snowpipe/Kafka (Datastream) 把 SaaS、Oracle、MySQL、日志等增量数据灌进 Snowflake 表或 Iceberg 表；Zero-Copy 方式直接读 SAP/Salesforce/Workday 数据，不复制 |
| 2. 转换/数据工程 | 数据工程师用 dbt 项目、Dynamic Tables、Serverless Tasks 跑 ELT 管道清洗建模；现在常让 CoCo 在 Snowsight 或 VS Code 里写/改 dbt 模型和 SQL，CoCo 也会诊断慢查询 (这一步是仓库 credit 的大头) |
| 3. 治理与语义层 | 平台团队在 Horizon Catalog 打标签、设掩码和行级策略，建 semantic views (定义'收入''活跃客户'等口径)，给 agent 一个统一的数据真相；Horizon AI Guardrails 拦 prompt 注入 |
| 4. BI 与分析 | 分析师用 Tableau/Power BI/Sigma 连 Snowflake 仓库跑报表，或在 Snowsight Notebooks 里写 SQL/Python；交互式仓库供低延迟看板 |
| 5. 业务用户问答 (CoWork) | 销售/财务在 CoWork (原 Snowflake Intelligence)、Slack、Teams 里用自然语言提问，agent 通过 Cortex Analyst 生成 SQL 在仓库执行、用 Cortex Search 查文档，出图表和 Deep Research 报告 |
| 6. AI/ML 与应用 | 数据科学家用 AI_COMPLETE/AI_EXTRACT 在 SQL 里批量处理文本/文档，用 Snowpark、Container Services、Cortex Training 训练模型；开发者用 Streamlit、Snowflake Postgres、App Runtime 做内部应用和 agent 后端 |
| 7. 迁移项目 | 迁移团队用 AIM (SnowConvert AI + Datometry) 和 CoCo 把 Teradata/Oracle/Hadoop/Spark 代码自动翻译并验证，迁移完成后历史工作负载转为长期 credit 消耗 |
| 8. 成本管控 | FinOps/管理员设资源监视器、预算、每用户 AI 配额，用 Gen2/Adaptive 仓库、自动挂起和查询优化 (常借助 CoCo 或 SELECT 等第三方工具) 压低账单 |
从流程里看出的三件事：
- 最大、最稳定的消耗是夜里的数据加工和白天的 BI 查询。这部分是"核心数仓"，增速 20–25%，跟着数据量和迁移走。Iceberg 开放写入以后，这部分是最可能被别的引擎分流的。
- AI 让"问数据的人"变多。以前只有会写 SQL 的分析师在查，CoCo 让工程师写得更快，CoWork 让业务人员直接用自然语言问。每多一个人问，就多一些查询，这是 AI 拉动核心 credit 最直接的路径 (BFSI 客户的 200 多个分行人员就是这样)。
- "控制 Snowflake 的账单"是平台团队的日常工作。设资源监控、调仓库大小、找出最贵的查询，现在 CoCo 也会主动做。客户对用量的控制是常态，不是例外。
Snowflake 自己怎么演示
2026 年 9 月 Snowflake 在上海的 BUILD/开发者活动、CoCo 黑客松和 Data for Breakfast (Merit 录音转写)。这是 Snowflake 员工自己对客户的演示，能看到产品实际怎么用，但是卖方视角。
| 产品事实 | 原话 | 出处 |
|---|---|---|
| CoWork 可经 MCP (转写为 NCP) 连 Google Calendar、Slack 等，一句话自动创建会议、发 Slack 消息 | 它会通过 NCP Service 连接 Google Calendar，查看是否已经安排了会议。如果没有，它会直接在 Google Calendar 中创建 Meeting Event。 | Snowflake 上海 BUILD/开发者活动 Sessi L752 |
| CoWork 有移动端，体验与电脑端一致，可看连接的服务、历史对话、表格和生成的 SQL | 手机端的体验与电脑端完全一致，也可以查看相关表格以及编写的 SQL Query。 | Snowflake 上海 BUILD/开发者活动 Sessi L766 |
| CoWork demo 中能直接问'AI 的 Cost 和消费是多少'，说明 Snowflake 在 CoWork 中暴露 AI 成本/消费查询 (demo 数据：7 月增长) | 接下来，我想进一步询问更加深入的信息，例如AI的Cost和消费是多少。VP可能也会关心AI到底花了多少钱。 | Snowflake 上海 BUILD/开发者活动 Sessi L748 |
| 季度末 review 演示：过去需 3 周整合数据，CoWork 一句话出结果并可发 Slack 给 VP | 现在Max需要准备季度末Review，过去可能要花长达3周的时间整合所有数据，但在Snowflake Co-work的帮助下，他只需要用1句话向Co-work提出要求 | Snowflake 上海 BUILD/开发者活动 Sessi L746 |
| AI 架构三层：上层工作空间、中层 API/Agent/内置 AI Function (可抽取 PDF 图片表格)、底层企业上下文；转写把 Cortex 误识为 Vertex | 其背后是 Vertex AI。如果大家听过上午的 Keynote，应该已经看到这个 3层结构。它是一个端到端的推理结构，中间层可能是我们的 API。我们提供了 Vertex Agent，也内置了 AI Function。 | Snowflake 上海 BUILD/开发者活动 Sessi L770 |
| Agent 原生继承行级/列级权限，每个 agent 操作可追溯 | Snowflake拥有非常严格的行级、列级安全策略，AI Agent会原生继承每个人的权限。Max能够看到的数据，AI Agent才能访问。 | Snowflake 上海 BUILD/开发者活动 Sessi L734 |
| 安全护栏：Prompt Injection 防护、Data Leakage 防护，客户数据/提示词不被外部基础模型用于训练 | 企业的敏感信息、提示词以及所有商业逻辑都完全属于企业内部，绝对不会被外部基础模型用于训练。 | Snowflake 上海 BUILD/开发者活动 Sessi L778 |
| Cortex 每日主动推送对账异常、Deep Research 回答 VP 复杂问题、自动生成可交互 Slides | 早上9点，Cortex会通过手机端应用主动推送数据对账异常警报。到10点半，如果VP提出较为复杂的商业问题，我们可以直接使用Deep Research功能，在几秒钟内回答这些问题。 | Snowflake 上海 BUILD/开发者活动 Sessi L780 |
| Snowflake 训练 Text-to-SQL 模型时专门学习 Snowflake SQL dialect | 第 1层是 data。我们生成了大量 domain-specific 数据，让模型先进一步了解 Snowflake dialect。 | Snowflake 上海 BUILD/开发者活动 Sessi L820 |
| SQL auto-completion 专用模型已经上线 | 第 2个想分享的是 Snowflake 的 auto-completion model。该模型已经进入市场。 | Snowflake 上海 BUILD/开发者活动 Sessi L830 |
| 工具和 skills 按需逐步加载，长代码给 context reference，以控制上下文膨胀 (类似 Claude Code skills 设计) | Snowflake Copilot通过3种方式解决context exposure问题。第1，tools和skills是iteratively added的，不会1次性加入所有可用内容。 | Snowflake 上海 BUILD/开发者活动 Sessi L846 |
| 多 agent hybrid research：先独立探索、再辩论验证、最后综合，每条主张附来源 | 系统设置了一个门控模块，国内通常称之为 EDS，用于控制各方之间是否可以相互影响。 | Snowflake 上海 BUILD/开发者活动 Sessi L892 |
| Agent 身份感知的 masking：同一数据，人看原文、agent 只能看掩码；可规定任何 PII 都不进入 agent | 这样，相关部门便可以有把握地确认：在整个数据库中，任何 PII 以及任何被标记为 PII 的信息都不会提供给智能体。即使用户调用了智能体并拥有相应权限，也无法访问这些信息。 | Snowflake 上海活动 Session3-4：Hori L773 |
| 受限会话 (restricted session)：agent 运行时只能用被授权的角色，即使用户有 200 个角色 | 即使配置了 200 个角色，代理实际使用的也只有获得授权的角色。 | Snowflake 上海活动 Session3-4：Hori L845 |
| Cortex Guard/AI 运行时治理是目录级设置，一键开启后覆盖 CoCo (转写为 CodeGo) 和 agents | This is is a catalog setting, you enable it, and it will cover you across your AI services, so both CodeGo, for the agent as well as for our, right? | Snowflake 上海活动 Session3-4：Hori L953 |
| 语义视图限定 agent 只能查询模型内的数据元素，再叠加用户权限 | 如果将语义视图定义为一组数据元素，智能体只能访问该视图内的数据元素，而且这些元素还必须符合访问规则，这一点非常重要。 | Snowflake 上海活动 Session3-4：Hori L803 |
| Horizon 可治理 Iceberg 表、客户自存数据，以及 SAP/Salesforce/Workday 的外部数据和元数据 | 目前，我们可以从多种主流企业数据源接入更多数据和元数据，包括SAP、Salesforce和Workday。 | Snowflake 上海活动 Session3-4：Hori L733 |
| ABAC：用户属性由 Okta/Entra 等 IdP 推送，策略按属性执行 | 这些属性可以通过身份提供商进行设置，由 Octa 或 Idra 推送。 | Snowflake 上海活动 Session3-4：Hori L761 |
| Copilot 用自然语言生成 tag、classification profile、masking policy，Trust Center 扫描出未 mask 的敏感对象 | 我们只需通过自然语言清楚描述扫描需求，Copilot就会自动生成对应代码。 | Snowflake 上海活动 Session3-4：Hori L929 |
| Databricks 通过 catalog 读 Snowflake 的 HR 表时，Snowflake 定义的脱敏策略仍生效 | 现在，我在 Databricks 中进行查询，让 Databricks 通过 catalog 读取 Snowflake 的表。首先，它可以读取刚才的 HR 表；随后对 HR 表执行 select 时，同样可以看到整个 governance 策略已经生效。 | Snowflake 上海活动 Session3-4：Hori L937 |
| Snowflake 读写 Databricks 管理的 Iceberg 表只需 2 条语句 (Catalog Link to Database) | 完成嫁接只需要2条非常简单的语句。完成后，可以直接读取 Databricks Managed 中继承的表，Shipping 表已经可以完整读取。 | Snowflake 上海活动 Session3-4：Hori L899 |
| 对外部负载提供 Snowflake Connector for Spark 和 SQL Scan 两种技术 | Snowflake提供2种技术支持：1种是专门面向Spark负载的Snowflake Connector for Spark，另1种是Spark之外的SQL Scan。 | Snowflake 上海活动 Session3-4：Hori L925 |
| 自动分类功能在中国区已 GA | 该功能在中国区已经GA，因此建议使用中国区Snowflake的用户开启这一功能。 | Snowflake 上海活动 Session3-4：Hori L921 |
| Polaris 计划扩展到 Delta、Hudi 等表格式 | 未来，它不仅支持 Iceberg，还会扩展至 Delta、Hudi 及其他开放的 Table Format。 | Snowflake 上海活动 Session3-4：Hori L881 |
读法：演示的重点几乎全在"治理"上：agent 继承行级/列级权限、PII 不进 agent、每个 agent 操作可追溯、Databricks 读 Snowflake 的表时脱敏策略仍生效。这就是 Snowflake 想把锁定放到的那一层：数据可以开放，但"谁能看什么"由 Snowflake 管。另一个重点是 CoCo 融入开发者已有的工具 (VS Code、CLI、Git)，并能直接调 Marketplace 数据，和 Claude Code、Cursor 正面竞争。
客户案例和新产品
从证据里摘出的具体客户。正面案例多来自 Snowflake 的合作伙伴，反面案例来自客户自己。
| 客户 | 发生了什么 | 来源 |
|---|---|---|
| 一家 BFSI 客户 | 把 Cortex Agents/CoWork 开放给 200 多个分行人员，每月 Snowflake 消费从 $4 万到 $7 万，仍在增长 | Merit 专家 2026-09-09 |
| 北欧一家保险公司 | 从试点到多业务线生产，计算消耗涨 5–8 倍，主要来自理赔人员的查询；但 AI 负载只占总消耗 15–20% | Merit 专家 2026-08-26 |
| 一家零售客户 | AI 负载涨 2–3 倍，账户整体只涨约 40% | MeritForum 2026-08-27 |
| 一家网络设备/制造厂商 | Teradata 迁移计划不到 3 个季度完成，以前要 2–3 年 | Q2 业绩会、高盛会议 |
| 一家半导体公司 | 最初用 CoCo 是为了优化查询省钱，之后供应链部门扩展使用，公司称消费增加 (无数字) | Investor Day |
| 一家 Spark 遗留客户 | 用 CoCo 把代码转成 Snowpark，快 5 倍、成本更低 | Investor Day |
| Snowflake 自己 | 取消了所有 BI 和分析软件合同，用 CoCo + Streamlit 自己搭看板 | Piper 会议 (CFO) |
| 一家大型药企 | Snowflake 支出同比只增 10–15%，CoCo 尚未真正使用 | Merit 专家 2026-08-10 |
2026 年新产品
| 产品 | 是什么 | 怎么收费 | 状态 |
|---|---|---|---|
| Snowflake CoCo (原 Cortex Code) | 面向数据工程师/开发者的编码 agent，在 Snowsight、Desktop、CLI、VS Code、Excel、Claude Code 中写 SQL/dbt/Python、搭管道和应用；Cloud Agents 后台跑长任务 | 按 token 计 AI Credit (Sonnet 4.6 输入 1.65/输出 8.25 每百万)；CLI/Desktop 独立月度订阅含固定额度，超出后暂停；执行的 SQL 另计仓库 credit | Snowsight GA 2026-02 (2026-04-01 起计费)；CLI 独立订阅 2026-02-24；Desktop/VS Code/Cloud Agents GA 2026-06-02 (Summit)；Agent SDK/MCP 预览 |
| Snowflake CoWork (原 Snowflake Intelligence) | 面向所有员工的 agent 前端：自然语言问数、Deep Research、Artifacts 看板、个人记忆、MCP 连接 Slack/Drive/Salesforce | 按 token 计 AI Credit (Sonnet 4.6 输入 1.95/输出 9.76) + 工具调用 + 生成 SQL 的仓库费用；无席位价，有每用户配额 | Snowflake Intelligence GA 2025-11；更名 CoWork 并扩展功能 2026-06-02 (核心 GA，Deep Research/Artifacts 'GA soon'，User Memory 公开预览) |
| Cortex Sense | CoCo/CoWork 下面的运行时上下文层，从查询历史、看板、元数据学习业务语境以提升 agent 准确率 | 未公布 | 私有预览 2026-06；Investor Day 称难题准确率 86.3% vs 前沿 coding agent+SQL MCP 24.1%、月成本 $19.47 vs $78.08 |
| Openflow | 基于 Apache NiFi 的托管数据接入，连接数据库 CDC、SaaS、非结构化文件；给数据工程师 | Platform Credit 按 vCPU-小时 (BYOC 0.0225) 或 SPCS 节点小时；部分连接器另收许可费 (Oracle $70/月) | GA (Q4 FY26 电话会 2026-02 称 GA)；2026-06 Summit 称三朵云 GA，Oracle 连接器 GA |
| Snowflake Postgres | 托管 PostgreSQL (来自 Crunchy Data 收购)，给应用开发者和 agent 做事务库；可镜像到 Snowflake 分析 | 实例按 Platform Credit/小时 (STANDARD_M 0.0356) + Postgres 存储 $117.76/TB/月 | GA 2026-02；Data Mirroring 公开预览 2026-06 |
| Observe (by Snowflake) | 建在 Snowflake 湖仓上的可观测性平台 (日志、指标、链路 + AI SRE)，给运维/SRE，与 Datadog/Splunk 竞争 | 可用 Snowflake credit 抵扣，也可单独购买；具体单价未在 Snowflake 页面公开 | 收购于 2026 年初完成 (Q4 FY26 电话会称 'recent acquisition')；FY27 预计贡献约 1pt 产品收入增长 |
| Natoma | 企业 MCP 平台/网关，约 100 个连接器，控制 agent 能读写哪些 SaaS/数据库/工具，按工具调用审计 | 未披露 | 2026-05-27 宣布收购意向，待完成 |
| Horizon Catalog / Horizon Context | 治理与目录层 (标签、策略、血缘、Agent Identity)；Horizon Context 把治理元数据喂给 CoCo/CoWork/agents | 内含在平台中，不单独收费 (消费表无单列) | Horizon Context 2026-06 发布；AI Guardrails、Agent Identity GA 2026-06 |
| Polaris / Open Catalog | 开源 Iceberg REST 目录，让 Spark/Trino/Databricks 等外部引擎读写同一份 Iceberg 数据 | Open Catalog 0.5 credit/百万请求，限时免费 | Polaris REST 接口并入 Horizon Catalog，2026-06 GA |
| Iceberg tables (含 Iceberg V3、Snowflake 托管 Iceberg 存储) | 在开放格式 Apache Iceberg 上用 Snowflake 引擎读写；V3 支持删除向量、行血缘、variant | 计算照常收 Platform Credit；外部存储由客户付云厂商，Snowflake 托管存储按存储价 | Iceberg tables GA 2024；V3 与托管存储 GA 2026-06 (AWS/Azure，GCP 稍后)；External Engine Access Management 私有预览 |
| Gen2 标准仓库 | 第二代仓库硬件/软件，DML 和分析更快 | 每小时 credit 为 Gen1 的 1.35 倍 | GA 2025 |
| Adaptive Compute / Adaptive Warehouses | 自动为每个查询选择算力大小，取代手动选仓库尺寸 | Platform Credit，只在查询运行时计费，无固定费率 | Summit 2025 预览；Summit 2026 称 GA/即将上线 (2026-06) |
| AIM (AI-powered Migrations) | SnowConvert AI + Snowpark Migration Accelerator + Datometry 合成的迁移套件，CoCo 内有迁移 agent | 未公布 (迁移工具多为免费引流，收入来自迁移后的消费) | GA 2026-06 |
| Cortex Training | 托管 GPU 微调/强化学习开源模型 (Qwen、Mistral)，数据不出 Snowflake | 未公布；现有 fine-tuning 按训练 token 计 Platform Credit (llama3.1-70b 3.40/百万) + 推理 AI Credit | 2026-06 发布，私有预览 |
| Snowflake Datastream | 兼容 Kafka 的托管流式接入 | 未公布 | 私有预览 2026-06 |
竞品事实
| 对手 | 事实 | 来源 |
|---|---|---|
| Databricks | 2026-08-13：收入 run-rate >$70 亿，Q2 同比 >80%；Lakehouse (数仓/DBSQL) run-rate >$15 亿、同比 >100%；Lakebase (Serverless Postgres) >$1 亿；$1M+ 客户 >1,000、$10M+ >100；TTM 调整后 FCF 为正；$50 亿融资、估值 $1,900 亿。对比 SNOW FY27 产品收入指引 $60.7 亿 (+36%)，Databricks 规模已超 SNOW 且增速约为两倍，数仓业务增速 >100% | 链接 |
| Google BigQuery | 按需 $6.25/TiB 扫描 (每月首 1 TiB 免费)，或 Editions 按 slot-小时计费 (按扫描量计价，与 Snowflake 按仓库时长不同)；Google Cloud Q2 2026 收入 $248 亿、同比 +82%，backlog $5,140 亿 (未单独披露 BigQuery) | 链接 |
| Amazon Redshift | Serverless $0.375/RPU-小时 (美东)，预置 ra3.xlplus $1.086/节点小时；AWS 同时是 Snowflake 最大渠道 (AWS Marketplace 年销售 >$20 亿) 和 5 年 $60 亿基础设施承诺对象，更多是合作而非主要威胁；CoWork 可直接查询 Redshift 数据 | 链接 |
| Microsoft Fabric | FY26 Q4 (2026-07-29 电话会)：付费 Fabric 客户 >40,000、同比 >60%；Azure PostgreSQL 收入 +55%，连续三季加速；Fabric 收入未单独披露。Fabric 按容量单位 (F SKU) 打包，与 Power BI/M365 捆绑 | 链接 |
| ClickHouse | 2026-05-27 自披露年化收入 $2.5 亿；Sacra 估 2026-08 约 $3.5 亿 (2025 年末约 $1.6 亿)；客户 4,000；2026-01 D 轮 $4 亿、估值约 $150 亿；也推出 Postgres 服务和 Claude 驱动的 ClickHouse Agents，收购 Langfuse/HyperDX 做 AI 可观测性 (同时与 Observe 竞争) | 链接 |
视频导读
视频都打开 YouTube 页面核实过标题和日期。建议先看 Summit keynote 建立整体概念，再看 CoCo 和 CoWork 的演示，最后看 Iceberg/Polaris。
| 视频 | 频道 · 日期 | 看什么 |
|---|---|---|
| Snowflake Summit 2026 Opening Keynote With Sridhar Ramaswamy And Daniela Amodei | Snowflake Inc. · 2026-06-01 | CEO 讲 agent 时代战略和 Anthropic 关系，看'数据控制面'叙事原文 |
| Snowflake Summit 2026 Platform Keynote | Snowflake Inc. · 2026-06-11 | CoWork、CoCo、Adaptive Compute、Iceberg V3 的现场演示，看产品实际能做什么 |
| Snowflake Summit 2026 Builder Keynote | Snowflake Inc. · 2026-07-10 | 面向开发者：CoCo 跨数据开发、Postgres、应用，看工作负载如何落在 Snowflake 上 |
| Christian Kleinerman & Elena Fersman | Snowflake Summit 2026 | SiliconANGLE theCUBE · 2026-06-03 | 产品负责人+爱立信客户谈从 AI 试点到生产，看客户视角 |
| Snowflake CoWork Is Here – What's New in Enterprise AI | Snowflake Summit 2026 | Snowflake Developers · 2026-06-25 | CoWork 和 Cortex Agents 新功能演示，看每次问答背后触发多少 SQL |
| From One Prompt to Production: Building with Snowflake CoCo | Snowflake Developers · 2026-06-04 | CoCo 一个 prompt 跨 Snowflake/Databricks/GCP 搭流失预警系统——注意 CoCo 也在帮客户用别家引擎 |
| Behind the Cape - Cortex Code for Performance Optimization & AI Insights | Snowflake Developers · 2026-05-04 | CoCo 找瓶颈、给优化建议——看 AI 如何帮客户省 credit (反向因素) |
| Get work done in one place with Snowflake Intelligence | Snowflake Inc. · 2026-04-21 | 销售负责人用 Intelligence 做客户分析、会前简报、Deep Research 的完整流程 |
| Talk To Your Data: Snowflake Intelligence Demo | Snowflake Inc. · 2025-06-03 | 2025 年首次发布时的演示，对比一年后 CoWork 的进化 |
| [LIVE] Using Cortex Code with Snowflake Openflow | Snowflake Developers · 2026-03-12 | CoCo 部署 Openflow 连接器，看 AI 如何加速数据接入 (更多数据进 Snowflake) |
| Getting Started With Snowflake Openflow | First Look At How Adesso Uses Openflow | Snowflake Developers · 2025-10-17 | Openflow 搭 MySQL→Snowflake 管道的基础演示 |
| How Iceberg Tables Work In Snowflake | DEMO | Snowflake Developers · 2023-01-31 | Iceberg 表原理：Snowflake 引擎读外部存储的开放格式数据——理解计算可分流的基础 |
| Build AI Agents in Minutes with Snowflake Postgres & Vercel | Live Demo | Snowflake Developers · 2026-06-25 | Snowflake Postgres 作为 agent 应用后端，看 OLTP 新消费池 |
| Snowflake Gen2 Warehouses: They're faster, but should you switch? | SELECT · 2025-05-21 | 第三方成本优化公司测 Gen2 性能与成本，看性能改进带来的收入逆风 |
| It Depends 110: Snowflake Summit: Will Allen - Snowflake Cowork, Cortex Agents & the Era of Builders | Sanjeev Mohan · 2026-06-04 | CoWork/Cortex Agents 产品负责人访谈，看 agent 在生产中的真实进度 |
反过来想 (Invert)
不问"SNOW 怎么赢"，而问"怎样才会让持有 SNOW 的人在 2 年后亏掉一半"，再看每条路径今天有多少证据。
| 亏一半的路径 | 机制 | 今天的证据 | 可能性 |
|---|---|---|---|
| 1. 前置的加速退潮 + 估值回到历史区间 | 迁移提速、容量更快用完是一次性台阶；FY28 增速回到约 25%，EV/Sales 从 19 倍回到 10–12 倍 (2024–25 年的区间) | use case 增速 114% → 89%、每 AE 86% → 43%；RPO 38% → 30%；迁移时间已经降到约 6 个月，"无法降到零" | 中 |
| 2. AI 层越成功，毛利越低 | AI 收入占比从约 10% 升到 20–25%，毛利 35–55%，产品毛利率从 75% 漂到 71–72%；token 继续通缩 | 毛利率指引已降 1 个点；CEO 说愿意接受；token 半年降价一半 | 中 |
| 3. 开放架构让计算被分流 | Iceberg + Polaris 开放写入，ETL 这种最大的负载可以去 Databricks 或别的引擎；Snowflake 只保住治理和一部分查询 | CEO："迁出也一样变快"；新预算 60:40 偏 Databricks；Snowflake 自己宣传"可以彻底离开" | 中低，2–3 年以上 |
| 4. 2023 年重演：AI 成本优化浪潮 | 企业 AI 试验成本到了董事会层面，开始砍；用量制让节省立刻变成收入下降 | 一位专家说有公司 1 个月用完全年 AI 预算；CoCo 的成本优化技能是前十大；2023 年增速一年从 70% 掉到 38% | 中 |
| 5. 什么都没发生，但股价不涨 | 收入每年 +25–30%，SBC 慢慢降，估值从 19 倍收敛到 12–14 倍，2 年回报接近 0 | 这是我的基本情形：按每股 FCF 口径，$320 的 5 年年化回报约 5–6% | 高 |
反过来，什么会让我错：
- FY28 首次指引 ≥30%：管理层只把看到的消费行为放进指引，如果首次指引就有 30%，说明 run-rate 足够硬，前置成分没有我想的大。
- AI 毛利被拉回来：开源模型自己做推理、批量采购、CoCo 的高溢价，让 AI 产品毛利到 60–70%，产品毛利率稳在 74–75%。那第 2 层就不是"低质量收入"。
- SBC 降得比斜率更快：如果 FY28 SBC 降到收入的 18% 以下、FCF 率保持 25%，FCF − SBC 会到 7% 以上，口径 A 的价值会大幅上升。
- Postgres/Openflow 带来新的大负载：如果 Snowflake 从"分析"扩到"交易库"和"数据接入"，TAM 和增长的持续性都会更好。目前还在早期。
框架打分：计费单位三问和赢家五条
计费单位三问
| 问题 | SNOW 的答案 | 和 TEAM、DDOG 比 |
|---|---|---|
| 1. AI 让计费单位变多还是变少？ | 变多，但 AI 也在压单位消耗 | TEAM 按 seat，agent 替人可能让 seat 变少；DDOG 按主机/数据量，AI 让应用变多。SNOW 按 credit，AI 让查询和问数据的人变多；但 CoCo 写更高效的 SQL、帮客户优化，Gen2/Adaptive Compute 让每个查询用的 credit 更少 |
| 2. 被计量的东西是不是它自己拥有的？ | 一半是 | 核心 credit 是自己的引擎、自己的数据和治理；AI credit 有一大块是 Anthropic/OpenAI 的 token。存储和计算正在通过 Iceberg 开放给别的引擎，它保住的是治理和目录 (Horizon/Polaris)。比 DDOG 更"转售"，比 TEAM 好 |
| 3. 产出能不能数？ | 能 | credit、token、查询都能精确计数。但能数也意味着客户和 CoCo 自己能精确优化 |
赢家五条 (从 CRWD/DDOG/NET/SNOW 的反转里总结的)
| 条件 | SNOW 现在 | 判断 |
|---|---|---|
| ① AI 让计费单位变多 (不是 seat) | 用量制，几乎没有 seat 收入；AI 让查询、问数据的人、模型调用都变多 | 满足 |
| ② 报告增速连续 ≥2 个季度加速，指引上调 | 产品收入 29% → 30% → 34% → 37%；全年指引 27% → 31% → 36% | 满足 |
| ③ 不是靠一个 AI 客户 | 管理层说加速来自很广的客户，AI 原生只占收入一小部分 | 满足 |
| ④ 拥有被计量的东西 (毛利率漂移 ≤3–4pt) | 产品毛利率 77% → 74.7%，指引 74%，漂移约 2–3pt；AI 层是转售，存储/计算在开放 | 勉强满足，趋势不好 |
| ⑤ 和板块一起被错杀，而 1–2 条已经成立 | 2026 年 2–4 月确实如此 ($263 → $121)，当时 Q4 FY26 已经 +30%、RPO +42%；现在已涨 2.6 倍 | 已过时点 |
| 价值检验：FCF − SBC 和估值 | FCF − SBC 约收入 -8%，EV/Sales 约 19 倍 FY27 | 不满足 |
结论：SNOW 和 DDOG 一样是这个框架下的"赢家"样本，满足前三条，所以 4 月低点之后涨了 2.6 倍 (比 DDOG 的 2.5 倍还多)。它在第四条上比 DDOG 弱 (AI 层是转售、毛利在漂)，在价值检验上也更差 (SBC 30% vs 21%)。好公司已经被定价成好公司了。
估值和买点
现在的价格：$319.61 (2026-10-20 收盘)，非 GAAP 稀释股数约 380M (FY27 指引)，市值约 $1,215 亿；现金及投资 $43.3 亿减去可转债 $22.8 亿，净现金约 $20.5 亿，EV 约 $1,194 亿。FY27 总收入约 $63.3 亿 (产品收入指引 $60.7 亿加专业服务)，EV/FY27 收入约 19 倍；如果 FY28 增长 28%，EV/FY28 收入约 15 倍。过去 12 个月 FCF 收益率约 1.0%，扣 SBC 后 -0.4%。
两种口径
- 口径 A (你的默认口径)：把 SBC 当成真实成本。FY32 的 (FCF − SBC) × 终值倍数，按 10% 折现 5 年回今天，加净现金，除以今天的股数。
- 口径 B (市场在用的口径)：不扣 SBC，但把它当稀释：FY32 的每股 FCF (股数每年 +3%) × 终值倍数，再折现。这个口径偏乐观：它把约占收入 30% 的 SBC 从成本里拿掉，只用每年 3% 的股数增长来近似 SBC 的代价。SBC 越高，两种口径的差距越大，所以 SNOW 的 A/B 差距比 DDOG 大得多。
| 情景 | FY28–32 增速 | FY32 收入 | FCF 率 / SBC 率 (FY32) | 终值倍数 | 口径 A 价值 | 口径 B 价值 | 现价 5 年年化 (A / B) |
|---|---|---|---|---|---|---|---|
| 悲观 | 24/18/15/13/11% (复合 16%) | $134 亿 | 26% / 20% | 30x | $45 | $152 | -27% / -6% |
| 基本 | 30/25/22/19/16% (复合 22%) | $173 亿 | 30% / 17% | 35x | $134 | $262 | -8% / +6% |
| 乐观 | 34/30/26/23/20% (复合 27%) | $205 亿 | 33% / 14% | 45x | $292 | $435 | +8% / +17% |
FY27 总收入取约 $63.3 亿；折现率 10%；终值倍数是对 FY32 (FCF − SBC) 或每股 FCF 的倍数。基本情形的 FY28 +30% 和 Q3 指引 (+37–38%) 之后的自然减速一致；SBC 17% 是"41% → 34% → 27%"的斜率放缓后的结果。
反推：现价隐含了什么
要在 $320 买入还拿到 10% 年化：
- 口径 A，FCF − SBC 率 13%、35 倍：FY32 收入要到约 $423 亿，5 年复合 46%。
- 口径 A，SBC 降到 12% (FCF − SBC 率 18%)：收入要到约 $305 亿，复合 37%。
- 口径 B，FCF 率 30%、35 倍：收入要到约 $212 亿，复合约 27%。
所以市场是按口径 B、5 年复合约 27% 在定价，比我的基本情形 (22%) 高 5 个点，接近乐观情形。考虑到加速里有前置成分、AI 层在通缩、新预算在流向 Databricks，这个预期偏满。
两年一倍检验
$639，算上每年 3% 的稀释，市值约 $2,580 亿。FY29 收入按 FY28 +30%、FY29 +27% 算约 $105 亿，需要约 24.5 倍 EV/Sales。SNOW 只有 2021 年泡沫期到过这个水平。不满足。
买点
| 价格 | 口径 A 基本情形 5 年年化 | 口径 B 基本情形 5 年年化 | 怎么看 |
|---|---|---|---|
| $150 | 7.5% | 23% | 2024-09 的 $109 和 2026-04 的 $121 都曾出现 |
| $180 | 3.5% | 19% | |
| $200 | 1.3% | 16% | |
| $220 | -0.7% | 14% | 买点区间 |
| $240 | -2.4% | 12% | 买点区间上沿 |
| $280 | -5.5% | 8.5% | |
| $320 (现价) | -8% | 5.6% | 不买 |
| $350 | -9.7% | 3.7% |
结论：买点 $210–240，对应口径 B 基本情形 12–15% 年化。口径 A 下基本情形最多值约 $134，只有在乐观情形 (5 年复合 27%、SBC 降到 14%) 下，$210–240 才同时在口径 A 下有正回报。所以这个买点隐含的判断和 DDOG 一样：SBC 会按管理层的斜率下降，或者增长好于基本情形。SNOW 在这一点上比 DDOG 更有依据 (管理层给了 GAAP 盈利日期和 SBC 斜率)，这是我愿意接受比 DDOG 更高倍数的原因。
自己算
置信度检查
| 来源 | 检查范围 | 身份与视角 | 影响判断的问题 | 采用边界 |
|---|---|---|---|---|
| 2 场业绩会 (Q1、Q2 FY27，FactSet 全文) | 全文 | CEO Ramaswamy、CFO Robins、产品 EVP Kleinerman、IR | AI 收入、AI 毛利率、AI 采用者的消费提升都拒绝给数 | 可作主要证据 (限"管理层怎么说") |
| 5 场会议 (Investor Day、股东大会、高盛、花旗、Piper) | 全文 | CEO 或 CFO | FactSet 近实时转录有错字；"一半"在不同场合口径不同 | 可作主要证据 |
| Merit 音频 + MeritForum | 全文 | 北欧/北美 SI、渠道、合作伙伴、客户、前 Databricks 销售 | MeritForum 整理稿和音频有矛盾 (以音频为准)；合作伙伴有立场；对手一方有立场 | 需交叉验证 |
| Snowflake 上海活动录音 (Merit) | 全文 | Snowflake 自己的工程师和 SE | 卖方视角，但对产品和开放架构的说法是一手的 | 产品事实可用；对数字仅作线索 |
| GuidePoint / ThirdBridge | 全文 | 独立顾问、Thoughtworks、前云厂商、可观测性同业 | 部分专家主要看 Databricks | 需交叉验证 |
| Acecamp、Funda、TMTB | Funda/Acecamp 全文；TMTB 关键词段落 | 二手转述卖方、渠道和客户 | 看不到底稿；Funda 两位合作伙伴对 CoCo 毛利的判断相差很大 (52–55% vs 60%+) | 需交叉验证 |
| Q3、Q4 FY26 业绩会 (Snowflake IR 的 corrected transcript PDF) | 按主题抽取的逐字句子 (各约 100 句)，不是全文 | CEO、CFO、分析师 | 句子之间不连续，所以这两场的引文只有单句、没有补成完整段落；个别问答配对是抽取时配的 | 可作主要证据 (单句) |
| 公司新闻稿、10-Q、Investor Day 材料 | 季度数字和资产负债表 | 公司 | 10-Q 后半部分 (股权/债务附注) 没读到，回购余额用 Investor Day 数字推算 | 可作主要证据 |
影响判断的口径问题
| 原话或数字 | 问题 | 对结论的影响 |
|---|---|---|
| "约一半加速来自 AI 产品" vs "一半 beat 来自 AI 产品" | 同一个 CFO 在两个场合说法不同：一个是增速加速的拆分，一个是超预期的拆分 | 不能直接说"AI 贡献了 37% 里的一半" |
| AI 收入按 SKU 划分 | 包括 token 转售 | "AI 驱动"里有一部分是低毛利收入 |
| 专家的按产品拆分 (核心 60–65% / Snowpark 20–25% / AI 10–15%) | 单一专家的估算 | 我用它做了贡献拆分，只能当量级参考 |
| CoCo "约占收入 2–2.5%" vs "约占 consumption 3%" vs "15–20%" | 不同专家口径差很大，15–20% 可能指 Cortex 整体或增量 | 取 2–3% 作为 CoCo 本身的量级 |
| "去年约 200 家客户迁往 Databricks" | 对手一方的说法 | 只作为方向参照 |
| "采用者核心消费 +11%" | CEO 的 LinkedIn 帖子转述，没有基准 | 不作为量化证据 |
附：来源与阅读记录
共 76 份有效材料：管理层发言 9 场，其余为专家访谈、渠道调研、Snowflake 自己的活动录音、卖方和 Substack 摘要。分级：A 核心 37 / B 实质 26 / C 顺带提及 13 / X 未取到 1。证据 814 条，另有反方原话 151 条，全部经脚本逐字比对原文。部分 Merit 音频和 MeritForum 整理稿是同一通话，计数时没有剔除。
专家与外部来源背景
| 日期 | 来源 | 背景与局限 | 分级 | 可信度 |
|---|---|---|---|---|
| 2026-01-20 | MeritForum 纪要 (整理稿) | 自述'We are a global systems integrator with approximately 300,000 employees'，在其 AI 业务中做 Palantir 渠道/实施 (Palantir 占其 AI 业务 <5%)，与 Microsoft/Databricks/Snowflake 等都有合作；面向企业客户，熟悉欧洲公共部门。主题是 Palantir，Snowflake 只是顺带。视角局限：SI 视角、看重'锁定'叙事；无 Snowflake 具体数字。 | C | 仅作局部线索 |
| 2026-01-28 | Acecamp | 未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。 | B | 需交叉验证 |
| 2026-02-23 | MeritForum 纪要 (整理稿) | 正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 。 | A | 仅作局部线索 |
| 2026-02-24 | MeritForum 纪要 (整理稿) | 正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。 | A | 需交叉验证 |
| 2026-02-24 | Acecamp | Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。 | B | 需交叉验证 |
| 2026-02-26 | MeritForum 纪要 (整理稿) | 作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。 | B | 可作主要证据 |
| 2026-02-28 | MeritForum 纪要 (整理稿) | 作者 Mo Zhou (Merit)；无专家信息 | X | 仅作局部线索 |
| 2026-04-28 | GuidePoint | Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示/)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 ；成本数字是方向性估计而非报价 ；访谈问卷后半段问 Dataiku，与 Snowflake 无关。 | B | 仅作局部线索 |
| 2026-04-29 | GuidePoint | Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示)，领导约 3000 人公司的 AI 与数据部门 ；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 ；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 。 | B | 仅作局部线索 |
| 2026-04-29 | GuidePoint | Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示/、)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge')；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。 | B | 仅作局部线索 |
| 2026-05-08 | GuidePoint | Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，/)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍)。GuidePoint 视角标注为 Reseller/Franchisee 。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' ；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。 | A | 需交叉验证 |
| 2026-05-19 | GuidePoint | Harsh Singla，标题栏写 'Former Regional Sales Director - Dremio' ，但主持人介绍为前 Dremio APAC 企业客户经理 (约 2 年多)，现任 Confluent 区域销售总监约 3 年 ；更早在 Qubole。即离开 Dremio 约 3 年。视角局限：APAC 销售，离开 Dremio 已久；对 Snowflake 只有泛泛比较；Dremio 立场天然推开放湖仓、反数仓。 | C | 仅作局部线索 |
| 2026-05-20 | GuidePoint | Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示/)；GuidePoint 视角标注为 Industry Consultant 。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。 | A | 需交叉验证 |
| 2026-05-20 | TMTB (Substack) | TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。 | B | 仅作局部线索 |
| 2026-05-25 | Funda (Substack) | 作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。 | A | 需交叉验证 |
| 2026-05-28 | Funda (Substack) | 作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。 | A | 需交叉验证 |
| 2026-06-01 | Funda (Substack) | 作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。 | B | 仅作局部线索 |
| 2026-06-03 | Merit 专家电话 (音频转写) | 专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。 | A | 需交叉验证 |
| 2026-06-03 | Merit 专家电话 (音频转写) | 专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。 | A | 需交叉验证 |
| 2026-06-03 | MeritForum 纪要 (整理稿) | 同专家A。本稿标题给出身份“Snowflake离职专家”，作者 Lantian，没有给职位和年份。视角局限同专家A。 | C | 仅作局部线索 |
| 2026-06-03 | MeritForum 纪要 (整理稿) | 同专家B。本稿标题给出身份“海外头部咨询公司离职专家”，作者 Lantian，没有给职位和年份。视角局限同专家B。 | C | 仅作局部线索 |
| 2026-06-04 | Merit 专家电话 (音频转写) | 专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。 | A | 需交叉验证 |
| 2026-06-04 | MeritForum 纪要 (整理稿) | 同专家C。本稿标题标为“北美头部软件公司离职专家”，作者 Lantian；这个标签和转写里“our Snowflake practice / sister practice”的说法不符。视角局限同专家C。 | C | 仅作局部线索 |
| 2026-06-04 | TMTB (Substack) | TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。 | B | 仅作局部线索 |
| 2026-06-15 | ThirdBridge | Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 之后重复一遍，以前半部分行号为准。 | B | 仅作局部线索 |
| 2026-07-28 | ThirdBridge | Ashish Shah，前 Cloudera 高级产品经理 (2016 年起在 Cloudera 9 年多，先后任销售工程师、ISV 合作伙伴管理 2018-21、产品管理，最后三年负责 SDX 治理组件，明示)；现任 Dataiku Technical Account Manager 。视角局限：Cloudera 本地/混合视角；对 Snowflake 的认识来自竞争丢单和朋友转述 ('I don't have any direct evidence')；离开 Cloudera 时间约 2025-2026。 | C | 仅作局部线索 |
| 2026-07-30 | ThirdBridge | Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示/)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。 | B | 仅作局部线索 |
| 2026-08-03 | Acecamp | 未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。 | A | 需交叉验证 |
| 2026-08-04 | Merit 专家电话 (音频转写) | 未具名专家，自述曾在泰国一家传统大型企业、一家覆盖英美的中型科技公司 (AI 非唯一业务) 和美国小型初创工作 (明示，)，职位未披露；中型科技公司经历中参与过 RAG 早期建设 。视角局限：样本为个人经历的三家公司；更偏工程/AI 应用，未直接谈 Snowflake 账单。 | C | 仅作局部线索 |
| 2026-08-06 | Merit 专家电话 (音频转写) | 姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。 | A | 仅作局部线索 |
| 2026-08-06 | Merit 专家电话 (音频转写) | 姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。 | A | 需交叉验证 |
| 2026-08-06 | MeritForum 纪要 (整理稿) | Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。 | B | 仅作局部线索 |
| 2026-08-06 | MeritForum 纪要 (整理稿) | MeritForum 标为'北美软件渠道商专家'，但内容与 260806 Merit Snowflake季度调研III 转写一致 (Q4 29%/Q1 30%/Q2 31%/Q3 33-34%)，专家自述在北欧地区、是 Snowflake 合作伙伴兼经销商，标签应以转写为准。局限同该转写。 | C | 仅作局部线索 |
| 2026-08-10 | Merit 专家电话 (音频转写) | 姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。 | A | 可作主要证据 |
| 2026-08-12 | Merit 专家电话 (音频转写) | 姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。 | A | 需交叉验证 |
| 2026-08-12 | MeritForum 纪要 (整理稿) | MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。 | A | 需交叉验证 |
| 2026-08-12 | MeritForum 纪要 (整理稿) | MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。 | B | 需交叉验证 |
| 2026-08-12 | Merit 专家电话 (音频转写) | Databricks 现任销售 (明示，)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 ；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。 | B | 需交叉验证 |
| 2026-08-13 | Merit 专家电话 (音频转写) | 未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”、“one project that I have recently worked on from one of the BSI” (原文)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。 | B | 需交叉验证 |
| 2026-08-13 | MeritForum 纪要 (整理稿) | MeritForum 标题明示“databrick离职专家”，作者 Lantian ；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering,)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。 | B | 仅作局部线索 |
| 2026-08-14 | Merit 专家电话 (音频转写) | 微软前沿部署工程师 (FDE) 团队成员 (纪要明示其团队直接服务 14 家 FDE 客户、可观察约 70 家，)，现任，职级与年限未披露。视角局限：微软立场，倾向推荐 Azure 体系；对 Snowflake 的判断是推理而非账单数据。 | C | 仅作局部线索 |
| 2026-08-17 | Acecamp | 未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。 | A | 需交叉验证 |
| 2026-08-19 | Merit 专家电话 (音频转写) | 未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。 | A | 需交叉验证 |
| 2026-08-19 | MeritForum 纪要 (整理稿) | 标题'北美软件渠道商离职专家'；与 2026-08-19 Merit 音频转写为同一场访谈，身份局限见该条 | B | 仅作局部线索 |
| 2026-08-20 | GuidePoint | Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示/)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 ；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 ；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。 | A | 需交叉验证 |
| 2026-08-26 | Merit 专家电话 (音频转写) | 未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。 | A | 可作主要证据 |
| 2026-08-26 | MeritForum 纪要 (整理稿) | 标题写'北美软件渠道商离职专家'，但与 08-26 Merit 音频为同一场，转写中专家自述负责北欧区域的实施/咨询商；身份局限见转写条 | B | 仅作局部线索 |
| 2026-08-27 | MeritForum 纪要 (整理稿) | 标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。 | A | 需交叉验证 |
| 2026-08-31 | Merit 路演 (音频转写) | 九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。 | A | 需交叉验证 |
| 2026-09-01 | Funda (Substack) | 作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。 | A | 需交叉验证 |
| 2026-09-02 | TMTB (Substack) | TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。 | A | 可作主要证据 |
| 2026-09-03 | Funda (Substack) | 作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。 | A | 可作主要证据 |
| 2026-09-03 | Acecamp | 未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Databricks 现任员工，负责 GCR/APAC 区域 (给出 GCR/APAC 约八成项目仍在 PoC，了解内部 Omnigent '零号用户' 使用情况与销售策略)。视角局限：竞争对手内部人员；对 Snowflake 是外部观察 ('不太确定 Snowflake 的情况')，尤其是中国市场的 Snowflake；样本为亚太。 | B | 仅作局部线索 |
| 2026-09-07 | Funda (Substack) | 作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。 | B | 仅作局部线索 |
| 2026-09-08 | Merit 专家电话 (音频转写) | 不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher 。文末引出 Zscaler 亚太 CSO Lawrence ，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。 | B | 仅作局部线索 |
| 2026-09-08 | Merit 专家电话 (音频转写) | Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead ，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 ；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。 | B | 仅作局部线索 |
| 2026-09-08 | Merit 专家电话 (音频转写) | Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'、曾任大数据产品经理 ，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 ；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 ；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。 | B | 仅作局部线索 |
| 2026-09-08 | Merit 专家电话 (音频转写) | Snowflake 官方 Data for Breakfast 上海站录音。讲者：(1) Christian，Snowflake 高管 keynote ；(2) 冯金鹏，Snowflake 中国区解决方案顾问 。视角局限：高管愿景演讲 + 售前演示，无任何财务、用量或计费数据；演示企业为虚构/示例场景，转写质量很差 (中文演示部分大量缺字)。 | C | 仅作局部线索 |
| 2026-09-08 | Merit 专家电话 (音频转写) | 未具名专家，纪要无正式简介。内容围绕 Agentforce 渗透率、credit 计费、POC 免费 10-20 万 credit、实施项目经验 ，推测为 Salesforce 内部或 Salesforce 实施伙伴的 Agentforce 交付/解决方案人员。视角局限：Salesforce 视角，对 Snowflake 只是顺带；所称 'Snowflake、Databricks 增速 20% 以上、收入 70 亿与 67 亿' 与公开数据不符 (Snowflake 产品收入增速 30%+)，可信度低。 | C | 仅作局部线索 |
| 2026-09-09 | MeritForum 纪要 (整理稿) | 标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。 | A | 需交叉验证 |
| 2026-09-09 | GuidePoint | Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示、)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased')；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 说话人标签错乱 (DG/PM 互换)。 | A | 需交叉验证 |
| 2026-09-09 | Acecamp | 未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。 | A | 需交叉验证 |
| 2026-09-09 | Merit 专家电话 (音频转写) | 未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 ，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。 | A | 需交叉验证 |
| 2026-09-10 | Merit 专家电话 (音频转写) | 未具名专家，推测为 LVMH 集团内部数字化/IT 相关人员 (纪要多次以'集团'自称并举 LV、Tiffany 门店应用为例，,)，职位年限未披露。视角局限：甲方单一集团，Snowflake 采购为统一 AI infra 口径、金额不清；对 Snowflake 无具体用量数据。 | C | 仅作局部线索 |
| 2026-09-18 | TMTB (Substack) | TMTB Podcast Intel (由 Ethmos 研究 agent 生成的播客摘要，明示)。相关内容为 Databricks CEO Ali Ghodsi 在 Brian Halligan 播客 'Long Strange Trip' 中的说法，以及 Dan Niles (Niles Investment Management 创始人) 的观点。视角局限：只有一句话摘要，正文是图片未读到；Ghodsi 是直接竞争对手，有利益偏向。 | C | 仅作局部线索 |
| 2026-09-24 | Acecamp | 未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。 | A | 需交叉验证 |
| 2026-10-09 | GuidePoint | Frank Hernandez，前 Qlik 美洲区专业服务 Practice 与解决方案负责人 (约 2018 年至 2025 年底，明示/)；主持人确认其 2026 年起加入 Snowflake 并任职至今 ；25 年以上技术经验，约 20 年数据相关 。视角局限：现任 Snowflake 员工，对 Snowflake 的正面评价有明显利益偏向 (如 'every day I wake up and I'm super excited')；访谈主题是 Qlik，Snowflake 内容是侧面。 | B | 仅作局部线索 |
| 2026-10-13 | Acecamp | 未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。 | B | 需交叉验证 |
KB/ 开头的路径相对于你的 Knowledge Base 根目录；web/ 开头的是从网上取的业绩会转写。
A 2025-12-03 · 业绩会转录 (网上公开转写) · Q3 FY26 业绩会：AI 收入 run-rate 达 $1 亿、Anthropic $2 亿采购承诺、迁移只完成 15-20% 22 条
文件：web/SNOW_2025-12-03_Q3-FY26_call.txt
发言人：Sridhar Ramaswamy (CEO)、Brian Robins (CFO，本次为首场业绩会)，以及 13 位卖方分析师；身份为转写稿明示。局限：管理层口径，偏乐观；分析师问题只记首句
阅读范围：非全文：WebFetch 拒绝逐字输出全文 (单段引用上限约 125 字符)，按主题抽取约 100 句逐字短句 (L5-L106)，其中 12 句已用第二来源 (platformaeronaut) 交叉核对 可信度：可作主要证据 (公司 IR corrected transcript，关键句双源核对)；但缺上下文，需与全文交叉验证
产品收入 +29%，NRR 稳在 125%，RPO 加速到 +37%；AI 收入 run-rate $1 亿 (提前一季)，占约 $46 亿年产品收入仅约 2%，管理层把 AI 定位为拉动核心数仓的'驱动因素'而非独立收入。Anthropic $2 亿是 Snowflake 的买方承诺 (成本端)，迁移被称只完成 15-20%，季度 beat 2.5% 低于 Q2，管理层用迁移不均和 hyperscaler 宕机解释。
问题：非全文：只有按主题抽取的逐字短句，句与句之间不连续；行号是本摘录文件行号；分析师问题后的管理层回答由抽取模型配对，个别配对 (如 Hedberg 问 OLTP/OLAP 后接 'AI revenue is predominantly driven…') 可能不是直接回答；G2000 776 与新闻稿口径可能不一致 (新闻稿通常用 Forbes G2000)；Brian Robins 姓名在 roic 版本拼为 Robbins
C 2026-01-20 · MeritForum 纪要 (整理稿) · (PLTR 主题) 渠道商专家：数据平台按消费+固定/用户组件计价，一旦全押某平台每年涨价 10% 也难切换 2 条
文件：KB/MeritForum/2026-01/2026-01-20-qu-dao-shang-zhuan-jia-zeng-chang-momentumliang-hao-jing-zheng-jia-ju.md
发言人：自述'We are a global systems integrator with approximately 300,000 employees'，在其 AI 业务中做 Palantir 渠道/实施 (Palantir 占其 AI 业务 <5%)，与 Microsoft/Databricks/Snowflake 等都有合作；面向企业客户，熟悉欧洲公共部门。主题是 Palantir，Snowflake 只是顺带。视角局限：SI 视角、看重'锁定'叙事；无 Snowflake 具体数字。
阅读范围：全文读过，只记 Snowflake/Databricks/数据平台相关段 (L36-40, L175-250, L286-310, L345-360) 可信度：仅作局部线索
Snowflake 只在数据平台对比中出现：数据侧 Snowflake/Databricks/超大云/Palantir 解决的问题大致相同，Snowflake 和 Databricks 在数仓和大规模数据集成更强；所有平台都在加 agent。定价普遍是消费+固定或用户组件，折扣取决于承诺量和年限；客户全押一个平台后，每年涨价 10% 也不足以触发切换。
问题：主题为 Palantir，Snowflake 信息很少
B 2026-01-28 · Acecamp · GCP 数据仓储需求与竞争：OpenAI 从 BigQuery 转到 Snowflake，Snowflake 客户向 BigQuery/Databricks 迁移趋势 4 条
文件：KB/Acecamp/2026-01/2026-01-28-gcpshu-ju-cang-chu-xu-qiu-he-jing-zheng-zhan-wang-agentic-aidui-cpuxu-qiu-ying-xiang-pan-duan-aigcda-ke-gpuzhuan-tpuqu-s.md
发言人：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Google/GCP 现任或前任员工，熟悉 BigQuery、BigTable、Gemini 语音、TPU 客户 (称 BigQuery '招聘的门槛也是最高的')，可能是工程或产品岗。视角局限：GCP 立场，倾向强调 BigQuery 优势；对 Snowflake 是竞争对手视角；SNOW 相关内容只占一部分。
阅读范围：全文 L1-L120；SNOW 相关在 L24-70 可信度：需交叉验证
专家称 OpenAI 以前把 GPT-4 日志和 AB 测试放在 BigQuery，后来出于性能考虑转到 Snowflake；但整体上用户从 Snowflake 迁到 BigQuery 或 Databricks '是挺明确的趋势'，尤其是数据全在单一云上的公司，不过市场整体净增长、不是零和。Snowflake 跨云 join 和分析优势明显；BigQuery 在 Iceberg 上落后 Databricks。未来两年看好 BigQuery 和 Snowflake 的份额潜力。
问题：无专家简介，身份为推测；L69 'Database' 可能为 Databricks 转录错误；'BigTable 的增速预计比 Snowflake 更快' 可能指 BigQuery；L63 关于 Observe 的问题专家未正面回答
A 2026-02-23 · MeritForum 纪要 (整理稿) · (加速前) SNOW 4Q25 调研：样本消费增速由 Q3 32% 放缓到 Q4 27-29%；CoCo GA、Iceberg 成默认格式 16 条
文件：KB/MeritForum/2026-02/2026-02-23-snow-4q25ye-ji-diao-yan-ying-wen.md
发言人：正文无身份介绍。专家多处以'我们'代指 Snowflake ('Iceberg已成为我们的主要开放表格式'，'我们在一个季度内达成了1亿美元目标'，'我们与OpenAI和Anthropic的协议已到位')，又被问'基于您的样本'，推测为 Snowflake 现任或前任员工 (销售/解决方案方向) 或深度绑定的合作伙伴 (推测，依据为措辞)。视角局限：(1) 身份不明；(2) 所给收入/AI 收入数字 (CoCo Q4 1.14 亿美元等) 与公司实际明显不符，疑似把 $100M AI 收入 run-rate 与单产品混淆；(3) 有利益偏向，几乎全部为正面表述；(4) 对增速/绝对值概念混淆 (L285)。
阅读范围：全文 L1-L287 读过 可信度：仅作局部线索
加速前 (Q4 FY26 财报前) 的专家：其样本消费增速由 Q3 32% 放缓到 Q4 27-29%，归因于年末预算和假期季节性，但未履约义务在增长。CoCo 刚 GA，专家称其用同一份 credit 计费、受 RBAC 管控是卖点，也被用于性能调优/健康检查。Iceberg 已成默认格式，部分客户把数据放 S3 只在 Snowflake 上算，导致存储收入减少但扩大可触达计算；Snowpark Connect 直接读 Delta Lake，与 Databricks 比成本便宜 2-5 倍。内存涨价未传导到客户价格。
问题：CoCo Q4 收入 1.14 亿美元、占收入 2-3% 与 CoCo 刚 GA 不符，疑似把公司'AI 收入 $100M run-rate 提前达成'与单产品混淆，推断过度；收入预测数字 (Q1 12.5-13.2 亿美元) 与增速概念混淆，L285 '84 亿 vs 63 亿'示例混乱；身份不明且以'我们'代指 Snowflake，利益偏向；样本增速 27-29% vs 次日 02-24 专家给的 Q4 38%、Q1 40%，同期专家口径差距极大
A 2026-02-24 · MeritForum 纪要 (整理稿) · (加速前) SNOW 4Q25 调研 II：合作伙伴样本 Q4 消费 38%、Q1 有望 40%；Snowflake Intelligence 贡献 <5% 消费，Iceberg 约 5% 16 条
文件：KB/MeritForum/2026-02/2026-02-24-snow-4q25ye-ji-diao-yan-ii.md
发言人：正文无身份介绍。自述'我们与 Snowflake 客户的大部分工作涉及提供数据工程资源'，'任何让 Snowflake 不仅仅是一个云数据仓库的东西，都是我们通过咨询实现货币化的潜在收入来源'，'在我们作为合作伙伴或实施合作伙伴的圈子里'——即 Snowflake 咨询/实施合作伙伴 (明示角色、公司不明)，此前季度也接受过同一系列访谈 ('上次你提到第三季度…37%')。视角局限：(1) 增速是其项目样本，37-40% 远高于公司同期产品收入增速 (Q4 30%)；(2) 自称看不到 CoCo 具体数据；(3) 咨询商从 Snowflake 生态扩张中获益，偏乐观，但对定价/锁定有批评视角。
阅读范围：全文 L1-L231 读过 可信度：需交叉验证
加速前的合作伙伴：其样本 Q4 消费 38%、Q1 估 40% (可能 45%+)，AI 是驱动但 Snowflake Intelligence 及 Cortex 只占消费 <5%，年底也只到 6-7%；CoCo '几乎看不见'。Snowpark 约占 20%、增约 45%；Iceberg 增 40-45% 但只占约 5%。Snowflake 未涨列价，客户越来越多请合作伙伴优化存储分层和运行时；专家判断一旦客户锁定深，Snowflake 可能提价。OpenAI/Anthropic 的 agent 指向 Snowflake 数据是消费驱动，但最坏情景是它们取代智能层，让 Snowflake 退化为数据存储。
问题：样本增速 (Q4 38%、Q1 40%) 比公司产品收入增速 (Q4 30%、Q1 34%) 高 4-8pp，与前一日 02-23 专家 (Q4 27-29%) 方向相反；对 CoCo 无数据，'Snowflake Intelligence <5%'口径为其项目客户
B 2026-02-24 · Acecamp · 数据重力：AI 时代独立数据平台的生与死 (Databricks vs Snowflake vs Fabric 对比文章) 7 条
文件：KB/Acecamp/2026-02/2026-02-24-shu-ju-zhong-li-aishi-dai-du-li-shu-ju-ping-tai-de-sheng-yu-si.md
发言人：Acecamp 平台研究文章，未署名作者，不是专家访谈 (文中无专家简介)。数据引自 ETR 调研、SaaStr、Databricks/微软官方披露。视角局限：二手研究汇总，写于 SNOW 加速之前 (2026-02)，NRR 和 AI 收入数字可能已过时；部分数字来源不明。
阅读范围：全文 L1-L153 (L43、L116 为图片未读) 可信度：需交叉验证
文章认为同为约 $5B ARR，Databricks 增速 55% vs Snowflake 29%，NRR >140% vs 约 125% 且 Snowflake 趋势下滑，AI 收入 run-rate $1B+ vs 约 $100M (10 倍差距)，Databricks SQL 数仓收入 run-rate 已破 $1B 正进攻 Snowflake 腹地；Fabric 以捆绑定价压缩独立数据平台定价权。支持面：ETR 显示部署 Cortex Search 后采用其他组件概率升约 350%，SNOW 核心数仓支出意向反弹。
问题：写于 2026-02，早于 SNOW 加速，NRR/AI 收入可能已过时；Snowflake 的 '$5B ARR' 和 29% 增速口径未说明 (产品收入 run-rate?)；无专家身份，二手汇总
A 2026-02-25 · 业绩会转录 (网上公开转写) · Q4 FY26 业绩会：FY27 指引 +27%，CoCo GA、Intelligence 2,500 账户，AI 产品毛利低于核心 25 条
文件：web/SNOW_2026-02-25_Q4-FY26_call.txt
发言人：Sridhar Ramaswamy (CEO)、Brian Robins (CFO)、Christian Kleinerman (产品 EVP)，及约 10 位卖方分析师；转写稿明示。局限：管理层口径
阅读范围：非全文：按主题抽取约 110 句逐字短句 (L5-L120)，其中 12 句已用 platformaeronaut 交叉核对 可信度：可作主要证据 (公司 IR corrected transcript，关键句双源核对)；缺上下文需交叉验证
FY26 产品收入 +29%，Q4 +30%，RPO 加速至 +42% (含史上最大 $4 亿+ 单)。FY27 首次指引 $56.6 亿 (+27%，含 Observe 约 1pt)，毛利率指引 75%。CFO 承认 AI 产品毛利率低于核心，靠核心效率抵消；分析师问 AI ARR 时未回答。SBC 占收入 41%→34%→27%E。
问题：非全文：逐字短句摘录，缺上下文；'approximately $600 million' 一句在摘录中紧接 Observe 语境，但可能指其他收购，需对照全文；Q&A 中 Kirk Materne/Raimo Lenschow 由同事代问，按转写稿记名
B 2026-02-26 · MeritForum 纪要 (整理稿) · (Merit 研究员点评) SNOW 4Q25：收入 +30% 小幅 beat，FY27 指引 27% (有机 26%)，RPO 42%，裁员+AI 提效推 margin 8 条
文件：KB/MeritForum/2026-02/2026-02-26-snow-4q25ye-ji-dian-ping.md
发言人：作者 Mo Zhou (Merit 研究员，卖方/研究平台视角)，引用 Q4 FY26 业绩会管理层原话 (未注明具体发言人，推测为 CEO Sridhar Ramaswamy/CFO)。注意文中'2026年/2027年'为日历年口径，大致对应 FY27/FY28。视角局限：研究员观点+管理层摘录，非一手客户数据。
阅读范围：全文 L1-L62 读过 可信度：可作主要证据
Q4 FY26 收入 +30%、beat 2.4% (近期最小)，FY27 指引 +27% 其中有机 +26% (即约 1pp 来自并购)；RPO 增速 38%→42%。管理层承认新 AI 产品利润率低于核心业务，靠核心效率 (约 200 人裁员、Q4 净增员工仅 37 人) 抵消。研究员认为 SNOW 与 NET 一样是少数 AI 带来加速且无 margin 负面的软件股，问题在估值。
问题：'2026年/2027年'为日历年口径 (≈FY27/FY28)，与公司财年标注不同；管理层原话为中文转译，未注明发言人
X 2026-02-28 · MeritForum 纪要 (整理稿) · (空稿) 4Q25 SNOW 季度调研：只有 8 个议题提纲，无正文 0 条
文件：KB/MeritForum/2026-02/2026-02-28-4q25-snowji-du-diao-yan.md
发言人：作者 Mo Zhou (Merit)；无专家信息
阅读范围：全文 L1-L32 读过 (只有议题列表) 可信度：仅作局部线索
只有议题提纲 (销售完成度、新产品、存储涨价、裁员、AI Lab 客户、AI loser 风险、OpenAI/Anthropic 企业方案、季度看法)，没有任何内容，正文可能在 docx 附件里未转出。
问题：正文缺失，只有提纲；报告链接指向 docx，需另行获取
B 2026-04-28 · GuidePoint · 10 人 AI 初创 CTO 选型：Snowflake 输给 Databricks，因 AI/ML 工作流深度不足；预计月成本 Snowflake 1-2.5 万美元 vs Databricks 2.5-3 万美元 5 条
文件：KB/GuidePoint/2026-04/2026-04-28-data-management-and-ml-platforms-databricks-cto-highlights-platform-selection-process-lakehouse-scalability-and-long-ter.md
发言人：Sahitya Sehgal，Untrap Inc 联合创始人兼 CTO (现任约 2 年，明示 L47/L164-165)；公司约 10 人，为 MSP 行业做 AI 引擎 SaaS；此前在一家大型一级银行财富管理部门任 MD 兼数据与数字 VP 约 4-5 年 (明示 L166-167)。GuidePoint 视角 Customer。视角局限：极小型初创、单一选型样本、工作负载很小 (月成本万美元级)；在银行时已用过 Databricks，存在熟悉度偏向 (L245)；成本数字是方向性估计而非报价 (L321-323)；访谈问卷后半段问 Dataiku，与 Snowflake 无关。
阅读范围：全文 L1-L621 (Dataiku 部分 L446-L607 仅浏览，与 Snowflake 无关) 可信度：仅作局部线索
小型 AI 初创选型中 Snowflake 被评为'优秀的数仓'(7.5/10)，但因端到端 AI/ML、agentic analytics 灵活性不如 Databricks (8/10) 而落选——是 AI 原生新负载流向 Databricks 的一个单点例子。价格上两家都是最贵档，Snowflake 估计每月 1-2.5 万美元，略低于 Databricks 2.5-3 万美元，即并非价格因素输单。
问题：单一小型初创样本，工作负载规模对 Snowflake 收入无代表性；成本为方向性估计，非报价；专家在前东家已用过 Databricks，有熟悉度偏向
B 2026-04-29 · GuidePoint · Tavant 数据平台 VP 选型：Snowflake 每 TB 处理成本比 Databricks 贵约 40%，未管控查询会带来'惊吓账单'；最终选 AWS+Databricks 5 条
文件：KB/GuidePoint/2026-04/2026-04-29-data-management-and-ml-platforms-datarobot-customer-explores-multi-cloud-strategy-with-aws-poc-led-evaluation-and-cost-e.md
发言人：Prasun Kumar Mishra，Tavant Technologies 数据平台与 EBiz 副总裁 (现任，明示 L45)，领导约 3000 人公司的 AI 与数据部门 (L173-174)；公司为解决方案咨询公司，做大型数据平台和 agentic AI 落地。GuidePoint 视角 Customer，实际是代客户选型的 SI。视角局限：一次为期 2 周的客户选型；价格数字口径混乱 (L303 说每 PB $30-40，L313 改为每 TB $25 vs $35-40)；最终选型受 AWS 三年承诺折扣驱动；访谈主体是 DataRobot，Snowflake 只在选型对比段落出现 (L250-L391)。
阅读范围：全文 L1-L562；Snowflake 相关段落集中在 L250-L391，其余为 DataRobot 使用细节 可信度：仅作局部线索
SI 选型给出的价格感受：Snowflake 每 TB 处理约 $35-40，比 Databricks $25 贵约 40%，但免运维；Snowflake 评分 8/10，但'未管控的查询会吃掉大量钱，出现惊吓账单'。最终选 AWS+Databricks，原因是大量非结构化数据预处理和客户愿意给 AWS 3 年承诺换折扣——是云厂商承诺折扣+Databricks 组合压制 Snowflake 的例子。
问题：价格单位前后不一：L303 说每 PB $30-40，L313 说每 TB；且 '$22, let us say $25' 为随口估计；L374 'Databricks or Snowflake, sorry, AWS' 为口误；主题是 DataRobot，Snowflake 只是选型对比的一环
B 2026-04-29 · GuidePoint · 前 AtScale 区域销售总监：独立语义层会被 Snowflake/Databricks 吞并；Snowflake 最大 BI 负载来自 Excel/Power BI，AtScale 能把查询和计算'几乎无摩擦地'导入 Snowflake 7 条
文件：KB/GuidePoint/2026-04/2026-04-29-atscale-durability-of-a-universal-semantic-layer-in-an-ai-driven-data-stack.md
发言人：Dionys Gragousian，前 AtScale 区域销售总监 (EMEA，约 2023 年 5 月入职，在职约 20 个月，明示 L40/L146、L534-535)；此前在 SAS、DataRobot 做数据分析与 AI 销售，自 2016 年起 (L150-152)。GuidePoint 视角 Former Executive。视角局限：做销售不碰技术 (自称 'limited technical knowledge' L463)；离职已有一段时间，说法可能过时；主要在 EMEA；对 Snowflake 的认识来自与其现场销售合作；Snowflake 投资 AtScale 的动机属推测。
阅读范围：全文 L1-L744 (Snowflake 相关段落为主) 可信度：仅作局部线索
对 thesis 最相关的两点：(1) Snowflake 最大 BI 负载来自 Excel 和 Power BI，语义层把这些查询和计算导入 Snowflake，说明 BI 仍是 Snowflake 消费底盘；(2) 在 AI/agent 时代语义/context 层变得更核心，专家认为离数据最近的数仓会赢，但'押 Databricks'，因其 AI/ML 出身。另：Snowflake/Databricks 之间切换'极少'且'非常困难'，支持粘性；但语义层/AtScale 的智能聚合本身用于降低扫描量和计算账单。
问题：专家为销售，技术细节有限 (L463)；Snowflake 投资/收购 AtScale 为推测；'Snowflake views' 等产品名称不准确 (L191)
A 2026-05-08 · GuidePoint · Thoughtworks 数据与 AI 负责人渠道调研：Q1 需求超预期，但 Databricks 势头强、云厂商用数百万美元补贴把客户迁出 Snowflake 16 条
文件：KB/GuidePoint/2026-05/2026-05-08-snowflake-1q26-ai-and-data-partner-channel-check.md
发言人：Bilal Jaffery，Thoughtworks 美洲区高级副总裁兼 Data & AI 负责人 (现任，明示，L30/L127)；此前在 Blend360，更早为 Accenture AI 与 experience platform practice 负责人 (主持人介绍 L127-128)。GuidePoint 视角标注为 Reseller/Franchisee (L24)。视角局限：咨询/SI 渠道商，看到的是自己咨询业务的增速 (high 30s) 而非 Snowflake 消费；客户多数同时用 Snowflake 和 Databricks；自称 'Snowflake-biased' (L239)；云厂商补贴是一个案例 (AWS) 加泛化判断；咨询公司对 '迁移项目' 有利益 (迁出迁入都能收费)；没有给出任何 Snowflake 账单数字。
阅读范围：全文 L1-L590 可信度：需交叉验证
渠道需求强 (咨询业务指引从 high 20s 上调到 high 30s)，AI 落地在拉动数据现代化，这支持 AI 拉动核心数据平台的 thesis；但专家明确说 '最近是 Databricks 的秀'，且最担心云厂商：迁移补贴从 10-20 万美元升到 '数百万美元'，三家云都在做，并已有 AWS 迁出 Snowflake 的案例。另一个反方信号：50%+ 项目包含降 token/降算力步骤，粗糙原型可降 30-60% 算力。
问题：专家给出的 high 30s 增速是咨询业务口径，不能直接映射 Snowflake 产品收入；AWS 迁出案例只有一个且细节含糊 (L294-301 转录不完整)；GuidePoint 将其标为 Reseller/Franchisee，但实为 SI/咨询，与迁移双向利益相关
C 2026-05-19 · GuidePoint · 前 Dremio 销售谈 SAP 收购 Dremio：开放架构 (Iceberg/Delta) 成为主流，SAP 用零拷贝查询抢 Snowflake/Databricks 在 SAP 客户中的外部数据份额 2 条
文件：KB/GuidePoint/2026-05/2026-05-19-sap-dremio-acquisition-can-it-help-sap-win-the-enterprise-ai-data-stack.md
发言人：Harsh Singla，标题栏写 'Former Regional Sales Director - Dremio' (L37)，但主持人介绍为前 Dremio APAC 企业客户经理 (约 2 年多)，现任 Confluent 区域销售总监约 3 年 (L161-167)；更早在 Qubole。即离开 Dremio 约 3 年。视角局限：APAC 销售，离开 Dremio 已久；对 Snowflake 只有泛泛比较；Dremio 立场天然推开放湖仓、反数仓。
阅读范围：仅检索 Snowflake/Iceberg/Databricks 关键词段落 (L57、L139、L211-L230、L325-L345、L378-L445) 可信度：仅作局部线索
对 SNOW 的直接内容很少：专家认为市场普遍转向开放架构，Snowflake '仍是专有但已一定程度开放'；SAP 可借 Dremio 把 SAP 客户 30-40% 的非 SAP 数据纳入 SAP 生态，'挤掉 CSP 原生或 Snowflake/Databricks'——是 Iceberg/零拷贝引擎分流计算的一个潜在渠道，但为销售视角推测。
问题：标题栏职位 (Regional Sales Director - Dremio) 与主持人介绍 (Dremio APAC AE，现 Confluent RSD) 不一致；对 Snowflake 的判断泛泛，无一手案例
A 2026-05-20 · GuidePoint · Launch Consulting 数据与 AI VP 渠道调研 (只记 Snowflake)：Snowflake 业务 Q1 增 24%，Cortex Code 让交付效率提升约 40%，既增加消费也帮客户省钱 15 条
文件：KB/GuidePoint/2026-05/2026-05-20-snowflake-and-mongodb-may-2026-channel-check-with-a-vp-of-data-and-ai.md
发言人：Felix Chen，Launch Consulting Group 数据与 AI 副总裁 (现任，明示 L36/L124)；GuidePoint 视角标注为 Industry Consultant (L30)。视角局限：SI/实施合作伙伴，增速是其 Snowflake 实施业务 (项目收入) 而非 Snowflake 消费；项目大且少 (约 8 个 pipeline，单个约 3000 万美元)，样本块状；作为 Snowflake 合作伙伴对 Cortex Code 明显热情 (自称 'sing the praises')；未给出任何客户 credit 账单数字。本文只记录 Snowflake 部分，MongoDB 部分略过。
阅读范围：全文 L1-L575 (只记录 Snowflake 相关内容) 可信度：需交叉验证
渠道商 Snowflake 业务 Q4 增 30%、Q1 增 24%，全年目标 25%，主要受工程师产能限制；Cortex Code 被称为'数据与分析领域唯一好用的编码 agent'，交付效率提升约 40%，并通过新建 view/语义层拉动消费——支持第 (1) 层。但专家也直说 Cortex Code 帮客户找出最贵的 pipeline 并优化，'在吸走本该流向 Snowflake 的收入'；Fabric 在 RFP 中做正面对比。
问题：24%/30% 是 SI 实施业务增速，不是 Snowflake 收入/消费；40% 效率与 '100% 效率' 在同一段内口径不同 (L406 vs L420)；专家为 Snowflake 合作伙伴，对 Cortex Code 评价有明显正向偏向
B 2026-05-20 · TMTB (Substack) · TMTB 早报 (FY27Q1 前)：Benchmark 上调 SNOW 目标价到 $200；KeyBanc 调查显示前瞻支出意愿转弱 4 条
文件：KB/Substack/2026-05/2026-05-20-tmtb-morning-wrap.md
发言人：TMT Breakout (发件人明示)，转述卖方 Benchmark (Buy) 和 KeyBanc (OW) 的报告。视角局限：二手转述，卖方原报告未见。
阅读范围：仅检索 SNOW/Cortex/Databricks 等关键词段落 (L180-196) 可信度：仅作局部线索
FY27Q1 前卖方分歧：Benchmark 目标价 $190→$200，认为 SNOW 能在 FY27 维持约 30% 产品收入增速；KeyBanc 调查偏混合：4 月消费健康，但前瞻支出意愿明显恶化、伙伴 NTM 增长预期下降、F2Q 指引风险 (约 24% vs 街 26%)、CoCo 采用数据弱于管理层说法，维持 OW 和 $200 目标价。
问题：contrarian 的 red 来自 L194，与 quote 所在 L192 同一段落；事后看 KeyBanc 的 F2Q 指引风险判断错误 (实际 Q2 指引隐含约 30%)
A 2026-05-25 · Funda (Substack) · Funda AI 1Q26 (FY27Q1) 前瞻：访谈 6 家 Snowflake 销售伙伴，CoCo 是史上最快增长产品 25 条
文件：KB/Substack/2026-05/2026-05-25-preview-snow-1q26-cortex-code-drives-accelerating-growth.md
发言人：作者 FundaAI (发件人 fundaai@substack.com，明示)。一手材料来自 6 位 Snowflake 销售/实施伙伴专家：北美 Partner A-E + 欧洲 Partner (欧洲伙伴明示来自 Cognizant，L165)，均未具名，其余所属机构不详。渠道方法：伙伴按 Snowflake 给的 Consumption 指标考核 (与收入高度相关)，部分无 consumption 数据者用咨询/实施收入 (practice revenue) 代替 (L23-25)。注意 Funda 的 1Q26 = 日历 2026Q1 = Snowflake FY27Q1。视角局限：伙伴只看自己服务的客户，口径各不相同 (consumption vs 咨询收入)；伙伴有推销 Snowflake 新产品的利益偏向；部分数字 (如 CoCo 占 consumption 5-10%) 明显高于公司口径，样本偏向早期采用者。
阅读范围：全文 L1-L689 (L681-689 为图片，未读到图表内容) 可信度：需交叉验证
Funda 通过 6 家伙伴的 consumption 达成率追踪 SNOW：多数伙伴 1Q26 consumption 增速 25-30% (欧洲伙伴 39%)，并预计 2Q26 进一步加速，主因 CoCo 从试用转入计费 (多数账户 4 月 1 日开始计费)。同时有不利信号：CoCo 比 Claude Code 贵约 33%、单查询成本高达 $5,000 引发客户不满、大型金融客户优化拖累、CoCo 本身直接消费不高而是间接驱动、Snowpark 在 greenfield AI/ML 招标中输给 Databricks。Funda 预测季度增速约 31%，CoCo 2Q26 贡献约 2% 增量。
问题：Funda 季度命名用日历年 (1Q26 = FY27Q1)，FY2026E 指日历 2026；各伙伴口径不一：consumption、practice revenue、自身客户收入占比混用；CoCo 占比 3% / 5% / 5-10% 互相不一致且远高于公司层面；伙伴 B 称 CoCo 是 Claude Code 的像素级复制并抢占 GPU，伙伴 C 称 CoCo 不跑在 Bedrock 上而在 Snowflake 内部运行，伙伴 B 又建议 Snowflake 应自托管而非走 Bedrock，说法互相矛盾；L681-689 图表无法读取
A 2026-05-27 · FactSet 业绩会转录 · Snowflake Q1 FY27 业绩会：增速 30%→34%，CoCo 驱动指引上调到 +31%，承认 AI 毛利率更低 27 条
文件：KB/FactSet/2026-05/2026-05-28-corrected-transcript-snowflake-inc-snow-us-q1-2027-earnings-call-27-may-2026-5-00-pm-et.md
发言人：CEO Sridhar Ramaswamy、CFO Brian Robins、EVP Product Christian Kleinerman、IR Katherine McCracken (转录稿明示)。视角局限：管理层自述，倾向正面叙事；所有 AI 拉动核心的说法都没有量化
阅读范围：全文 L1-L376 可信度：可作主要证据
Q1 产品收入 +34%、NRR 126%，FY27 指引从 +27% 上调到 +31%，CFO 说 CoCo 是指引上调的最大驱动。管理层承认 AI 产品毛利率低于核心，75% GM 靠 AWS 合同省带宽成本维持；CEO 承认大规模铺开时 token 成本治理是问题，迁移周期从 2 年缩到 1–2 季度。加速没有量化拆分。
问题：$1M+ 客户数转录为 79，应为约 779/780 (Q2 为 828=780+48)，疑转录漏字；FY27 指引从 27% 上调到 31% (L42-43)，不是 36%；36% 是 Q2 时的上调；'CoCo's double developer productivity as measured by pull requests and lines accord' 疑转录错误 (lines of code)；全程未提 Iceberg、Databricks 竞争、价格、SBC；AI 收入金额和 AI 毛利率未披露；Snowflake Intelligence 账户只说环比翻倍，未给绝对数 (Q2 改名 CoWork 才给 5,800)
A 2026-05-28 · Funda (Substack) · Funda AI 1Q26 (FY27Q1) 业绩点评：CoCo 驱动加速得到确认，产品收入 +34% 11 条
文件：KB/Substack/2026-05/2026-05-28-review-snow-1q26-cortex-code-driving-acceleration-reconfirmed.md
发言人：作者 FundaAI (发件人明示)。内容是 Funda 对 FY27Q1 业绩的点评，引用管理层电话会原话。视角局限：Funda 在前瞻里已看多，点评有自我验证倾向 (多次强调 'our calls validated')；季度命名用日历年。
阅读范围：全文 L1-L157 可信度：需交叉验证
Funda 称 FY27Q1 产品收入 +34%，超预期 5.3% (剔除 Observe 仍超 4%)，为近年最大 beat；FY 指引从 27% 上调到 31%；CoCo 用户超 7,100 个账户，Snowflake Intelligence 账户环比翻倍。管理层说 CoCo 毛利率略低，但用 AWS 合同降低带宽成本，全年产品毛利率维持 75%。Funda 判断 CoCo 两个月达到收入 1%，单季可贡献 2-3% 收入，把 SNOW 和 DDOG 并列为最好的 AI SaaS 标的。
问题：Funda 'FY26 guidance raised from 27% to 31%' 实指 Snowflake FY27 (截至 2027-01)，与任务背景所说 FY27 最终上调到 36% 是不同时点；引用的管理层原话未标注说话人；Funda 自我验证倾向强
B 2026-06-01 · Funda (Substack) · Funda AI 周报：SNOW CoCo 拐点得到验证，并回顾 Funda 历次 SNOW 多空判断 4 条
文件：KB/Substack/2026-06/2026-06-01-weekly-welcome-fabian-onboard-snow-cortex-code-inflection-mrvl-optical-raise-smtc-beat-mdb-panw-next-llm-upg.md
发言人：作者 FundaAI (发件人明示)。周报汇总，SNOW 部分是对 5/25 前瞻和 5/28 点评的复述，加上 Funda 过往 SNOW 判断记录。视角局限：自我宣传性质，无新一手数据。
阅读范围：全文 L1-L176；只记 SNOW 相关段落 (L40-47, L58, L169-174) 可信度：仅作局部线索
复述 FY27Q1 产品收入 +34%、全年指引从 27% 上调到 31%，Funda 把 SNOW 和 DDOG 列为最喜欢的两个 AI SaaS。给出 Funda 历次 SNOW 立场：3Q24 起看多 (新 CEO 改革)、3Q25 看空 (Gen2 优化 + Snowpark Connect + 前沿实验室压预算)、4Q25 看多、1Q26 很看多 (CoCo 吸收 Claude Code 溢出需求)。
问题：无新一手数据，是此前报告的复述
A 2026-06-02 · FactSet 会议转录 · Snowflake 2026 Investor Day (Summit)：宣布 4Q FY28 GAAP 盈利、SBC 路径、CoCo/CoWork 战略，但没有给 FY 收入/FCF 中长期目标 32 条
文件：KB/FactSet/2026-06/2026-06-02-raw-transcript-snowflake-inc-snow-us-investor-day-2-june-2026-4-00-pm-et.md
发言人：IR 负责人 Katherine McCracken 主持；CEO Sridhar Ramaswamy (战略)、产品 EVP Christian Kleinerman (产品发布，台上称 CK)、CFO Brian Robins (入职不到一年，财务框架)；Q&A 分析师：Carl (机构未说明，推测 UBS Karl Keirstead)、Sanjit Singh (Morgan Stanley)、Alex Zukin (Wolfe)、Ittai Kidron (Oppenheimer)、Rob Owens (Piper)、Brad Zelnick (Deutsche Bank)、一位未具名分析师 (问 Databricks 增速)、Mike Cikos (Needham)、Adam Tindle (Raymond James)。均为明示。视角局限：公司自我推介场合，客户案例都是精选正面样本；RAW (未校对) 转录，有多处错字 (cloud code=Claude Code、Christine=Christian 等)。
阅读范围：全文 L1-L740 可信度：可作主要证据
Investor Day 唯一硬目标是 4Q FY28 实现 GAAP 盈利 (只靠 opex 和 SBC 两个杠杆，明确说不是 FY28 收入指引)，SBC/收入 41%→34%→27% 趋势延续；没有给 FY 收入、FCF 或 AI 收入目标，TAM 由 $225B 到 5 年后 $460B。产品端核心是 CoCo 加速迁移和消费、CoWork 仍需证明，Iceberg 继续开放 (Snowflake 托管 Iceberg 存储 GA)，Adaptive Compute 按收入中性定价；管理层承认 AI 产品毛利更低、CoCo 也让客户更易优化成本、多数客户仍只有一个用户在用。
问题：没有给 FY 收入、FCF、FCF 利润率、AI 收入的中长期目标；唯一硬目标是 4Q FY28 GAAP 盈利 (单季，不是全年)，且明确不含收入假设；SBC 41%/34%/27% 的年份对应未明说，按上下文推断为 FY25/FY26/FY27 指引；TAM 口径：Investor Day 说 $225B→$460B (5 年)，Piper 会议 CFO 说 roughly going to $300B in the next five or six years，前后不一致；“10 个月→7 个月”是新客户开始消费时间，Piper 会议的“10-11 个月→6 个月”是迁移时间，两者口径不同；RAW 转录错字多：cloud code=Claude Code，Christine=Christian，saved over $500 疑为漏单位，(00:13:15) 处缺词；Carl 的机构未在原文出现
A 2026-06-03 · Merit 专家电话 (音频转写) · 前 Snowflake 员工谈 Cortex Code：8,000+ 客户、计入 credit、标价贵 40% 但 TCO 低、毛利远低于核心业务 17 条
文件：KB/2026-06-03/260603 - Merit Snowflake产品进展调研Ⅰ(2026-06-.md
发言人：专家A：转写里没有专家自我介绍 (开头只有合规声明)，身份来自同场 MeritForum 摘要稿标题“Snowflake离职专家” (forum/3232)，即前 Snowflake 员工；职位、年份都没给。转写里他把 Snowflake 称作“we” (“we've trained, Snowflake has trained cortex code”；“If you look at Snowflake, we have Cortex”)，熟悉 Snowsight/Notebooks/skills/RBAC 这类产品细节，推测是售前、解决方案架构或产品技术岗，不是财务岗。视角局限：(1) 没有任何一个具体客户的账单或消费数据，讲的基本是产品卖点，和官方营销口径高度一致，有明显亲 Snowflake 倾向；(2) 关于 Claude Code 每代模型都要客户“重新训练”的说法技术上站不住 (把上下文/skills 和模型训练混为一谈)，技术判断要打折；(3) 有事实错误：说 AWS 协议是“6 million dollar” (实际 60 亿美元)，GA 时间一会儿说上季度中旬一会儿说上个月；(4) 已离职，不清楚离开多久，对当前定价/毛利只引用公开说法。
阅读范围：全文 L1-L393 (L13-L173 为平台生成的摘要，L177-L387 为中文整理转译，L393 为中英混杂的原始转写，单行约 60KB，已通读) 可信度：需交叉验证
前员工把 Cortex Code (CoCo) 描述成原生、按 credit 计费、靠拉动计算/存储/token 来贡献收入的工具：8,000+ 客户在用，标价比 Claude Code 贵约 40% 但企业客户 TCO 更低，毛利“远低于”核心业务约 70% 的水平。对 thesis 的关键反面信息是：他自己说 CoCo 写的 SQL 更高效、帮客户省 credit；客户把数据放进 S3 上的 Iceberg 来省 Snowflake 存储费；40% 价差会随微调/小模型缩小；Databricks 追平“只是时间问题”；Snowpark 增长停滞。整篇没有任何量化的消费增速。
问题：转写里没有专家自我介绍，“前 Snowflake 员工”的身份只来自 MeritForum 标题；英文原话把 AWS 协议说成“6 million dollar”，中文整理写 60 亿美元，应为 60 亿美元；GA 时间说法不一致：L183 说上季度中旬，L193 说上个月；毛利率：平台摘要 L111 写“低于约 75%”，专家原话是远低于约 70%；“Cloud Code”是 Claude Code 的转写错误；“Codex Code”是 Cortex Code 的转写错误；Claude Code 每代模型都要客户“重新训练”的说法技术上不准确，属推断过度；这一段反复重复，信息量低；40% 溢价的来源 (按 token？按用户？) 全篇没有说清，是访谈员先提出、专家附和
A 2026-06-03 · Merit 专家电话 (音频转写) · 前头部咨询公司专家：Q1 消费 +30%、Q2 跟踪 >35%，加速主要来自 Cortex AI 而非传统数仓；Databricks 增速 >50% 22 条
文件：KB/2026-06-03/260603 - Merit Snowflake产品进展调研Ⅱ(2026-06-.md
发言人：专家B：转写里没有自我介绍，身份来自同场 MeritForum 摘要稿标题“海外头部咨询公司离职专家” (forum/3233)，即前海外头部咨询公司员工；职位、年份都没给。转写里他说“our snowflake consumption”“migrated onto our platform”“if I look at my portfolio”“most of the customers I speak with”“what I see in customer deployment”，推测他负责咨询公司的数据/AI 业务或一组 Snowflake 客户 (组合视角)，不是 Snowflake 员工。视角局限：(1) 30%/35% 到底是他所在公司/客户组合的消费，还是 Snowflake 公司整体，原话含糊；数值和 Snowflake 公开的产品收入增速 (Q1 FY27 34%) 接近，可能部分在复述公开数字；(2) 产品贡献排序和 15-25% 这类比例来自他自己的客户组合，样本未知；(3) 标为“离职”但用现在时说 Q2 跟踪数据，时间点存疑；(4) 咨询公司同时和 Snowflake、Databricks 合作，没有明显偏向，但对 Databricks 的增速 (>50%) 来自“我看到的最新数据”，像是公开数字。
阅读范围：全文 L1-L376 (L15-L190 为平台摘要，L194-L370 为中文整理转译，L376 为中英原始转写，已通读) 可信度：需交叉验证
这是三篇里唯一给出消费增速的：Q1 同比约 30%，Q2 跟踪略高于 35%；他说 Q2 加速主要来自 Cortex AI (agents/search 约占增量 50%、CoCo 15-25%、Snowpark 10-20%)，核心数仓贡献只有“中等”，而且“核心平台扩张对增长的贡献在下降”。这对“第 (1) 层核心数仓是加速主体”的条件不利，不过他也说 CoCo 直接消耗很小、消耗来自用它建出来的管道/agent/Snowpark (这部分是 Snowflake 自有 credit)。反面信息：Databricks 增速 >50%；Iceberg 让两家读同一份数据；客户对话的主题是怎么少花钱；大部分 agent 仍在试验阶段。
问题：消费增速的主体不明 (他的公司/客户组合还是 Snowflake 整体)，可能在复述公开数字；15-25% 原话是“growth comes from Cortex code”(增长贡献)，同时又说 CoCo 直接消耗很小，篇内矛盾；平台摘要和 MeritForum 都写成“消费占比”；Claude Code 月成本前后不一：L342 说 200-250 美元，L352 说 200-500 美元；Snowpark 贡献与专家A说法矛盾 (A 说增长停滞)；标为“离职”专家，却用现在时报告 Q2 跟踪数据；Genie 定位：前面说 Genie 是对话式 BI，不是 CoCo 的对标，和其他专家不同
C 2026-06-03 · MeritForum 纪要 (整理稿) · 专家A (Snowflake 离职专家) 访谈的论坛摘要稿 0 条
文件：KB/MeritForum/2026-06/2026-06-03-snowflakechi-zhi-zhuan-jia-cortex-codezai-snowflakesheng-tai-zhong-kuai-su-pu-ji-ping-jie-shen-du-ji-cheng-shang-xia-wen.md
发言人：同专家A。本稿标题给出身份“Snowflake离职专家”，作者 Lantian，没有给职位和年份。视角局限同专家A。
阅读范围：全文 L1-L245 可信度：仅作局部线索
内容和 260603 Merit 调研Ⅰ转写基本一一对应 (CoCo 8,000+ 客户、40% 溢价/TCO、毛利低于核心计算、Genie、Snowpark 停滞、Iceberg/S3、CSP 竞合、RBAC)。唯一新增的是给出了专家身份 (Snowflake 离职专家)；没有其他新的数字或判断。
问题：与转写重复，证据都记在转写里；L197“然而，生成式工具的增长正在影响这一价值主张”在转写里找不到对应原话，可能是摘要加工；L223 写 AWS 60 亿美元协议，修正了转写里的“6 million”口误
C 2026-06-03 · MeritForum 纪要 (整理稿) · 专家B (海外头部咨询公司离职专家) 访谈的论坛摘要稿：标题把 CoCo 增量贡献写成“贡献 Q2 消费 15-25%” 1 条
文件：KB/MeritForum/2026-06/2026-06-03-hai-wai-tou-bu-zi-xun-gong-si-chi-zhi-zhuan-jia-snowflake-2025nian-aigong-zuo-fu-zai-qu-dong-xiao-fei-zeng-chang-q1tong.md
发言人：同专家B。本稿标题给出身份“海外头部咨询公司离职专家”，作者 Lantian，没有给职位和年份。视角局限同专家B。
阅读范围：全文 L1-L229 可信度：仅作局部线索
内容和 260603 Merit 调研Ⅱ转写对应。没有新的信息，但标题和要点把 CoCo 的“增长贡献 15-25%”写成了“贡献 Q2 消费 15-25%”，把 Snowpark 的 10-20% 写成“约 20%”，引用时要以转写为准。
问题：L2/L32/L71 把“增长贡献”误写为“消耗占比”，标题有误导性；L176 Snowpark 写“约 20%”，转写为 10-20%；L136 赞助人写“CTO、CDO，有时还有 CIO”，转写为 CDO、CIO、有时 CFO；L220 写“在两年内”，转写为“couple of years”
A 2026-06-04 · Merit 专家电话 (音频转写) · IT 服务商 Snowflake practice 负责人：业务环比约 +10%，CoCo 是间接驱动，4 月推出 AI Credits，Iceberg 约占 10-15%，Databricks AI 规模约 10 倍 17 条
文件：KB/2026-06-04/260604 - Merit Snowflake产品进展调研Ⅲ(2026-06-.md
发言人：专家C：转写里没有自我介绍。MeritForum 摘要稿 (forum/3244) 标为“北美头部软件公司离职专家”，但转写内容和这个标签不符：访谈员问“how did your Snowflake practice perform”，他说 Databricks 是“a sister practice to me... we work together”，“we do a lot of finance and accounting and shared services”“we have insurance clients and there we do a lot of policy administration”，用的是现在时。更像是在一家 IT 服务/咨询/系统集成公司负责 Snowflake 交付 practice (同时有 Databricks practice)，客户集中在金融会计共享服务和保险。职位、年份都没给。视角局限：(1) 他说的“环比 10%”是自家 practice 的业务增速，不等于 Snowflake 的消费增速；(2) Databricks 规模/增速对比也是两个 practice 的对比；(3) 涨价 8-15% 来自“其他案例”的经验，是推测；(4) 服务商从 Snowflake 项目中获益，可能偏乐观。
阅读范围：全文 L1-L316 (L15-L184 为平台摘要，L188-L310 为中文整理转译，L316 为中英原始转写，已通读) 可信度：需交叉验证
服务商视角：CoCo 直接贡献“很小”，作用是让开发者更快建管道、间接拉动计算；Snowflake 4 月 (Q2 初) 推出 AI Credits，把 AI 定价和版本定价拆开，40% 溢价从现有 credit 池里扣，客户心理上容易接受。这支持“第 (1) 层自有 credit 是主体”，但没有给出任何量化拆分。Iceberg 估计约占 consumption 的 10-15%；Databricks 在 AI 上的规模约是 Snowflake 的 10 倍；他推测 Snowflake 可能对传统产品涨价 8-15%，这和专家B“没有涨价、方向是降本”的说法相反。
问题：身份标签矛盾：MeritForum 写“北美头部软件公司离职专家”，转写内容像是在职的服务商 practice 负责人；practice 业务增速 (环比 10%) 不能当作 Snowflake 消费增速；Databricks 10 倍、6-7% 的口径也不清楚；Iceberg 10-15% 是“put a gun to my head”式的估计，而且没说清是消费占比还是用量占比；涨价 8-15% 来自其他厂商的经验，与专家B“没有涨价、方向是降本”相反；“Codex Code”/“Cloud Code”为转写错误
C 2026-06-04 · MeritForum 纪要 (整理稿) · 专家C 访谈的论坛摘要稿 (标为北美头部软件公司离职专家) 0 条
文件：KB/MeritForum/2026-06/2026-06-04-bei-mei-tou-bu-ruan-jian-gong-si-chi-zhi-zhuan-jia-snowflake-q2biao-xian-you-yu-q1-cortex-codeyi-40-jie-ge-yi-jie-ti-she.md
发言人：同专家C。本稿标题标为“北美头部软件公司离职专家”，作者 Lantian；这个标签和转写里“our Snowflake practice / sister practice”的说法不符。视角局限同专家C。
阅读范围：全文 L1-L166 可信度：仅作局部线索
内容和 260604 Merit 调研Ⅲ转写对应 (环比 10%、CoCo 间接驱动、AI Credits 4 月推出、95% 开发者、Databricks AI 10 倍、Iceberg 10-15%、涨价 8-15%、agent 用例)。没有新增信息。
问题：与转写重复，证据都记在转写里；L153 写风险“可能在大约七年之后”，转写原话是“several years out”，疑似听写错误；L115 把 Replit 写成“LabVIEW”，属转写错误；身份标签与转写不一致
B 2026-06-04 · TMTB (Substack) · TMTB 早报 (Summit/Investor Day 后)：Benchmark 目标价上调到 $290，Jefferies 提高对 CoCo 的信心 5 条
文件：KB/Substack/2026-06/2026-06-04-tmtb-morning-wrap.md
发言人：TMT Breakout (发件人明示)，转述 Benchmark (Buy) 与 Jefferies 在 Snowflake Summit 和 Investor Day 之后的报告。视角局限：二手转述。
阅读范围：仅检索 SNOW/Cortex/Databricks 等关键词段落 (L218-238) 可信度：仅作局部线索
Investor Day 后 Benchmark 目标价 $270→$290，称 SNOW 正成为企业 agentic AI 控制面，AI 产品加速迁移、提升消费、缩短 time-to-value；Benchmark 与 Jefferies 都提到 FY28 实现 GAAP 盈利的路径。Jefferies 渠道称 CoCo 自 4 月后可靠性明显改善，在 Snowflake 原生工作流 (text-to-SQL、治理数据) 上优于第三方编码 agent，并指出 SNOW 相对 DDOG 估值仍有明显折价。
问题：Investor Day 具体中期目标数字未在本文出现
B 2026-06-15 · ThirdBridge · 前 Qlik AI 与分析产品副总裁：Snowflake (Cortex/语义 view/Snowflake Intelligence) 和 Databricks 已基本端到端，点工具被挤出；有客户放弃 Qlik AI 改用 Cortex 自带模型 4 条
文件：KB/ThirdBridge/2026-06/2026-06-15-qlik-can-saas-click-under-on-prem-inertia-deconstructing-the-ai-pivot-semantic-gaps-margin-durability.md
发言人：Ian Crosland (ThirdBridge 标题栏显示姓名，正文以 [Expert] 匿名)，前 Qlik Technologies AI 与分析产品管理副总裁 (明示 L15-16)；在 Qlik 担任多个职位，最后负责分析业务产品管理，带约 15 名 PM、数百名研发；现为独立顾问；入行约 30 年，早年做数仓/BI 顾问 (L25)。视角局限：Qlik 视角看 Snowflake/Databricks 是作为合作伙伴兼竞争者；离职时间未明；对 Snowflake 的产品了解是二手 ('I think it's called')。文件内容在 L130 之后重复一遍，以前半部分行号为准。
阅读范围：仅检索 Snowflake/Iceberg/Databricks 关键词段落 (L75、L83-L87；L130 之后为重复内容) 可信度：仅作局部线索
对 thesis 的含义：Snowflake 借 Cortex、语义 view、Snowflake Intelligence 形成'几乎端到端'的分析+AI 栈，挤压 Qlik 等中间层，这支持 Snowflake 抢占 AI 分析层 (E05)。一个中东客户放弃 Qlik 的'黑盒' AI，改用 Snowflake Cortex 作为 AI studio 并自带 LLM——即 Snowflake 获得计算，但模型可能不是 Snowflake 转售的 (E03)。另：Databricks 销售愿意给前端打折，因为'他们只要计算'。
问题：'Snowsight 已被弃用' 与公开信息不符，可能是专家记错或指其中某项功能；文件在 L130 后重复全文，行号取前半部分；专家对 Snowflake 产品名不确定 ('I think it's called')
B 2026-06-29 · FactSet 会议转录 · Snowflake 2026 股东大会：say-on-pay 未通过，多数票选董提案通过 3 条
文件：KB/FactSet/2026-06/2026-06-29-raw-transcript-snowflake-inc-snow-us-annual-general-meeting-29-june-2026-12-00-pm-et.md
发言人：CEO 兼会议主席 Sridhar Ramaswamy 主持，CFO Brian Robins、GC Jeff True 在场；股东提案人 James McRitchie (预录发言)。程序性会议，无问答 (没有收到问题)
阅读范围：全文 L1-L83 可信度：可作主要证据
程序性股东大会，无业务讨论。高管薪酬咨询投票未通过，股东要求董事多数票选举的提案获通过，记录了 3.466 亿股流通股；对 E07 股东经济学/治理是一个负面信号。
问题：RAW 转录 (未校正)，如 'come toward' (应为 come to order)、'fabrications' (应为 revocations)；投票结果为初步结果，最终以 8-K 为准；say-on-pay 未通过的原因未在会上说明
C 2026-07-28 · ThirdBridge · 前 Cloudera 高级产品经理：本地客户把数据管道接入 Snowflake/Databricks 跑 AI 与现代分析，Snowflake 难拿走本地存量负载；客户最担心 Snowflake 账单不可预测 2 条
文件：KB/ThirdBridge/2026-07/2026-07-28-cloudera-silver-linings-private-clouds-balancing-legacy-moats-ai-strategy-churn-realities.md
发言人：Ashish Shah，前 Cloudera 高级产品经理 (2016 年起在 Cloudera 9 年多，先后任销售工程师、ISV 合作伙伴管理 2018-21、产品管理，最后三年负责 SDX 治理组件，明示 L40)；现任 Dataiku Technical Account Manager (L11)。视角局限：Cloudera 本地/混合视角；对 Snowflake 的认识来自竞争丢单和朋友转述 ('I don't have any direct evidence' L152)；离开 Cloudera 时间约 2025-2026。
阅读范围：仅检索 Snowflake/Iceberg/Databricks 关键词段落 (L28、L56-L58、L146-L160) 可信度：仅作局部线索
两点与 SNOW 相关：(1) Cloudera 本地客户的高级/AI 用例流向 Snowflake 和 Databricks，客户建管道把本地数据喂给它们 (增量负载流入，但本地存量迁不走)；(2) 在经济下行时客户最担心 Snowflake/Databricks 账单不可预测。
问题：专家对 Snowflake 的判断为间接观察
B 2026-07-30 · ThirdBridge · 前 Monte Carlo 战略客户经理：Snowflake 销售更愿推 CoCo/自家可观测套件以拉动计算；承诺额消耗不足是消费模式通病，第三方只能用 10-20% 承诺额 5 条
文件：KB/ThirdBridge/2026-07/2026-07-30-monte-carlo-betting-the-house-on-agent-observability-dynamics-of-commercial-technical-scalability.md
发言人：Terry Lucy，前 Monte Carlo Data 战略客户经理 (Strategic Account Executive)，伦敦办公室区域第二号员工，负责企业/战略客户，并与 Snowflake、Databricks、Google、Amazon 等生态伙伴合作 (明示 L38/L42)；任职年限未明 ('a number of years')。视角局限：数据可观测性 ISV 销售视角；对 Snowflake 内部激励和 RPO 的描述是外部观察和推断；样本偏欧洲。
阅读范围：仅检索 Snowflake/Iceberg/Databricks 关键词段落 (L29、L42、L48、L56、L60、L78、L96-L104、L154-L162、L198、L206、L214) 可信度：仅作局部线索
与 SNOW 相关的观察：(1) Snowflake 的 CoCo 和原生监控挤压独立可观测性工具，客户可'几乎免费'开启监控、只付 SQL 检查的计算——平台原生功能把第三方价值转为 Snowflake 计算；(2) Snowflake AE 倾向推 CoCo 和自家可观测套件因为会驱动计算；(3) 专家认为市场普遍存在'承诺额用不完'(承诺 200 万只用 100 万)，marketplace 是消化承诺的手段，客户通常只能把 10-20% 承诺额用于第三方——提示 RPO 与实际消费存在缺口。
问题：RPO 2m/1m 是示意性例子，不是 Snowflake 披露；Snowflake AE 激励为外部推断
A 2026-08-03 · Acecamp · Snowflake 竞对 (Databricks) 调研：AI 产品年增 184%，客户 token 消耗按月翻倍，Databricks vs Snowflake 竞争与价格 11 条
文件：KB/Acecamp/2026-08/2026-08-03-snowflakejing-dui-diao-yan-aichan-pin-nian-zeng-jin-2bei-tokenxiao-hao-an-yue-fan-bei-cong-ke-kou-ke-le-dao-minimaxde-ag.md
发言人：未给出专家简介。从内容推测是 Databricks 内部人员 (用 '公司的核心策略'、'Databricks 内部' 等口吻，能说出客户 AgentBricks 月费用和 MiniMax 是 Databricks 全球第三大 AI 客户)，所服务客户为亚太/大中华区 (太古可口可乐、OOCL、渣打、泰国银行、AXA、FWD、腾讯、台积电、联发科)，推测为 Databricks 亚太区解决方案架构/现场工程或销售岗位，职级和年份不详。视角局限：竞争对手内部人员，对 Snowflake 的技术评价有明显偏向；样本以亚太和数字原生客户为主；价格单位 ('一个 core 约 0.3 元') 口径不清，可能是转录问题。
阅读范围：全文 L1-L166 可信度：需交叉验证
Databricks ML/GenAI 产品同比 +184%；存量客户新增收入数据和 AI 各半 (agent 上线后 ETL 从每周变每日，数据收入翻倍)，新客户 80-90% 收入来自 AI 产品 (如 MiniMax 通过 AI Gateway 调多家模型)；Databricks 转售 Anthropic/OpenAI 模型的价格与直连相同、利润很薄 (对 SNOW token 层毛利的直接参照)。对 Snowflake：客户选 Snowflake 因易用、部署快、与 AWS 深度绑定 (AWS 销售卖 Snowflake 拿同等激励)，存量客户迁出到 Databricks 是重大决策；单价上 Databricks 约为 Snowflake 两倍，但宣称性能带来更低 TCO；Snowflake Horizon/Open Catalog 治理割裂。客户 token 消耗基本按月翻倍。
问题：无专家简介，身份为推测；价格单位 '0.3 元/core' 口径不清，可能是转录或单位错误 (实际 Snowflake credit 约 $2-4)；专家称 Snowflake Horizon Catalog '仅能处理 Iceberg'，属竞争对手说法，需核实；标题称 'Snowflake 竞对调研'，内容几乎全是 Databricks 视角
C 2026-08-04 · Merit 专家电话 (音频转写) · 多类企业工作经历专家：大型企业不会因 RAG 新采购 Snowflake/Databricks，中型科技公司才会评估 2 条
文件：KB/2026-08-04/260804 - Merit 企业AI投入调研V(2026-08-04 15_0.md
发言人：未具名专家，自述曾在泰国一家传统大型企业、一家覆盖英美的中型科技公司 (AI 非唯一业务) 和美国小型初创工作 (明示，L709-L715)，职位未披露；中型科技公司经历中参与过 RAG 早期建设 (L877)。视角局限：样本为个人经历的三家公司；更偏工程/AI 应用，未直接谈 Snowflake 账单。
阅读范围：仅检索关键词段落：L505-L530, L569, L709-L717, L857-L881 可信度：仅作局部线索
反方线索：大型企业已与 AWS/Azure/GCP 签折扣协议，内部 RAG 用量小，'为什么不能继续用 AWS 的数据库' 让新采购 Snowflake/Databricks 很难；中型科技公司在 RAG 放量后才评估新数据库，且不只看 Snowflake/Databricks。
问题：专家经历公司未具名，难以核实
A 2026-08-06 · Merit 专家电话 (音频转写) · 前 Snowflake 员工：AI 推动消费加速、token 转售毛利仅 10-15%、闭源/开源模型 70:30 18 条
文件：KB/2026-08-06/260806 - Merit Snowflake季度调研I(2026-08-06.md
发言人：姓名未披露。转写无开头自我介绍；MeritForum 摘要稿 (forum/3485) 标为'Snowflake离职专家'，即前任 Snowflake 员工，职位与年份未披露。全程以'我们'代称 Snowflake (如'Snowflake Cowork, which is our version of Anthropic Claude Desktop'、'we have acquired')，并提到'在亚太地区我们也与不少政府机构合作'，推测曾在亚太区 (印度/东盟) 做售前/解决方案类岗位。视角局限：(1) 自称只讨论公开信息，所引数字几乎全部是 FY26 (2025 年) 的公司披露数字——Q1 收入 $997M/+26%/NRR 124%、Q2 +32%/NRR 125%、AI 账户 5,200→6,200、Iceberg 1,200 家均对应 FY26 Q1/Q2，并非 FY27 (2026) 本年数据；(2) 立场明显偏向 Snowflake，几乎每个问题都回答为'利好消费'；(3) 亚太视角，对开源模型占比的判断可能高估北美；(4) 多处数字口误 (Q2 收入 $1.909B、'30-35% 指引'、新增客户 5,000 或 2,000)。
阅读范围：全文 L1-L803 (L15-L601 为纪要总结，L603-L797 为中文转译，L803 为中英原文单行)；证据取自中文转译段 可信度：仅作局部线索
前员工把 AI 讲成 Snowflake 消费的放大器：agent 让可用 Snowflake 的员工从 40-50 人扩到 500+，开源模型降价只会带来更多 SQL 查询。但他也给出对 thesis 第 (2) 层不利的关键事实：token 转售约 85-90% 付给模型商，毛利很低；客户在 POC 后因 ROI 不足从 Opus/GPT 转向 Kimi 等开源模型 (token 收入通缩)。所引增速数字是 FY26 旧数，不能当作 FY27 Q2 渠道调研。
问题：关键口径问题：专家所称'今年 Q1 26%/Q2 32%/NRR 124→125%/AI 账户 5,200→6,200/Iceberg FY26 Q2 1,200 家/RPO $6.7B'与 FY26 (2025 年) Q1/Q2 公司披露数一致，不是 FY27 本年数据；访谈日 2026-08-06 早于 Q2 FY27 财报，不可能已有 FY27 Q2 实际数；篇内数字矛盾：Q2 收入说成 $1.909B；新增客户'5,000 名或 2,000 名'；'未来 2 季度指引 30-35%'无出处；立场偏向 Snowflake，几乎所有问题均答为利好消费；agent 消耗倍数被问到时未给数字；MeritForum 版本 NRR 写成 123%，与转写 125% 不符
A 2026-08-06 · Merit 专家电话 (音频转写) · 北欧 SI/经销商：Q4 29%→Q1 30%→Q2 31%，Q3 预期 33-34%，温和加速而非曲棍球杆 16 条
文件：KB/2026-08-06/260806 - Merit Snowflake季度调研III(2026-08-.md
发言人：姓名与职位未披露，无开头自我介绍。从对话推断：现任欧洲 (北欧) 系统集成商兼 Snowflake 合作伙伴与经销商——原话'我们既是 Snowflake 的合作伙伴，也是经销商，但大多数客户会直接向 Snowflake 采购'、'至少在我工作的北欧地区'、'我主要从欧洲视角进行分析'。MeritForum 摘要 (forum/3490) 标为'北美软件渠道商专家'，与本人自述的北欧/欧洲视角矛盾，以转写为准。视角局限：(1) 只看到本公司客户 (传统受监管大企业)，本人承认'我的看法可能存在一定偏向'；(2) 大多数客户直签 Snowflake，看不到签约/backlog；(3) 作为 SI 对 Cortex Code 压缩人力有利益冲突；(4) 增速为'我们的估算'，未说明是收入还是消费口径。
阅读范围：全文 L1-L764 (L15-L568 纪要总结，L570-L758 中文转译，L764 中英原文单行)；证据取自中文转译段 可信度：需交叉验证
渠道观察到的 Snowflake 增速只是温和加速 (29%/30%/31%)，Q3 预期 33-34%，明年 30-45%；AI 是加速来源但'采用速度可能比预期更慢'。Iceberg 被描述为曾经的生存威胁、现在是防御性模块，尚未见大型受监管客户换成自建湖仓；大企业采购最在意消费可见性与折扣。Cortex Code 让开发提速 30-40%，同时压缩 SI 的计费工时。
问题：身份标签矛盾：MeritForum 标'北美软件渠道商'，转写自述北欧/欧洲视角；增速口径未明：是渠道客户支出、Snowflake 公司整体还是估算未说明；'在Q2之前，增速约为28%至30%。这是全年增速，但按季度来看'表述含糊；Q3 预期 33-34% 与公司随后公布的 Q2 FY27 37% 相比偏低，说明渠道样本滞后或低估；第二段 agent 消耗倍数被问到时未给数字
B 2026-08-06 · MeritForum 纪要 (整理稿) · MeritForum 摘要 (前 Snowflake 员工)：与 260806 季度调研I 为同一专家，新增 CoCo 单用户成本上限 $100 等 4 条
文件：KB/MeritForum/2026-08/2026-08-06-snowflakechi-zhi-zhuan-jia-snowflake-2026-cai-nian-di-er-ji-du-shou-ru-tong-bi-zeng-chang-32-zhi-19-yi-mei-yuan-yuan-cha.md
发言人：Snowflake 离职专家 (MeritForum 标题明示'Snowflake离职专家')，与 260806 Merit Snowflake季度调研I 转写为同一场访谈 (问答顺序、数字一致)。职位与年份未披露。局限同该转写：所引增速与账户数为 FY26 旧期数字，立场偏 Snowflake。
阅读范围：全文 L1-L263；重复内容不记，只记转写中找不到的新增表述 可信度：仅作局部线索
摘要稿新增了转写中没有的三点：Cortex Code 有成本控制确保单用户成本不超过 $100；超过 50% 采用 AI 的客户在用 Cortex Code；Cortex Code 下拉菜单有 20+ 模型，客户消耗 $1,000 token 时约 85-90% 付给模型商。标题把增速明确写成'2026 财年第二季度'，印证专家用的是 FY26 旧数。
问题：'单用户成本不超过 $100'、'超过 50% AI 客户用 Cortex Code'、'20+ 模型下拉'在转写中均找不到，可能来自另一录音版本或摘要加工，需核实；NRR 写作 124%/123%，与转写 124%/125% 不一致；迁移成本写成'数十亿美元'，转写为'数百万美元'，属摘要误译
C 2026-08-06 · MeritForum 纪要 (整理稿) · MeritForum 摘要 (北欧 SI/经销商)：与 260806 季度调研III 为同一专家，新增'Cortex Code 使用率约 10%' 1 条
文件：KB/MeritForum/2026-08/2026-08-06-bei-mei-ruan-jian-qu-dao-shang-zhuan-jia-snowflake-di-er-ji-du-xiao-fei-zeng-chang-da-31-yu-ji-di-san-ji-du-wei-33-34-mi.md
发言人：MeritForum 标为'北美软件渠道商专家'，但内容与 260806 Merit Snowflake季度调研III 转写一致 (Q4 29%/Q1 30%/Q2 31%/Q3 33-34%)，专家自述在北欧地区、是 Snowflake 合作伙伴兼经销商，标签应以转写为准。局限同该转写。
阅读范围：全文 L1-L201；重复内容不记，只记新增 可信度：仅作局部线索
与转写基本重复。唯一新增的是'目前 Cortex Code 的使用率大约是 10%'(标题写成'使用率维持在 10%，采用率约 50%')，该句在转写中英文原文均找不到。
问题：身份标签与转写自述 (北欧) 矛盾；'使用率 10%'不见于转写，来源不明
A 2026-08-10 · Merit 专家电话 (音频转写) · 大型制药客户：Snowflake 支出同比 10-15%，Cortex Code 尚未使用 (Cursor 为全球标准)，Databricks 降为遗留平台 14 条
文件：KB/2026-08-10/260810 - Merit Snowflake Q2业绩调研(2026-08-.md
发言人：姓名与职位未披露，无开头自我介绍。从对话推断：现任一家全球大型制药公司的 IT/数据平台负责人级别人员——原话'我们是一家制药公司，处于受监管的环境中'、'几年前，公司就开始制定并推进 cloud first strategy，我也在其中进行了较大推动'、'Snowflake CEO与我们的CDO、CIO也进行过深入交流'，公司是 Snowflake 公开客户案例。视角局限：(1) 单一客户样本，且是 Snowflake 的标杆/忠实客户，偏乐观；(2) 制药行业保守、审批慢，采用新产品节奏不代表科技/数字原生客户；(3) 增速口径 (同比 10-15%、环比 3-5%) 是个人估计、按 PO/里程碑付款，不精确；(4) 中文口语转写，部分术语混乱 (如 'Identical AI')。
阅读范围：全文 L1-L895 (L15-L649 纪要总结，L651-L889 中文转译，L895 原文单行)；证据取自中文转译段 可信度：可作主要证据
一家把 Snowflake 当全球标准数据平台的大型药企，支出同比只增 10-15%，明年增速可能只快 1-2 个点，远低于公司整体 30%+。Cortex Code 尚未使用，原因是 Cursor 已是公司级标准、审批流程长；Databricks 被降级为遗留平台，新项目优先 Snowflake，商业类 AI 应用至少 80% 跑在 Snowflake 上。说明成熟大客户的增长主要来自整合份额，AI 产品的拉动还很早。
问题：单一客户样本，且为 Snowflake 标杆客户 (CEO 曾采访其 CDO/CIO)，不代表整体；增速为个人估计，按 PO/里程碑付款，季度波动大；访谈称 Cortex Code '2 月发布'、'约 3 个月'，与其他专家说法 (CLI 约 6 个月前) 不一致
A 2026-08-12 · Merit 专家电话 (音频转写) · 北美渠道商：Snowflake 消费稳定在 25-27%，Cortex Code 加速要到明年；可观测性难敌 Datadog 14 条
文件：KB/2026-08-12/260812 - Merit Snowflake Q2业绩调研 III(2026.md
发言人：姓名与职位未披露，无开头自我介绍。MeritForum 摘要 (forum/3507) 标为'北美软件渠道商离职专家'；但转写中一直以现在时说'我们进入市场时'、'我们尚未向其中许多客户交付硬件'、'今年年初我们围绕Snowflake的可观测性平台开展了大量市场推广培训'，像是仍在 IT 分销/经销商 (同时卖硬件与 Snowflake、MongoDB) 的人员，现任/前任存疑。视角局限：(1) 只看到经该渠道采购的客户；(2) 对 MongoDB 的发言带有经销商自身目标 ('我们希望...留存率 95-97%')，偏利益方视角；(3) 数字多为'我们估计'，篇内有矛盾 (Cortex Code 尚未带动加速，却占消耗 15-20%)；(4) 录音在 Databricks 竞争问题处截断。
阅读范围：全文 L1-L665 (L15-L475 纪要总结，L477-L660 中文转译，L665 原文单行；录音在 Databricks 选型问题处截断) 可信度：需交叉验证
渠道看到 Snowflake 消费过去三季稳定在 25-27%、年底前维持约 25%，Cortex Code 尚未带来加速，原因是治理合规、今年预算已用完；把翻倍到 40-50% 的希望放在明年。新客户 29-30%、老客户合同规模/期限 +20%+，FDE 让迁移周期缩短 50%。Snowflake 可观测性 (Observe) 今年难敌 Datadog/Dynatrace。
问题：篇内矛盾：称 Cortex Code 尚未带来加速，却估计其占消耗 15-20%；身份存疑：MeritForum 标'离职'，转写口吻像现任经销商；'明年翻倍到 40-50%' 与 '混合模型下用量增长 50-60%' 是愿望性预测，无依据；录音在 Databricks 选型问题处截断；60/30 胜率只见于 MeritForum 摘要；渠道口径 25-27% 与公司 Q1/Q2 FY27 公布的 34%/37% 差距很大
A 2026-08-12 · MeritForum 纪要 (整理稿) · 北美咨询公司前专家：按产品拆分 Snowflake 消费——核心 60-65% 增 20-25%，Cortex AI 10-15% 增 70%+；重度 CoCo 用户增速 45% vs 30% 20 条
文件：KB/MeritForum/2026-08/2026-08-12-bei-mei-zi-xun-gong-si-chi-zhi-zhuan-jia-snowflake-xiao-fei-zeng-su-cong-2026-nian-di-yi-ji-du-de-32-fang-huan-zhi-di-er.md
发言人：MeritForum 标为'北美咨询公司离职专家'，即前任北美咨询公司人员，姓名、职位、年份未披露；本次分配文件中无对应的完整转写 (可能对应未分配的 Q2业绩调研 II)。从行文看服务一个客户组合 ('在我的组合里'、'基于我所服务客户的判断')，同时覆盖 Snowflake 与 MongoDB。视角局限：(1) 只看自己服务的客户组合；(2) 已离职，现状判断可能滞后；(3) 产品拆分比例是'大致归属'，无账单支撑；(4) 季度口径混乱 (称 2026 年 Q1，又说 Q2 仍在进行中)；(5) 摘要稿为二手整理，非逐字转写。
阅读范围：全文 L1-L309 (Snowflake 部分 L45-L182、L226-L248；其余为 MongoDB) 可信度：需交叉验证
这是本组里唯一给出 Snowflake 消费按产品拆分的来源：核心数仓 60-65% 份额、增 20-25%；Snowpark/ML 20-25%、增约 40%；Cortex AI/agents 10-15%、增 70%+，并强调 AI 驱动的消费有一部分落在传统类别里，所以 AI 贡献可能超过 15%。重度使用 Cortex Code 的客户消费增速约 45%，可比客户约 30%；50%+ 客户用过 CoCo，但只有 25-30% 用到有实质消费影响。他也承认开源模型压缩了 Cortex 推理的溢价。
问题：本篇无对应的完整转写 (不在本次分配文件中)，只能依赖摘要稿；季度口径混乱：称'2026 年第一季度'，又在 8 月 12 日说'第二季度仍在进行中'；文中又说'如果第一季度是 34%'，与前文 32% 不一致；专家预计 Q2 较 Q1 回落 2-5 个点，与公司随后公布的 Q2 FY27 37% (加速) 相反；产品份额与增速为'大致归属'，无账单数据支撑；'收购了 Country Data' 应为 Crunchy Data，转录/翻译错误
B 2026-08-12 · MeritForum 纪要 (整理稿) · MeritForum 摘要 (北美渠道商)：与 260812 Q2业绩调研 III 为同一专家，新增 Snowflake 对 Databricks 正面胜率 60:30 3 条
文件：KB/MeritForum/2026-08/2026-08-12-bei-mei-ruan-jian-qu-dao-shang-chi-zhi-zhuan-jia-snowflake-zui-jin-ji-ge-ji-du-xiao-hao-liang-zeng-chang-wen-ding-zai-25.md
发言人：MeritForum 标为'北美软件渠道商离职专家'，内容与 260812 Merit Snowflake Q2业绩调研 III 转写一致 (25-27%、新客户 29-30%、CoCo 15-20%、明年 40-50%)。转写录音在 Databricks 选型问题处截断，摘要稿保留了后续回答。局限同该转写：渠道样本、利益方视角、现任/离职存疑。
阅读范围：全文 L1-L257；重复内容不记，只记转写截断后的新增回答 可信度：需交叉验证
新增转写中截断的 Databricks 竞争回答：正面交锋 Snowflake 赢 60%、Databricks 赢 30%；但 Microsoft 靠 EA 承诺支出成为重要挑战者，Google 也在跟进，超大云厂商越来越难对付。
问题：胜率 60:30 只见于摘要稿，转写录音截断无法核对
B 2026-08-12 · Merit 专家电话 (音频转写) · Databricks 比荷卢区销售：agent 时代 Unity Catalog 赢治理层，Snowflake 在 agent 项目中很少遇到 9 条
文件：KB/2026-08-13/26.08.12 Databricks 销售(audio).md
发言人：Databricks 现任销售 (明示，L9-L17)：在 Databricks 约 4 年，先做制药行业 AE，再做公共部门，过去 1.5 年负责比荷卢地区数字原生 (2000 年后云原生公司) 与新兴企业 (营收 $2.5亿-10亿) 团队。原文英文 L1116 说 “I'm leading the digital native and emerging enterprise team”。视角局限：竞争对手销售，天然唱多 Databricks、贬 Snowflake；样本限于比荷卢中小型数字原生客户，专家自己承认存在样本偏差 (L571-L577)；做销售不碰技术细节，“95%”“很少输”等为个人观察非官方数据。
阅读范围：纪要总结 L1-L828 全读；中文转译 L830-L1112 按 Snowflake/上下文/支出关键词检索并读上下文；英文原文 L1116 (单行 5.5 万字) 仅检索 可信度：需交叉验证
从 Databricks 一线销售视角：agent 项目里“很少看到 Snowflake、正面竞争基本没输过”，客户从 Snowflake 迁出在增加但周期长、重投入 Snowflake 的客户不轻易迁；数据仓库正被“商品化”，竞争焦点转向治理/上下文层 (Unity Catalog)。同时给出 agent 规模化的消费弹性：从试点到 1000 人生产，支出 10-20 倍；但只有 5-10% 客户已把 agent 投产。对 SNOW 的 Iceberg/治理层威胁 thesis 是反方证据，但来源偏向明显。
问题：竞争对手销售，利益偏向明显；样本仅比荷卢数字原生/新兴企业，专家自认样本偏差 (L571-L577)；'95%' 为个人观察，非官方数据 (L337, L914)；英文原文为单行超长文本 (L1116)，引文取自中文转译段
B 2026-08-13 · Merit 专家电话 (音频转写) · Databricks 现场工程专家：Databricks 同比 80% vs Snowflake 35%，去年近 200 家客户从 Snowflake 迁移，SQL 成本低 50-60% 8 条
文件：KB/2026-08-13/260813 - Merit AI infra Q2进展调研(2026-08-1.md
发言人：未具名专家。纪要无正式简介；原文自称参与“field engineering”迁移项目、“我本人评估过超过 20 个用例”(L650)、“one project that I have recently worked on from one of the BSI” (原文 L800)，推测为 Databricks 现场工程/解决方案架构侧人员；同一场访谈在 MeritForum 摘要中标为“databrick离职专家”，现任/前任存在出入，但全程以“我们 Databricks”自称。视角局限：竞争对手立场，Snowflake 迁移数字 (近 200 家、前 6 大客户) 无法核实且可能只是部分工作负载；SQL 成本比较为其自评用例，偏向明显。
阅读范围：纪要总结 L13-L590 精读与 Snowflake 相关段落 (L17-L190, L275-L316, L387-L412, 表格 L575-L590)；中文转译 L592-L796 按 Snowflake 关键词检索；英文原文 L800 仅检索 可信度：需交叉验证
同一 Databricks 专家的完整音频版本：Databricks 消费同比约 80% 是 Snowflake 约 35% 的两倍多；从 Snowflake 迁往 Databricks 是第二大迁移来源，去年近 200 家客户、甚至 Snowflake 前 6 大客户也在迁 (主要是 AI/ML 和流处理)；理由是 Lakebridge 工具、Genie 准确度高于 Snowflake Agents、Snowflake 需维护多份数据副本导致消费更高。对 E04 是有份量的反方证据，但为竞争对手一面之词。
问题：与 MeritForum 2026-08-13 摘要为同一场访谈，勿重复计数；提问者口误 Snowflake 增速 '330%' (L598)；'Snowflake 前 6 大客户在迁移' 可能仅指部分工作负载 (总结 L101 措辞为'部分工作负载')，原文未区分；专家现任/离职身份不一致
B 2026-08-13 · MeritForum 纪要 (整理稿) · MeritForum 摘要：Databricks 离职专家称 Databricks 增速为 Snowflake 两倍，Snowflake 前六大客户也在迁移 4 条
文件：KB/MeritForum/2026-08/2026-08-13-databrickchi-zhi-zhuan-jia-databricks-xiao-fei-zeng-chang-cong-2026-nian-di-yi-ji-du-yue-12-hui-sheng-zhi-di-er-ji-du-20.md
发言人：MeritForum 标题明示“databrick离职专家”(L2)，作者 Lantian (L11)；内容与 260813 Merit AI infra Q2 音频为同一场访谈，推测为前 Databricks 现场工程人员 (原文称“字段工程工作”即 field engineering, L74)。视角局限：竞争对手前员工，数字为个人观察；摘要经编辑改写，比音频纪要更概括。
阅读范围：全文 L1-L220 可信度：仅作局部线索
与 260813 音频同源的论坛摘要，标题直接把“Databricks 同比 80% 为 Snowflake 35% 两倍以上”和“从 Snowflake 迁移”作为卖点；补充了“即便 Snowflake 前六大客户也在迁移”“Snowflake 多份数据副本导致消费更高”等原句。另提到 Databricks 与 Snowflake 同样进军 Postgres。作为 E04 反方证据，但不应与 260813 重复计数。
问题：与 260813 音频为同一场访谈，重复来源；'字段工程' 为 field engineering 误译；摘要 L37 把迁移原因归纳为'较低的数据重复与消费成本'，系编辑概括
C 2026-08-14 · Merit 专家电话 (音频转写) · 微软 FDE 团队专家：FDE 不必然增加 Snowflake 用量，可能用组织度低的数据绕开 Snowflake，微软 FDE 甚至可能屏蔽 Snowflake 2 条
文件：KB/2026-08-14/260814 - Merit 企业端AI渗透进展调研III(2026-08-14.md
发言人：微软前沿部署工程师 (FDE) 团队成员 (纪要明示其团队直接服务 14 家 FDE 客户、可观察约 70 家，L23-L27)，现任，职级与年限未披露。视角局限：微软立场，倾向推荐 Azure 体系；对 Snowflake 的判断是推理而非账单数据。
阅读范围：仅检索关键词段落：L505-L529, L610, L804-L814 可信度：仅作局部线索
FDE/agent 工作流对 Snowflake 用量的影响不确定：数据已在 Snowflake 时会增加，但 FDE 能串联非结构化和组织度低的数据，可能蚕食'先把数据整理进 Snowflake'的需求；微软 FDE 还可能优先推 Azure 竞品。是 E02 '外部 agent 绕开 Snowflake' 的反例线索。
问题：推理为主，无具体客户账单数字
A 2026-08-17 · Acecamp · Databricks Genie vs Snowflake CoCo 获客与收入对比：Genie 打折抢 CoCo 用户，Foundation Model API 利润很低 8 条
文件：KB/acecamp 2/2026-08-18/260818 - Genie和Snowflake CoCo在获客和收入贡献层面对.md
发言人：未给出专家简介 (标注 '专家观点汇总')。推测为 Databricks 现任员工 (能给出 Databricks AI 收入占比约三成、$1M+ 客户 820 多家、$10M+ 70 家、Genie 定价调整、内部销售目标)，客户案例为腾讯、CLP 中电，推测属亚太区。文件时间 2026-08-17 23:55，存放于 2026-08-18 目录。视角局限：竞争对手内部人员；对 CoCo 的了解来自官网 ('官网显示有 7000 多个')。
阅读范围：全文 L1-L73 可信度：需交叉验证
Databricks AI 业务约占总收入三成多；Foundation Model API 按 token 收费、与客户直购第三方模型价格一致、利润很低 (靠规模折扣赚差价)，对 SNOW token 层毛利是直接参照。为抢 CoCo 吸引的用户，Databricks 对 Genie 大幅打折 (5 月起半价，暂无实质收入)，即价格战。专家认为 CoCo 比 Genie 更能拉新客户 (Genie 主要拉存量)，但 CoCo 治理不足，普通员工能查到其有权访问的所有内部数据。
问题：无专家简介，身份为推测；'GML 5.2' 应为 GLM 5.2，转录问题；专家把 Mosaic 归为非 AI，口径特殊；date 字段取文件内时间 2026-08-17，目录日期为 2026-08-18
A 2026-08-19 · Merit 专家电话 (音频转写) · 渠道商专家：自有客户 Snowflake 消费增速 22-23%→26%→29-30%，全年 32-34%；新业务 40-45% 与 AI/智能体相关 18 条
文件：KB/2026-08-19/260819 - Merit Snowflake、MongoDB Q2业绩调研V.md
发言人：未具名。MeritForum 同场摘要标题称"北美软件渠道商离职专家"(离职=前任)；转写中专家自称"partners like us that have global professional services capabilities…We are an extension of their FDE program…we also work with Snowflake and data bricks"，即具全球专业服务能力的系统集成/渠道合作伙伴，同时服务 Snowflake、Databricks、MongoDB。转写开头无完整履历介绍，职位/任职年份不明。视角局限：(1) 增速是其所在合作伙伴所服务客户样本的消费增速，不是 Snowflake 全公司口径，且绝对值比公司产品收入增速 (30%/34%/37%) 低 7-8pp；(2) 访谈在 Q2 FY27 财报 (8 月下旬) 之前，Q2 数字为专家观察/估计；(3) 渠道商从 Snowflake 新业务中获益，存在乐观偏向；(4) 产品占比 (仓库 55-60%/Snowpark 15-20%/CoCo+CoWork 15%) 为粗估，不看客户账单明细；(5) 对 MongoDB 部分略过。
阅读范围：全文 L1-L632 读过 (L632 为英中混合原始转写单行约 3.4 万字符，按关键词检索 Snowflake 段)；L321-L459 MongoDB 部分及原文对应段按要求略过 可信度：需交叉验证
渠道商专家称其客户在 Snowflake 上的消费逐季加速 (22-23%→26%→29-30%)，预计全年 32-34%，下半年快于上半年；驱动是遗留数仓迁移/数据整合 + 智能体上线后新增数据和负载 + CoCo/CoWork。CoCo 在企业客户中试用 50-60%，但仅 20-30% 形成显著消耗、注册用户中只有 30-40% 会用，实质采用客户消耗增长约 15-20%。竞争上 Databricks 抢新增负载 (金融/医疗/制药)，但没有整体迁出；Observe 对用量贡献仅 2-3%。
问题：增速为专家所在合作伙伴客户样本口径，比公司披露的产品收入增速 (Q4 30%/Q1 34%/Q2 37%) 低 7-8pp，只能看方向；CoCo+CoWork 约占 15% 的说法与 2 月专家 (<5%)、公司 AI 收入规模都对不上，可能把 AI 拉动的核心消费算进去，推断过度；CoCo 客户'消耗增长 15-20%'口径不清：MeritForum 摘要译为'每月/每期'，转写原文为'right now'，未说明是同比还是相对未采用客户；转写中有错字：'Covert'(应为 CoWork)、'JAI'(应为 GenAI)、'OpenAI Cortex'(应为 Codex)；访谈在 Snowflake Q2 FY27 财报之前，Q2 为专家估计；转写翻译稿 (L481-627) 没有收入模型层问答的完整答案 (多云、自带模型、MCP)，MeritForum 摘要稿 L73-74 有
B 2026-08-19 · MeritForum 纪要 (整理稿) · (摘要稿) 渠道商专家 Snowflake 消费 22-23%→29-30%，全年 32-34%——与 08-19 Merit 音频同场 2 条
文件：KB/MeritForum/2026-08/2026-08-19-bei-mei-ruan-jian-qu-dao-shang-chi-zhi-zhuan-jia-snowflake-xiao-fei-zeng-chang-you-qu-nian-di-si-ji-du-de-22-zhi-23-ti-s.md
发言人：标题'北美软件渠道商离职专家'；与 2026-08-19 Merit 音频转写为同一场访谈，身份局限见该条
阅读范围：全文 L1-L200 左右读过；与转写重复的内容不再记，只记转写翻译稿缺失的模型层问答 可信度：仅作局部线索
与 08-19 Merit 音频同场的摘要稿，数字一致。唯一新增：模型层问答里专家称 Snowflake 支持自带模型和开源模型、RAG、A2A、MCP，认为模型多样化会长期增加 CoCo/CoWork 需求。
问题：与转写重复；摘要把'15-20% 消耗增长'译为'每月/每期'，转写原文没有该限定，翻译可能失真
A 2026-08-20 · GuidePoint · Thoughtworks 8 月渠道回访：H1 增长 40%+ 且满负荷；新增平台预算 Databricks:Snowflake 约 60:40，CoCo 转化快于 CoWork，AI 计费难预测导致决策拖延 18 条
文件：KB/GuidePoint/2026-08/2026-08-20-snowflake-august-2026-ai-and-data-partner-channel-check.md
发言人：Bilal Jaffery，Thoughtworks 美洲区 SVP 兼 Data & AI 负责人 (现任，明示 L30/L130)；25 年以上经验，曾在 IBM、Bell Canada、Deloitte、Accenture、Blend360 (L130-137)；历任 Snowflake 合作伙伴，同时也是 Databricks 及各云厂商合作伙伴 (L137-139，明示)。与 5 月那篇为同一专家。视角局限：咨询 SI 渠道；他自己说 Thoughtworks 偏复杂工程项目，天然多 Databricks 业务，放大到全市场他会按 50-50 建模 (L272-276)；增速为咨询业务口径；预算拆分为主观估计；同时是两家及云厂商合作伙伴。
阅读范围：全文 L1-L517 可信度：需交叉验证
渠道层面 AI 治理/agent 工作在把数据基础重建拉成多年项目，'Snowflake 和 Databricks 都在受益'；但新增预算 Databricks 60:40 占优 (Thoughtworks 样本偏工程，全市场他按 50-50)，两家架构'已经等同'，护城河向 context/权限层转移且 Atlan 等也在抢。CoCo 早期转化快于 CoWork；客户对 AI token 成本'防御性'，ad hoc 查询难以预测导致决策拖延，客户更想要固定费率；云厂商补贴更激进，但客户'没上当'。
问题：60:40 预算拆分是 Thoughtworks 偏工程样本，专家自己说全市场按 50-50 (L276)，两种口径需并列；L481 'I was working at Snowflake' 疑为转录误差 (应为 working with)，不能据此认为其曾任职 Snowflake；增速为咨询业务口径，不是 Snowflake 消费
A 2026-08-26 · Merit 专家电话 (音频转写) · 北欧实施商专家：保险客户 AI 负载从试点到生产 Snowflake 计算 5-8 倍，但 AI 只占其消耗 15-20%，其余 80-85% 每年增 20-30%；消耗增长很多来自架构低效 17 条
文件：KB/2026-08-26/260826 - Merit Snowflake Q2调研(2026-08-26.md
发言人：未具名。转写中自述"my region as you know is is Northern Europe"，"with us as implementation firm and advisors"，并称部分理赔人员"以近岸模式外包给我们"，即北欧区域的 IT 实施/咨询/外包服务商 (系统集成商) 人员，现任 (推测，依据为用现在时描述在做的项目)。MeritForum 同场摘要标题写作"北美软件渠道商离职专家"，与转写自述的北欧区域不一致，以转写为准。视角局限：(1) 核心论据来自两个案例 (北欧寿险/养老金公司用 Snowflake；荷兰电子制造商用 MongoDB)，单客户样本；(2) 保险客户直接与 Snowflake 签约，专家看不到账单 ("they contract directly with snowflake, so I don't know exactly")，15-20% 等为估计；(3) 专家所在公司做架构优化/降本服务，强调'消耗来自架构低效'可能有业务偏向；(4) 北欧市场不代表北美。
阅读范围：全文 L1-L681 读过 (L681 为英中混合原始转写单行约 4.4 万字符，按关键词检索) 可信度：可作主要证据
北欧保险客户从试点到多业务线生产后 Snowflake 计算增长 5-8 倍，主要由理赔人员查询量驱动；但这只是 AI 相关负载 (狭义口径约占其 Snowflake 消耗 15-20%)，其余 80-85% 仓库负载每年增 20-30%，整体增速远低于倍数。专家警告大量消耗来自架构低效 (无缓存、编排差)，5-10 倍不可能长期被接受，客户会通过架构优化、采购谈判、在 Snowflake 旁建开源数据湖来压成本 (2-3 年后成重点)。75-80% 的 Snowflake 客户在加容量。
问题：MeritForum 摘要标题称'北美软件渠道商离职专家'，转写中专家自述负责北欧，身份口径矛盾；token 成本'每月'与'每年'在追问中前后矛盾，转写纪要已自行更正为每年；FDE 与 FTE 混用 (转录问题)；保险客户直接与 Snowflake 签约，专家不掌握账单，15-20%、20-30% 等为估算；单客户案例，'5-8 倍'只适用于 AI 负载，不能外推为整体消耗；建模倍数 (5-6 倍或 7 倍) 只出现在 MeritForum 摘要稿 L185，转写翻译稿截断未包含
B 2026-08-26 · MeritForum 纪要 (整理稿) · (摘要稿) 北欧保险客户 AI 相关 Snowflake 消费 5-8 倍、整体远低于倍增——与 08-26 Merit 音频同场 1 条
文件：KB/MeritForum/2026-08/2026-08-26-bei-mei-ruan-jian-qu-dao-shang-chi-zhi-zhuan-jia-qi-ye-ji-ren-gong-zhi-neng-gui-mo-hua-cong-10-zhi-20-de-ke-hu-shang-she.md
发言人：标题写'北美软件渠道商离职专家'，但与 08-26 Merit 音频为同一场，转写中专家自述负责北欧区域的实施/咨询商；身份局限见转写条
阅读范围：全文 L1-L188 读过；重复内容不再记，只记转写截断后的最后一问 可信度：仅作局部线索
与 08-26 转写同场。新增一点：建模时不宜用 10 倍上限，5-6 倍或 7 倍更合适；行业间差异不大，差异主要来自成本控制。
问题：标题身份 (北美) 与转写 (北欧) 矛盾
A 2026-08-27 · MeritForum 纪要 (整理稿) · 北美 IT 咨询专家：零售客户 AI 相关 Snowflake 负载 2-3 倍，但账户整体只 +40%；部分 AI 检索绕开数据平台 13 条
文件：KB/MeritForum/2026-08/2026-08-27-bei-mei-itzi-xun-gong-si-chi-zhi-zhuan-jia-qi-ye-ji-ai-sheng-chan-cai-yong-lu-ke-neng-zai-nian-di-da-dao-40-zhi-50-bing.md
发言人：标题'北美IT咨询公司离职专家'(前任，IT 咨询/实施)；正文无公司名、职位和年份。视角局限：(1) 只有摘要稿、无转写，无法核对原话；(2) 主案例是一家大型消费品与零售客户；(3) RPO、大客户占比等是复述 Snowflake 公开披露，数字有错 (见 issues)；(4) 作为离职人员，对当前季度的'观察'多为推测。
阅读范围：全文 L1-L218 读过 (仅摘要稿) 可信度：需交叉验证
零售客户从 PoC 到生产 6-12 个月内 AI 相关 Snowflake 负载涨 2-3 倍，但账户整体只增加约 40%，因为既有分析负载基数大。专家明确指出知识类 AI 可直接检索 SharePoint/ServiceNow 等，AI 增长不会一对一转成数据平台消费，Snowflake 必须成为治理层才能赢得位置。增长主要来自大客户扩张 (约 95% 收入来自存量客户)；RPO 在消费之前加速，是领先指标也有执行风险。
问题：同篇内大客户占比矛盾：L148 称 $1M+ 客户约占 5%，L174 称约 10%；收入占比 L175 称约 70%，L186 称约 60%；RPO 增速与公司披露不符：02-26 点评引公司数据 Q4 FY26 RPO 增速 42%，专家说 Q1 '30%+'、Q2 '中高 30%' 未经核对，Q2 时尚未发布却给出估计；只有摘要稿，无转写，无法核对逐字原话；零售案例收益数据 (检索时间 -40-50%、40 倍) 专家自称未经审计
A 2026-08-31 · Merit 路演 (音频转写) · 九千论坛周会：Snowflake 渠道调研偏正面，预计 Q2 产品收入 +35-36%，生产客户 AI 消费 5-8 倍 16 条
文件：KB/2026-08-31/260831 - Merit 公开路演：科技_制造周度调研重点更新(2026-0.md
发言人：九千/Merit 论坛周会 (路演)，软件部分由主讲分析师总述、同事“蓝天老师”(即 MeritForum 作者 Lantian) 转述 4 位头部渠道商专家 (3 北美、1 欧洲) 和 1 位咨询公司专家的访谈 (L897-L899)。视角局限：卖方式二手转述，非专家原话；渠道商只看到自己客户；Q2 预测发布于财报前，后续实际为 37%；把 '超 40% 消费与 AI 相关' 这类单一专家估计当成趋势；Databricks ARR 数据来自 '披露' 但口径混乱。
阅读范围：仅检索关键词段落：纪要总结软件/Snowflake/MongoDB/Databricks 部分 L107-L325、表格 L799-L801；转译段 L881-L1009；其他章节 (宏观、硬件、光纤等) 未读 可信度：需交叉验证
财报前的渠道调研：整体正面但'实际增长仍温和、超预期主要在 RPO 等前瞻指标'；给出两个生产化案例 (北欧养老险 AI 消费 5-8 倍、传统数据工程 +20-30%；北美消费品 AI 负载 2-3 倍、整体 +40%)、CoCo 重度用户消费 +10-20%，并明确提到模型路由降低单次成本、'价格存在变化所以金额不一定数倍增长'。增长弹性排序 Databricks > Snowflake > MongoDB。
问题：二手转述，转录错误多：MangoDV=MongoDB、Codex Code/Context Code=Cortex Code、IPO=RPO、QR/QE=Q2/Q1、Databreak=Databricks、DBG=数据库；'4540 到 50 个人'、'超 4500 个人' 数字明显转录错误 (L927)；Merit 说'财报发布后股价积极反应'(L883) 指的是其他已发财报公司，Snowflake Q2 尚未发布；L897 称部分案例消耗 5-10 倍、传统数据库 30%+，与 L915/L919 的 5-8 倍、20-30% 口径不一；Databricks ARR '20 到 50 至 60' 口径混乱 (L1007)
A 2026-09-01 · Funda (Substack) · Funda AI 2Q26 (FY27Q2) 前瞻：伙伴渠道显示 CoCo 继续驱动加速，企业开始把更多数据归集到 Snowflake 28 条
文件：KB/Substack/2026-09/2026-09-01-preview-snow-2q26-cortex-code-continues-to-drive-strong-growth.md
发言人：作者 FUNDA/FundaAI (发件人明示)。一手材料来自：美国 Partner A-D (Snowflake 销售/实施伙伴，未具名；Partner D 只有 practice revenue，覆盖科技/网络安全/零售/制造/金融客户)、欧洲 Partner (上期明示为 Cognizant)、一家英国政府机构、一家北美大型药企的联系人。追踪方法同上期：伙伴按 consumption 考核，跟踪其 consumption 达成率和产品构成。视角局限：伙伴只看自己的客户群且有推广 Snowflake 新产品的动机；Partner C 的 CoCo 增量占比与 Partner D 的总占比口径不同；部分为“方向性估计”(专家自己声明)。
阅读范围：全文 L1-L521 (L501-521 为图表，未读到内容) 可信度：需交叉验证
多数伙伴 2Q26 consumption 超预期 (Partner C +32% vs 预期 27-28%)，CoCo 5-7 月消费是 1-4 月近 3 倍、仍月增 15-20%，占增量 consumption 约 12-15%；但 Partner D 的实施收入显示加速主体仍是核心数仓迁移 (约 60%)，CoCo 只占约 15%。不利点：AWS 正把销售资源转向 Databricks 和 Anthropic、CoCo 毛利率约 52-55% (低于公司 70-75%)、token 成本将受采购审查、Observe 进展有限、Iceberg/Snowpark 被看作防御性产品。Funda 基准预测 FY27Q2 增速 35% (超指引 4%)，3Q 加速到 36%，即使落在或略低于 bogey 也视为正面。
问题：Partner C 称 CoCo 占总消费接近两位数 (L193)，Partner D 称约 3% (L292)，欧洲伙伴称约 2-2.5% 收入，差异很大；欧洲伙伴说 CoCo 在 '4Q2026' 占比远低于 1% (L365)，应为 4Q2025 笔误；Funda 季度命名用日历年 (2Q26 = FY27Q2)；图表 L501-521 无法读取
A 2026-09-02 · FactSet 业绩会转录 · Snowflake Q2 FY27 业绩会：增速 37%、指引 +36%，AI 约贡献一半加速，产品毛利率指引降到 74% 29 条
文件：KB/FactSet/2026-09/2026-09-02-corrected-transcript-snowflake-inc-snow-us-q2-2027-earnings-call-2-september-2026-5-00-pm-et.md
发言人：CEO Sridhar Ramaswamy、CFO Brian Robins、EVP Product Christian Kleinerman、IR Katherine McCracken (转录稿明示)。视角局限：管理层自述；关键拆分 ('约一半') 为 CEO 口头粗略估计，口径宽
阅读范围：全文 L1-L438 可信度：可作主要证据
Q2 产品收入 +37%，FY27 指引上调到 $6.07B/+36%，Q3 指引 +37–38%。CEO 说宽口径 AI 产品 (CoCo/CoWork/AI functions/AI Gateway) 约贡献一半加速；产品毛利率指引因 AI 组合从 75% 降到 74%，CFO 明说先追采用再管毛利。反方信号：RPO 增速 38%→30%、use case 领先指标放缓、CoWork 账户环比仅 +11%、CoCo 被大量用来省钱、AI 采用者消费提升不披露。
问题：'约一半'中的 AI 产品口径包含 AI functions 和 AI Gateway，不等于 thesis 的 token 转售层；另一半含 Notebooks/Streamlit/迁移，不等于纯核心数仓；Q1 叫 Snowflake Intelligence，Q2 改称 CoWork，账户口径是否一致未说明；'the mix between our AI products and the course changed' 应为 core，转录错误；'150 Snowflake on Snow' 疑转录错误；全程未提 Iceberg、Databricks、BigQuery、价格/折扣、性能改进逆风、回购、SBC、Anthropic 承诺金额；GRR '相对平稳' 无数字；'3 个月达 80% 采购容量' 指标无数字
A 2026-09-02 · TMTB (Substack) · TMTB 收盘点评：SNOW FY27Q2 盘后 +20%，产品收入和 Q3 指引都远超街和买方 bogey 5 条
文件：KB/Substack/2026-09/2026-09-02-tmtb-eod-wrap-snowflake-snow-broadcom-avgo-hpe-first-takes.md
发言人：TMT Breakout (TMTB，发件人 tmtbreakout@substack.com 明示)，买方视角的盘后速评，记录街预期和买方 bogey。视角局限：财报当晚第一反应，未经电话会细节核实；称 SNOW 是软件里最拥挤的名字。
阅读范围：全文 L1-L182；只记 SNOW 相关段落 (L20-39)，另记 MDB 一句作对照 可信度：可作主要证据
SNOW 盘后 +20%：FY27Q2 产品收入 $1.49B (+37%)，街预期 $1.42B (+31%)，买方 bogey 36%；Q3 指引 $1.591B (+37.5%)，街 $1.50B (+29%)，买方 bogey 32-33%；FY27 增速从 31% 上调到 36%。TMTB 称 Q3 很可能提前一个季度达到 40%+，管理层把远超当季超额的部分加进全年指引，多头完全占据主动。
问题：L30 图表 (CoCo + core 拆解) 无法读取；原文有拼写错误 'explaisn the CoCO'
A 2026-09-03 · Funda (Substack) · Funda AI FY27Q2 点评：超越单一产品的数据飞轮，产品收入 +37%，FY27 指引上调到 36% 19 条
文件：KB/Substack/2026-09/2026-09-03-review-snow-fy27q2-beyond-a-single-product-a-data-flywheel-that-merits-a-longer-horizon.md
发言人：作者 FUNDA/FundaAI (发件人明示)。引用的是 FY27Q2 电话会管理层原话 (未标注具体说话人，CoCo/飞轮部分推测为 CEO Sridhar Ramaswamy，毛利率/招聘部分推测为 CFO)。模型数字来自图表 alt 文本 (L194-208)。视角局限：Funda 一贯看多 SNOW，叙事带自我验证；模型只在 alt 文本里看到概要。
阅读范围：全文 L1-L210 (图表只读到 alt 文字描述) 可信度：可作主要证据
FY27Q2 产品收入 +37%，超卖方一致预期 5.2%、超买方预期 1%；Q3 指引隐含 37-38%，按 3% beat 推算约 42%；FY27 指引从 31% 上调到 36%；产品毛利率指引降到 74%，但 OPM 指引从 13.5% 升到 14.5%。管理层称 AI 产品贡献约一半加速，另一半来自核心平台、迁移、Notebooks、Streamlit/React 应用；新客户用完 80% 容量的速度显著变快 (未量化，即提前消费)。CoCo 的 top 10 技能里有成本管理 (帮客户省钱)。Funda 模型：CoCo 五个季度从 $30k 到 $59.4m，数仓占产品收入从 96% 降到 83%，产品增速峰值 42%。
问题：'一半加速来自 AI' 未说明计算方法；新客户容量消耗加快未量化；模型数字只来自图表 alt 文本，具体季度和 FY28 数字未知；电话会引语未标注说话人
B 2026-09-03 · Acecamp · Databricks 视角：OLTP/OLAP 融合上 Snowflake 更快更成熟但两条路线定位模糊，CoCoEvolve 可迁移 Databricks/dbt/Airflow 工作流 5 条
文件：KB/Acecamp/2026-09/2026-09-03-bei-mei-ai-saaslong-tou-ke-hu-luo-di-tiao-zhan-tou-hao-agenticxiang-mu-xu-qiu-fan-kui-shu-ju-ku-rong-he-dui-huo-ke-de-yo.md
发言人：未给出专家简介 (标注 '专家观点汇总')。从内容推测为 Databricks 现任员工，负责 GCR/APAC 区域 (给出 GCR/APAC 约八成项目仍在 PoC，了解内部 Omnigent '零号用户' 使用情况与销售策略)。视角局限：竞争对手内部人员；对 Snowflake 是外部观察 ('不太确定 Snowflake 的情况')，尤其是中国市场的 Snowflake；样本为亚太。
阅读范围：全文 L1-L100 可信度：仅作局部线索
专家认为在 OLTP/OLAP 融合上 Snowflake 的 hybrid table 迭代快、比 Databricks 7 月才 GA 的 Lakebase 成熟，但 Snowflake 有 hybrid table 和 Snowflake Postgres 两套体系，多个客户反映定位模糊、内部互相竞争。CoCoEvolve 能把 Databricks、dbt、Airflow 的工作流迁到 Snowflake 并创建业务 context，Databricks 用 Agent Bricks 和自己的迁移 agent 应对。Snowflake 管理层执行力强，推出 CoCo 等能迅速获客。
问题：无专家简介，身份为推测；'LTAP' 可能是 Lakebase/HTAP 的转录混淆
B 2026-09-07 · Funda (Substack) · Funda AI 周报：SNOW 数据飞轮与 CoCo，软件空头回补接近尾声 4 条
文件：KB/Substack/2026-09/2026-09-07-weekly-snow-data-flywheel-coco-optics-scale-in-be-delivery-debate-enterprise-ai-vol-2-crdo-mdb-panw.md
发言人：作者 FUNDA/FundaAI (发件人明示)。SNOW 部分是 FY27Q2 前瞻/点评复述 + 市场环境评论。视角局限：周报复述，无新一手数据。
阅读范围：全文 L1-L231；只记 SNOW 相关段落 (L16-30, L65) 以及与 SNOW 相关的企业 AI 预算/MDB 段落 可信度：仅作局部线索
Funda 称 SNOW 是渠道数据最干净的体现：产品收入 +37%、超预期 5.2%、全年指引 31%→36%，CoCo 账户 9,100、CoWork 5,800，客户把交易型数据迁到平台；认为估值应基于数据飞轮而非单一 SKU。同时提示软件空头回补基本结束、软件跑赢已显拉伸 (利于判断仓位)；企业 AI 研究显示增量在 API 和生产工作流，部分 CFO 已冻结 IT 总预算。
问题：无新一手数据
A 2026-09-08 · FactSet 会议转录 · 高盛 Communacopia：CEO 谈推理转售不创造价值、愿为 CoCo 牺牲毛利、迁入变快迁出也变快 17 条
文件：KB/FactSet/2026-09/2026-09-08-corrected-transcript-snowflake-inc-snow-us-goldman-sachs-communacopia-technology-conference-8-september-2026-1-50-pm-et.md
发言人：主持：Gabriela Borges (Goldman Sachs 软件研究负责人)；台上 CEO Sridhar Ramaswamy、CFO Brian Robins。均为明示 (校对版转录)。视角局限：Q2 FY27 财报后一周的卖方会议，管理层宣传语气；数字很少。
阅读范围：全文 L1-L289 可信度：可作主要证据
这场最有价值的是 AI 收入质量：Sridhar 说单纯转售模型容量是“fake news”，偏好顺序是 CoCo/CoWork > 卖模型容量 > 只做 Claude 的数据后端 (勉强接受)，并公开表示若 CoCo 大规模普及拖累毛利他愿意解释。竞争端罕见直白：迁入快也意味着迁出快，托管 Iceberg 后价值必须上移；Gen2 的价格通缩被量和新作业抵消但没给数字。
问题：Gen2 性能改进逆风只有定性说法，未给收入影响比例；AWS $6B 合同原文没说年限；CoCo 毛利护栏未给任何数字；转录有小错 (I want to underestimate 疑为 I don't want to underestimate)
B 2026-09-08 · Merit 专家电话 (音频转写) · Snowflake 上海 BUILD/开发者活动 Session1-2：CoWork 演示、Raven 内部用例、Text-to-SQL 专用小模型与 context engineering 11 条
文件：KB/2026-09-08/SNOW CoCo Part1(audio).md
发言人：不是专家访谈，而是 Snowflake 自办的上海开发者活动 (同日下午 CoCo 黑客松前的 5 场 session) 现场录音转写。主讲：(1) Sally，自称'负责中国的国际方案工程师'(L682，明示)；(2) 姚政伟，Snowflake AI 实验室 Principal Researcher (L788 自我介绍；主持人 L786 称'首席科学家周伟老师'，名字/头衔有识别出入)。文末引出 Zscaler 亚太 CSO Lawrence (L930)，其内容不在本文件。视角局限：全部是 Snowflake 员工的产品宣讲/营销口径，面向中国开发者和客户，无计费、毛利、消费量等商业数据；演示数据为 demo，所有效率/成本数字为公司自报内部测评，未经第三方验证；中国区是 Snowflake 很小的市场，不代表北美大客户用法。
阅读范围：全文 L1-L938 (L1-L667 为纪要总结，L668-L930 为整理版转译，L938 为单行原始转写；quote 主要取自整理版转译段) 可信度：仅作局部线索
Snowflake 自己的产品宣讲：卖点是'把 AI 带到数据里'——自建语义层让 agent 少找错表、少全表扫描，宣称 token 用量降低 20-50%、上下文优化成本降 33%，并用自训 8B Text-to-SQL、4B 自动补全等专用小模型。对 thesis 的含义是双向的：一方面承认大部分企业现在用外部 agent 经狭窄管道连 Snowflake、只把它当存储和管道 (外部 agent 也带来全表扫描的计算消费)；另一方面 Snowflake 主动把'帮客户省 token/省扫描'当卖点，说明 AI 带来的单次查询效率提升会抵消部分量的增长。没有任何关于 CoCo 计费方式、毛利或拉动倍数的数据。
问题：非专家访谈，是 Snowflake 官方活动宣讲；所有数字为公司自报内部测评，营销属性强；可能转录问题：Cortex 被识别为 'Vertex AI/Vertex Agent' (L770)，MCP 被识别为 'NCP' (L752)，DeepSpeed/ZeRO 团队被识别为'微软 system team' (L826)，'ContentSense/Podcast Center' 可能是 Cortex 相关产品名误识 (L732/L774)；篇内矛盾：主持人称讲者'首席科学家周伟'(L786)，讲者自称'姚政伟 Principal Researcher'(L788)；Text-to-SQL 模型参数'8B / 300万 / 小于30亿'前后矛盾 (L816)；benchmark 成本'3毛钱 vs 0.14'与'降低33%'方向不一致 (L852)；token 降低 20-50% 未说明基线、样本和计量口径；没有涉及 CoCo/CoWork 的定价、计费 (credit/token)、毛利或对数仓计算的拉动倍数
B 2026-09-08 · Merit 专家电话 (音频转写) · Snowflake 上海活动 Session3-4：Horizon 的 AI/agent 治理，以及 Iceberg+Polaris 开放架构与 Databricks 双向读写演示 10 条
文件：KB/2026-09-08/SNOW CoCo Part2(audio).md
发言人：Snowflake 官方活动现场录音。主讲：(1) Sebastian Weitzner，Snowflake Security Practice Lead (L715/L953 自我介绍，明示)，称已做过 100+ 次客户交流；(2) Bruce 陈，Snowflake 中国区高级解决方案工程师 (L857，明示)。视角局限：均为 Snowflake 员工、产品宣讲口径；Sebastian 明确说部分功能未 GA、属前瞻性陈述 (L717)；Bruce 面向中国区客户，讲的是开放架构卖点，不涉及收入、定价或客户迁移数据；没有提到 CoCo 计费。
阅读范围：全文 L1-L953 (L1-L709 纪要总结，L711-L947 整理版转译，L953 为单行原始转写) 可信度：仅作局部线索
前半段是 Horizon 把治理延伸到 agent (agent 身份感知的 masking、受限会话、Cortex Guard 覆盖 CoCo)，Snowflake 想把'治理'做成 AI 时代的控制面。后半段 Snowflake 中国区 SE 正面宣传开放：Polaris/Iceberg 让 Spark/Trino/Databricks 对同一份数据双向读写，已支持 9 个外部引擎读、现在也支持写，并明说'客户可以迁移甚至彻底离开'、'性能没优势就会被淘汰'、'过去不开放，是被激烈竞争逼的'——这直接支持我们'Iceberg 让计算可分流、单价承压'的担忧，同时 Snowflake 的应对是把治理 (Horizon) 留在自己手里，让外部引擎读也要走它的策略。
问题：非专家访谈，Snowflake 官方宣讲；Sebastian 明确部分功能未 GA、属前瞻性陈述 (L717)；可能转录问题：'CodeGo' 应为 CoCo/Cortex Code，'Quality AI Governance' 可能为 Cortex AI Guard 相关名称，'Octa/Idra' 为 Okta/Entra，'普渡' 可能为某 catalog 名，'RSS' 可能为 restricted session；'Polaris 于 2020 年捐献'(L877) 与 '2024 年捐献'(L891) 篇内矛盾 (实际为 2024)；'计划于 Q2 推出'的 AI 治理仪表板，与 2026-09 的日期不符，可能是复用的旧材料或指 Snowflake 财年季度，口径未明；演示数据和流程为 demo，未提供客户实际把负载迁出/迁入的数据
B 2026-09-08 · Merit 专家电话 (音频转写) · Snowflake 上海 CoCo 黑客松：AWS 合作分享 + Dash Desai 现场演示 CoCo 的用法 (CLI/Desktop/VS Code/Snowsight、skills、plan mode、自动选模型) 10 条
文件：KB/2026-09-08/SNOW CoCo Part3(audio).md
发言人：Snowflake 官方黑客松现场录音。讲者：(1) 沈晶，自称'来自亚马逊'(L529)、曾任大数据产品经理 (L543)，AWS 合作伙伴视角；(2) Louis Lee，Snowflake 亚太区产品市场总经理 (L599)；(3) Dash Desai，Snowflake Principal Developer Advocate，20 年以上经验，来自旧金山总部 (L609-L611)；(4) Lin Xiao，Snowflake 高级解决方案工程师，讲账号领取 (L733)。视角局限：这是面向开发者的产品演示和活动说明，展示的是 CoCo 能做什么 (demo 用 Marketplace 数据)，不是客户实际怎么用、用了多少；没有任何定价、计费、毛利、消费量数据；AWS 讲者转写质量很差，大量断句缺字。
阅读范围：全文 L1-L767 (L1-L523 纪要总结，L525-L761 整理版转译，L767 为单行原始转写) 可信度：仅作局部线索
这是理解 CoCo 产品形态最直接的一篇：CoCo 是一个 agentic 编码/构建工具，有 CLI、Desktop、VS Code 插件和 Snowsight 右侧面板四种入口，能用自然语言建数据管道、Streamlit 类应用、ML 模型 (经存储过程接入 Snowflake Intelligence)，支持 skills、plan mode、Git 导入、多会话、Gmail/Slack 连接，模型默认 auto 由 Snowflake 选择。对 thesis：CoCo 的产出 (管道、应用、ML 模型、定时任务) 都运行在 Snowflake 计算上，机制上支持'第(1)层 AI 拉动核心计算'；默认 auto 选模型意味着 Snowflake 掌控模型路由、可调节 token 成本。但没有任何量化。AWS 讲者给出 Snowflake 约 70% 客户与 AWS 相关 (亚太 75%)，并强调'不可能 all in 单一平台'、数据可留在 S3。
问题：AWS 讲者部分转写严重缺字断句 (L767 原文大量'……包括，包括'），70%/75% 的主语缺失 (原文'70%的客户是把...')，'31亿、85亿→96亿'口径完全不明，只能作线索；可能转录问题：'Core' 可能是 Snowflake Intelligence 或 CoWork 的误识；'Jenkins 的计划模式'(L703) 实为 plan mode；'Google Data Stream'(L713) 可能是某 Snowflake 流数据产品名误识；'3D 服务'应为第三方服务；全部是 demo 场景，没有客户实际使用、用量、账单或计费信息；未披露 CoCo 底层模型 (只说 auto)；日期：文件夹日期 2026-09-08，活动为上海首场 Snowflake 开发者活动，具体举办日未说明
C 2026-09-08 · Merit 专家电话 (音频转写) · Snowflake Data for Breakfast 上海站：'智能体企业'四要素 keynote 与经营分析/治理/每日简报演示 2 条
文件：KB/2026-09-08/SNOW - Part1(audio).md
发言人：Snowflake 官方 Data for Breakfast 上海站录音。讲者：(1) Christian，Snowflake 高管 keynote (L433'谢谢Christian'，明示名字；推测为 Christian Kleinerman，仅凭名字和讲 product 愿景推测，未经确认)；(2) 冯金鹏，Snowflake 中国区解决方案顾问 (L433，明示)。视角局限：高管愿景演讲 + 售前演示，无任何财务、用量或计费数据；演示企业为虚构/示例场景，转写质量很差 (中文演示部分大量缺字)。
阅读范围：全文 L1-L513 (L1-L401 纪要总结，L403-L507 整理版转译，L513 为单行原始转写) 可信度：仅作局部线索
高管 keynote 把 Snowflake 定位为'AI 数据云'，提出智能体企业四要素 (模型、企业数据与上下文、应用集成、agentic control plane)，并说'部署 agent 数据会自己理顺'是一年前的普遍误区——这与市场'Snowflake 是 AI 时代企业数据控制面'的叙事一致，但纯属定性宣传。演示部分无与 thesis 相关的数字。
问题：'Observer' 很可能是 Snowflake Intelligence 或 CoWork 的转写误识，无法确认；中文演示部分转写严重缺字 (如'荧光是新生代的党群创建'L473、'北京当前的成功率、之后5年有5'L481)，内容不可靠；Christian 的身份仅凭名字推测，未经确认
C 2026-09-08 · Merit 专家电话 (音频转写) · Salesforce/Agentforce 生态专家：AI 预算以 token 为大头，增量流向 Data Cloud/Snowflake/Databricks，Zero Copy 不耗 credit 2 条
文件：KB/2026-09-08/260908 - Merit 企业AI部署进展调研(2026-09-08 16_.md
发言人：未具名专家，纪要无正式简介。内容围绕 Agentforce 渗透率、credit 计费、POC 免费 10-20 万 credit、实施项目经验 (L19-L45, L908)，推测为 Salesforce 内部或 Salesforce 实施伙伴的 Agentforce 交付/解决方案人员。视角局限：Salesforce 视角，对 Snowflake 只是顺带；所称 'Snowflake、Databricks 增速 20% 以上、收入 70 亿与 67 亿' 与公开数据不符 (Snowflake 产品收入增速 30%+)，可信度低。
阅读范围：仅检索关键词段落 (Snowflake/Databricks)：L119, L383-L425, L545, L837, L906-L914, L992-L1030, L1108 可信度：仅作局部线索
企业 AI 费用中 token/credit 是大头，增量预算流向 Data Cloud、Snowflake、Databricks 等数据底座；Salesforce Zero Copy 与 Snowflake 的 server-to-server 直连'不会消耗太多 Credit'——零拷贝共享对 Snowflake 计算消费可能是中性偏负。
问题：'Snowflake 与 Databricks 增速 20% 以上、收入 70 亿与 67 亿' 与公开数据不符，专家对 Snowflake 的数字不可靠
A 2026-09-09 · MeritForum 纪要 (整理稿) · Snowflake 离职专家：BFSI 贷款审批用例上线 co-work 后月消耗 4 万→7 万美元；CoCo 先行、co-work 在后；自动路由控制 token 成本 18 条
文件：KB/MeritForum/2026-09/2026-09-09-snowflakechi-zhi-zhuan-jia-snowflake-de-ai-kuo-zhan-jiang-yong-li-tui-xiang-sheng-chan-huan-jing-sheng-chan-yong-li-shu.md
发言人：标题'Snowflake离职专家'(前 Snowflake 员工)，正文无职位和年份；从回答看熟悉解决方案/售前架构 (Document AI、OpenFlow、Snowpark Streaming 等细节)，推测为解决方案工程师或架构师类岗位 (推测)。视角局限：(1) 已离职，大量数字是复述公开披露并有明显错误 (见 issues)；(2) 案例数字 (BFSI、汽车零部件) 为其接触客户，单客户样本；(3) 前员工对平台价值主张有认同偏向，几乎不谈竞争和价格压力；(4) 只有摘要稿。
阅读范围：全文 L1-L255 读过 (仅摘要稿) 可信度：需交叉验证
一家 BFSI 客户把贷款申请处理从 4 天降到约 5 分钟，日处理约 1 万份；开放 co-work 给 200 多名网点人员后月消耗从 4 万美元升到 7 万美元 (该用例 PoC 时约 5 千美元/月)，消费增长主要来自'更多用户能直接查数据'。另一家水泥/制造客户用 co-work 替代数百张 Power BI 报表，消费增长 25-36%。专家强调 Snowflake 的价值在数据上下文、引用和模型选择，自动路由把任务分给 GLM/Qwen/DeepSeek 等便宜模型以控 token 成本——这对 token 转售收入是压力。专家还称 AI 负载毛利率 34%。
问题：复述公开数字多处可疑：'89% 用例转入生产，去年也 89%'自相矛盾；co-work '2,500 客户、同比 11%'与新产品不符；'828 家 $1M+ 客户≈828 家福布斯 2000 客户'数字重复；'14,500 客户'未核实；'AI 负载毛利率 34%'无出处，未说明是公司披露还是推测，需用公司 10-Q 核对；BFSI 案例口径不一致：+25-30%、+近 40%、4 万→7 万 (+75%) 三个数字；GPT-6、Claude Opus 5、GLM 5.2 等模型名称来自访谈原文，未核实；仅摘要稿
A 2026-09-09 · GuidePoint · 前 Databricks 十年现场工程师谈 Databricks vs Snowflake：非结构化数据、Unity Catalog、Iceberg 摩擦与模型网关 16 条
文件：KB/GuidePoint/2026-09/2026-09-09-databricks-playbook-a-field-perspective-on-product-momentum-consumption-drivers-and-the-snowflake-hyperscaler-battle.md
发言人：Peyman Mohajerian，现任 Mistral AI Partner Solution Architect (明示 L37)；2016 年至 2025 年 12 月在 Databricks 任现场工程/解决方案架构师约 10 年，覆盖零售行业 (明示 L167、L178)。视角局限：前 Databricks 员工，对 Databricks 天然偏向 (自己说 'I may be biased' L514)；对 Snowflake 的了解是二手且有时滞 (多次说 'from my understanding'、'at least up to 6 months ago')；现任 Mistral 对开源模型/网关话题有利益；未给 Snowflake 任何账单数字。转录中 L394-L401 说话人标签错乱 (DG/PM 互换)。
阅读范围：全文 L1-L628 可信度：需交叉验证
从 Databricks 视角看，消费扩张来自更多数据、更低延迟 (实时 ETL)、更多用户/数据共享和 agent；Databricks 赢在非结构化数据、开放格式统一治理、给工程师的控制力，Snowflake 赢在易用、BI/仓库、非技术型初创。关于 Iceberg：Snowflake 要获得最佳性能仍需转成专有格式 (截至 6 个月前存在摩擦)。AI 函数/批量推理两家'几乎相同'，都调用 Anthropic/OpenAI 端点，没有差异化——对 Snowflake AI token 收入质量是反面信号。
问题：L394-L401 说话人标签错乱：L398 的问题实为主持人、L401 起的回答实为专家；专家对 Snowflake 的描述多为二手且自承有时滞和偏见 (L386、L514)；标题中 'Lakehouse' 在 L242-L280 应为 Lakebase (Databricks Postgres)，疑为转录错误
A 2026-09-09 · Acecamp · Snowflake 与 Databricks AI 商业化调研：Snowflake 靠新产品和 AWS 算力绑定吸引客户迁入；Cortex 问数产品留存存疑 14 条
文件：KB/Acecamp/2026-09/2026-09-09-snowflakeyu-databricks-aishang-ye-hua-diao-yan-snowflakejie-xin-chan-pin-yu-awssuan-li-bang-ding-xi-yin-ke-hu-qian-ru-da.md
发言人：未给出专家简介。从内容推测为 Databricks 亚太区 (APJ/GCR) 内部人员：能给出 Databricks 2026 年 8 月分区域增速 (北美 Enterprise 64%、北美平均 72%)、GCR 客户集中度、内部模型测试结果，客户包括 FunPlus、腾讯、Swire Coca-Cola、Reliance、InMobi。可能与 2026-08-03 那篇是同一位或同类专家。视角局限：竞争对手内部人员；对 Snowflake 的了解是外部视角；对 Snowflake 事实有明显错误 (称 Cortex '刚刚推出两个月'、'新增客户超过 7,000' 混淆 CoCo 账户数)；亚太视角，而 Snowflake 重心在北美。
阅读范围：全文 L1-L171 可信度：需交叉验证
专家认为 Snowflake 37% 增长更多来自份额获取而非行业整体，新 CEO 后 Hybrid Table、Postgres、Cortex 等新产品甚至吸引部分 Databricks 客户迁入，原因之一是借 Snowflake 与 AWS 的绑定拿到稀缺 GPU。但同时提出几个不利点：Snowflake 的 AI 场景与数据场景割裂、AI 对存量客户的拉动与新客户无异 (NRR 持平)；Cortex 类问数产品粘性弱、获客速度在下降、早期增长部分是尝鲜效应 (Genie 100 个 space 中 90% 不活跃)；Snowflake 可能把 AI 应用产生的数据用量单独计为 AI 收入，口径与 Databricks 不同。
问题：无专家简介，身份为推测；专家对 Snowflake 事实有误：称 Cortex '刚刚推出两个月' (L39)、'新增客户超过 7,000' 应为 CoCo 账户数；Snowflake 产品名 'Postgres Page'、'Cortex Evolve' 可能是转录错误；模型名称 (K3、GLM-5.3、Claude 5.1) 无法核实；Snowflake 创始人背景说法 (来自 BigQuery 团队) 与公开信息不符 (创始人来自 Oracle)，可信度存疑
A 2026-09-09 · Merit 专家电话 (音频转写) · Snowflake 现任一线专家谈 CoCo/CoWork 采用、客户结构与两个 AI 消费案例 17 条
文件：KB/2026-09-09/260909 - Merit 企业AI进展调研II(2026-09-09 10_.md
发言人：未具名专家。全程以“我们/Snowflake”自称，并说“在我近期参与的一个项目中”亲自交付 BFSI 贷款审批和汽车零部件报价项目 (L673, L725)，推测为 Snowflake 现任解决方案架构师/现场工程或交付人员，任职年限未披露；纪要未给出正式简介。视角局限：站在 Snowflake 立场推销口径明显 (模型中立、Token 更省、全托管)；公司层面数字多为复述财报且口径混乱 (89% 用例进生产 同时说“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 vs 740；74% 毛利率“完全源于 AI 工作负载”的因果解读有误)；客户案例是单一客户样本，且自己是交付方，有利益偏向。
阅读范围：关键词检索后精读：纪要总结 L13-L569 全读，转译段 L571-L795 全读；原文英文 L799 仅检索核对个别句子 (74% 毛利率、692 新客户) 可信度：需交叉验证
Snowflake 内部人给出最正面的消费故事：CoCo 约 7,100 账户、CoWork 单季 2,500+，BFSI 客户上线 Cortex Agents/CoWork 后月消费从 $40K 到 $70K，单用例从 POC $5K/月到生产 $30K/月，汽车零部件客户消费同比 +35-36%。但同一人也明确说 CoCo 生成的查询“Token 消耗量明显更少”、自动路由到 GLM/Qwen/DeepSeek 以降低客户成本，且开源模型成本仅为 1/20——说明 AI 层的 token 收入有通缩/让利机制；公司层面数字多为复述且口径错乱，不能当一手数据。
问题：公司层面数字为专家复述且多处矛盾：89% 用例进生产与“去年也超过 89%”；生产比例提升 18% 又说 43%；新增客户 692 又说 740 (L575-L579, L639)；“毛利率指引 74% 完全源于 AI 工作负载”因果表述含糊，应理解为 AI 拉低毛利率 (L583)；BFSI 案例增幅 25-30%/近 40% 与 $40K→$70K (+75%) 不一致 (L685 vs L693)；CoWork '同比增长约11%' 与产品刚上线矛盾，可能转录或口径错误 (L581)；转录错误：Astra 应为 GPT/某模型，Deep Sea=DeepSeek，Covert=CoWork，Cloud Desktop=Claude Desktop，MCB=MCP；专家为 Snowflake 现任员工且是案例交付方，存在明显利益偏向
A 2026-09-10 · FactSet 会议转录 · 花旗 TMT 会议：CFO 解释 AI 收入按 SKU 计、CoCo 本身不收费、新产品指引打折、毛利指引 74% 17 条
文件：KB/FactSet/2026-09/2026-09-10-raw-transcript-snowflake-inc-snow-us-citi-global-technology-media-and-telecom-conference-10-september-2026-12-35-pm-et.md
发言人：台上仅 CFO Brian Robins (明示)；主持为花旗分析师，转录标为 Unverified Participant，未写名字 (推测为花旗软件分析师 Tyler Radke，无原文依据)；IR Katherine McCracken 在场。视角局限：CFO 单人、RAW 未校对转录，数字主要是重复财报口径。
阅读范围：全文 L1-L276 可信度：可作主要证据
关键是 AI 收入口径：Snowflake 的“AI”按 SKU 计，CoCo 本身不收费、只收它驱动的消费，所以 AI SKU 收入和 AI 带动的核心消费是两回事；采用 CoCo 的客户核心消费更多，但未给数字。指引上 CoCo 首季没纳入、第二季打折纳入，新产品基本不计入；毛利率指引 74% (此前约 75%)，承认 AI 产品毛利更低；上半年净增 330 人 (含 Observe 约 180 人) vs 去年 950 人。
问题：RAW 转录：JB, Jonathan is our new CFO 实为新 CRO；主持分析师姓名未出现；毛利率 74% 是全年指引，与 Investor Day 的“持平约 75%”相比已下调约 1pt，CFO 未解释幅度；AI 按 SKU 计意味着 CoCo 带动的数仓消费计入核心，“一半来自 AI 产品”取决于这个口径
C 2026-09-10 · Merit 专家电话 (音频转写) · 奢侈品集团 (LVMH 系) 专家：集团统一采购 Snowflake 等 AI infra，agent 增多带动数据底座扩容 1 条
文件：KB/2026-09-10/260910 - Merit 企业AI进展调研IV(2026-09-10 16_.md
发言人：未具名专家，推测为 LVMH 集团内部数字化/IT 相关人员 (纪要多次以'集团'自称并举 LV、Tiffany 门店应用为例，L47, L333)，职位年限未披露。视角局限：甲方单一集团，Snowflake 采购为统一 AI infra 口径、金额不清；对 Snowflake 无具体用量数据。
阅读范围：仅检索关键词段落：L317-L325, L403-L409, L569, L689-L691, L723-L727 可信度：仅作局部线索
大型非科技企业确有采购 Snowflake 作为数据底座，并认为 agent 增多会让基础设施消耗'指数级'增长，但单项金额不清，属方向性表述。
问题：无量化，'指数级增长'为定性判断
A 2026-09-15 · FactSet 会议转录 · Piper Sandler 会议：CFO 给出 CoCo 9,100 账户、采用后核心消费 +11%、超预期一半来自 AI 产品 17 条
文件：KB/FactSet/2026-09/2026-09-15-corrected-transcript-snowflake-inc-snow-us-piper-sandler-growth-frontiers-conference-15-september-2026-9-30-am-et.md
发言人：主持：Rob Owens (Piper Sandler 技术研究负责人，覆盖基础设施和安全)；台上仅 CFO Brian Robins (明示)，IR Katherine McCracken 在场。视角局限：CFO 单人，11% 数字转述自 CEO 的 LinkedIn 帖子，方法未披露。
阅读范围：全文 L1-L204 可信度：可作主要证据
给了几个关键数字：约 9,100 个账户用 CoCo、5,000 多个用 CoWork；采用 CoCo/AI 的客户核心业务比原本多 11%；Q2 超预期一半来自 AI 产品、一半来自核心 (注意是“超预期”而不是“加速”)；迁移时间从 10-11 个月降到约 6 个月。CFO 否认 GAAP 盈利目标约束投入，上半年净增 330 人；第四季度续约集中但称无担忧。
问题：11% 增量的口径 (对照组、时间窗、是否含迁移) 未披露，且转述自 CEO 的 LinkedIn；50/50 是 beat 拆分，与“加速约一半来自 AI”的市场说法口径不同；迁移时间 10-11→6 个月与 Investor Day 新客户消费时间 10→7 个月是不同指标；TAM $300B 与 Investor Day $460B 不一致 (Brian 说不含分析和 BI)；员工数：Citi 说 Observe 180-ish、去年 950；Piper 说 170-180、去年 940，基本一致
C 2026-09-18 · TMTB (Substack) · TMTB 播客摘要：Databricks CEO Ghodsi 谈击败 Snowflake 的打法 (开放格式、AI、宣称 3 倍 TCO 优势) 2 条
文件：KB/Substack/2026-09/2026-09-18-tmtb-podcast-intel-databricks-ghodsi-on-beating-snowflake-and-staying-private-dan-niles-on-the-ai-speed-bump-p.md
发言人：TMTB Podcast Intel (由 Ethmos 研究 agent 生成的播客摘要，明示)。相关内容为 Databricks CEO Ali Ghodsi 在 Brian Halligan 播客 'Long Strange Trip' 中的说法，以及 Dan Niles (Niles Investment Management 创始人) 的观点。视角局限：只有一句话摘要，正文是图片未读到；Ghodsi 是直接竞争对手，有利益偏向。
阅读范围：全文 L1-L87 (正文简报为图片，只读到文字摘要) 可信度：仅作局部线索
Databricks 最近一轮以约 $200 亿需求募资 $50 亿；Ghodsi 的反 Snowflake 打法是开放格式、AI 以及宣称的 3 倍 TCO 优势。Dan Niles 提到 token 价格自 5 月以来下降一半、份额转向开源模型，这与 SNOW AI 产品 token 收入的通缩风险相关。
问题：正文简报为图片无法读取，只有一句话摘要
A 2026-09-24 · Acecamp · Snowflake 收入构成、AI token 计费与毛利、折扣、与 MongoDB/Oracle/Databricks 竞争、Anthropic 合作和 CoCoEvolve 11 条
文件：KB/Acecamp/2026-09/2026-09-24-bei-mei-shu-ju-hu-cang-long-tou-he-anthropiche-zuo-jin-zhan-yu-databrickde-chai-yi-hua-jing-zheng-ke-hu-cong-mongodbhe-o.md
发言人：未给出专家简介 (标注 '专家观点汇总')。从 '内部资料并没有明确写出具体比例' (L40) 推测为 Snowflake 现任或前任内部人员，偏技术岗 (熟悉 Warehouse、Cloud Service 10% 免费阈值、Horizon/Open Catalog 细节)，具体职位和年份不详。视角局限：技术视角，对财务数字 ('具体比例我没有查到') 和定价成本不掌握；对 AI 拉动 warehouse 消费有明确看法但未给数字。
阅读范围：全文 L1-L119 可信度：需交叉验证
专家拆解 Snowflake 收入为 Warehouse 计算 (主体)、存储、Cloud Service (不超过 warehouse 10% 免费) 和 AI token。AI 使用会同时增加 token 费和 warehouse 节点用量：大模型在数据库中跑 SQL、用户反复调整 prompt，比人工写精准 SQL 更费 credit，这是第 (1) 层消费的具体机制。毛利率收缩源于 AI 推理和云基础设施成本，credit 价格一般不调整；折扣无大的调整。MongoDB 迁到 Snowflake 只是零星个案、Oracle/Teradata 迁移不多，主要对手是 Databricks；客户多选 Claude 模型；Snowflake 向 Anthropic 支付的成本比例内部资料也没写。
问题：无专家简介，身份为推测；'原生模型 Architect' 应为 Arctic，转录错误；模型名 'FABLE 5.1' 无法核实，可能是转录问题；标题提到 Open Catalog sunset，但正文未回答其影响
B 2026-10-09 · GuidePoint · 前 Qlik 专业服务负责人 (现任职 Snowflake)：Snowflake 等平台用 AI + 按用量计费吞并 ETL/目录/BI 点工具；Iceberg 让数据目录成为争夺焦点 4 条
文件：KB/GuidePoint/2026-10/2026-10-09-qlik-where-can-this-data-governance-and-analytics-provider-continue-to-deliver-premium-value-in-an-ai-focused-analytics.md
发言人：Frank Hernandez，前 Qlik 美洲区专业服务 Practice 与解决方案负责人 (约 2018 年至 2025 年底，明示 L37/L214-216)；主持人确认其 2026 年起加入 Snowflake 并任职至今 (L214-216，具体职位未说明)；25 年以上技术经验，约 20 年数据相关 (L44-45)。视角局限：现任 Snowflake 员工，对 Snowflake 的正面评价有明显利益偏向 (如 'every day I wake up and I'm super excited' L504)；访谈主题是 Qlik，Snowflake 内容是侧面。
阅读范围：仅检索 Snowflake/Iceberg/Databricks 关键词段落 (L205-L218、L290-L375、L450-L515、L550-L590、L705-L718) 可信度：仅作局部线索
支持'平台吞并点工具'：Snowflake 提供管道、目录、数据质量和洞察，按用量收费，客户无需单独许可 ETL/BI/目录，企业在压缩重叠技术；SI 围绕 Snowflake 建 COE 是增长信号。同时指出 Iceberg 互操作让'谁赢得数据目录'成为新战场——这正是 Iceberg 分流计算的机制。注意专家现任职 Snowflake。
问题：专家现任职 Snowflake，存在利益冲突；Snowflake 内容是 Qlik 主题访谈的侧面，偏定性
B 2026-10-13 · Acecamp · Databricks 调研：与 Snowflake 竞争优劣、Snowflake 大数据量客户成本阶梯跳升导致流失、AI 时代数据增长与云合作 7 条
文件：KB/Acecamp/2026-10/2026-10-13-databricksdiao-yan-shu-ju-yun-ping-tai-cspji-da-mo-xing-gong-si-zai-aian-quan-fang-mian-tou-ru-dui-bi-aishi-dai-ke-hu-sh.md
发言人：未给出专家简介。从口吻 ('公司已明确决定'、'Databricks 采购的 GPU') 推测为 Databricks 现任员工，熟悉采购与区域客户 (提到中国客户、无大陆办公室)，可能与此前几篇 Acecamp 的 Databricks 亚太专家是同类来源。视角局限：竞争对手内部人员，对 Snowflake 的成本劣势说法有明显利益偏向；部分事实有误 (称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月' 上任)。
阅读范围：全文 L1-L120 可信度：需交叉验证
专家称 Snowflake 优势是易用、新 CEO 后 AI 步伐加快，Cortex 约 7,000 用户；客户在两家之间双向迁移一直存在。不利点：Snowflake 客户数据量超过 2PB 且慢查询约占 15% 时成本急剧上升 (集群从中型升大型价格翻倍)，是部分客户转到 Databricks 的核心原因，Databricks 会专门找大数据量老客户做成本 POC；AI 模型粘性最弱，换模型只需改 URL/endpoint。数据有机增长月环比 6%，非结构化数据一次性导入高峰已过。
问题：无专家简介，身份为推测；专家称 Snowflake '主要在 AWS 上售卖'、CEO '2024 年 6 月上任' (实为 2024 年 2 月)，事实有偏差；'Cortex 已拥有 7,000 个用户' 疑为 CoCo 账户数混淆；模型名 'Astra' 无法核实