2026.10.08 | 9K | GOOG | Google Q3 Channel Check

音频信息

音频标题: 261008 - Merit Google Q3业绩调研

相关标的：undefined

议题：

完整音频链接：http://mw-meeting.oss-cn-shanghai.aliyuncs.com/audio/1791433508209-261008%2520-%2520Merit%2520Google%2520Q3%25E4%25B8%259A%25E7%25BB%25A9%25E8%25B0%2583%25E7%25A0%2594_20261008_122446.mp3

音频纪要总结

专家访谈纪要

1. Google搜索广告增长与放缓原因

a) Q3搜索广告增速如何，是否低于Q2？

• 受访者披露，搜索及广告业务本季度增速约为14.2%，低于上季度约16.8%的水平，整体仍保持较强增长。
• 访谈最初问题采用同比口径，但回答中使用了“环比”表述，因此14.2%与16.8%的具体统计口径仍需进一步确认。
• 增速放缓并非源于搜索产品受欢迎程度下降，主要受到两方面影响：

1. 汇率变化产生显著负面影响。
2. Google主动降低部分AI搜索广告价格，以推动AI广告产品在美国以外市场的普及。

b) 除汇率外，Google为何主动降低AI搜索广告价格？

• Google正在降低AI Overview及嵌入Google搜索的其他AI模式广告成本。
• 对全球首次使用AI广告的广告主，Google提供约15%—20%的折扣。
• 该策略主要用于：

1. 降低海外广告主尝试AI广告产品的门槛。
2. 加快AI广告产品向全球市场拓展。
3. 提高美国以外地区广告主的采用率。

• 海外市场扩张本身又受到汇率影响，因此折扣策略与汇率变化共同拖累了短期收入增速。

c) AI搜索增长情况如何？

• AI Overview环比增速约为8%。
• AI Mode环比增速约为2%。
• 两类AI搜索产品的增速均有所放缓，但仍处于增长阶段。

2. 搜索广告加载率与AI搜索商业化

a) 常规搜索、AI Overview和AI Mode当前广告加载率分别是多少？

• 常规Google搜索：

1. 上季度广告加载率约为20%。
2. 本季度提升至约22%。
3. 长期目标为逐步达到40%。

• AI Overview：

1. 上季度约为7.2%。
2. 本季度超过8%，整体约在8%—9%。
3. 美国市场约为12%，其他地区约为7%。
4. 下一季度计划提升至约9%。

• AI Mode：

1. 上季度约为1.4%。
2. 当前披露口径约为2.5%—3%。
3. 年底计划提升至3.5%—4%。
4. 长期目标约为20%。

b) AI Overview广告渗透率包含哪些广告位置？

• AI Overview上方和下方的广告均计入其广告渗透率。
• 当前约8%的渗透率并非仅指AI生成内容内部的广告位，而是包含AI Overview上下方的广告展示。

c) Google为何不快速提高AI搜索广告加载率？

• Google不希望为了单季度业绩而快速提升广告加载率，避免短期收入突然增长80%或100%后难以持续。
• 管理层更希望未来4—5年的广告收入保持稳定、持续增长，而非出现一次性大幅跃升。
• 广告加载率提升需要兼顾：

1. 用户体验。
2. 客户反馈。
3. 新广告形式的成熟度。
4. 广告主接受程度。

• Google正在推进UCP、购物广告及AI Max等新产品，因此没有必要通过激进提高加载率来提前变现。

d) 后续广告加载率将如何提升？

• Q4作为重要购物季，Google计划额外提升约2个百分点。
• 受访者提到，2026年Q4常规搜索广告加载率目标约为24%。
• 此后计划每季度提升约1.5个百分点，预计后续年度末达到约30%。
• AI Overview将比常规搜索更积极地提升广告加载率。
• AI Mode则继续采取渐进式提升方式，避免影响用户体验和投资者预期。
• 访谈中关于AI Overview“明年上半年达到4%、下半年达到15%”的表述与当前超过8%的数据存在冲突，可能存在转录或口径问题，需进一步核实。

3. AI Mode用户渗透率与搜索使用频次

a) AI Mode目前占Google搜索总查询量的比例是多少？

• 以Google搜索全部用户为基础，约34%的用户每月至少使用一次AI Mode。
• 在这些用户中，约12%的搜索通过AI Mode完成。
• 按34%×12%计算，目前约4%的Google搜索查询通过AI Mode完成。
• 上季度该比例约为3.47%，本季度提升约0.5个百分点。
• 相对增幅约为14%—15%，在低基数下增长较为明显。

b) AI Overview和AI Mode的用户规模分别是多少？

• 全球AI Overview用户规模约为26.8亿人。
• 全球AI Mode用户规模约为12亿人。
• 美国AI Overview用户约为2.8亿人。
• 美国AI Mode用户约为1.14亿人。
• 美国用户占AI Mode用户的比例约为10%—12%。

c) AI Overview和AI Mode的人均每日使用频次如何？

• AI Overview：

1. 全球用户人均每日使用约2.1次。
2. 美国用户人均每日使用约3.6次。

• AI Mode：

1. 全球用户人均每日使用约0.554次。
2. 美国用户人均每日使用约0.88次。

• AI Mode使用频次明显低于AI Overview，主要因为AI Overview会在搜索结果中直接展示，而AI Mode需要用户主动触发并进入。

4. AI Max覆盖范围与广告收入贡献

a) AI Max是否正在扩大查询词覆盖范围？

• AI Max正逐步覆盖更多模糊查询、长尾搜索词以及广告主未主动选择的关键词。
• 受访者披露，AI Max的采用率约为7%，另有“同比提升至13%”的表述，具体是采用率达到13%还是同比提升13%，仍需确认。
• 对广告主未选择的长尾关键词，AI Max的覆盖范围提升约15%。

b) AI Max将为Google广告收入带来多大增量？

• 本季度预计AI Max可推动广告收入增长约1.5%。
• 明年保守预计推动广告收入增长约2.3%。
• 乐观情况下，贡献幅度可达到约2.8%。
• 综合预计，明年AI Max对广告收入的提升约为2.3%—2.8%。

5. Google个人AI Agent产品规划

a) Google是否计划推出个人AI Agent？

• Google正在开发与Meta旗下Muse类似的个人AI Agent产品。
• 该产品不会主要依赖用户单独下载安装，而是将嵌入：

1. Android。
2. Chrome。
3. Google各操作系统内核。
4. Search、Gmail、YouTube、Maps、Photos及Gemini等产品。

• 用户仅需选择开启或关闭，即可使用相关能力。
• Google将利用现有分发网络，降低个人Agent的获客和安装门槛。

b) Google个人Agent的模型和产品能力如何？

• 个人Agent将由Pro 4支持，同时Google正在开发4.5和5版本。
• 在B2C、任务执行和生产力场景中，Google认为自身产品具有较强竞争力。
• 在与Muse的基准测试中表现较好，有望从Muse手中争取用户。
• 但在B2B编程场景中，Google当前能力仍弱于主要竞争对手。

c) 个人AI Agent会对传统搜索广告产生多大冲击？

• 智能体可代替用户完成的购买，初期主要集中在：

1. 旅行。
2. 本地消费。
3. 重复购买。

• 上述支出约占个人总支出的28%。
• 在这部分交易中，受访者预计由智能体完成的比例为：

1. 2027年约3%。
2. 2028年约8%。
3. 2029年约14%。

• Google认为，对现有搜索广告收入的最大影响约为3%—4%。
• 由于智能体、UCP及其他新型广告市场也会带来新增收入，因此预计对Google整体收入不会产生实质性负面影响。
• 访谈中将“约30%的可代理支出×3%的智能体渗透率”计算为总支出的9%，该计算与常规数学结果不一致，相关预测口径需进一步确认。

6. 个人AI Agent的广告与商业化模式

a) 个人Agent是否会扩大广告市场？

• Google认为广告市场不会消失，而会演变为新的广告形态。
• 智能体可能跳过传统搜索、比价和交易漏斗，直接完成商品购买。
• 传统竞价搜索广告可能减少，但对话式推荐、直接报价及任务执行过程中的广告将形成新的广告空间。

b) Google正在开发哪些Agent广告形式？

• Google正在开发超过20种适用于个人AI Agent的新型广告，包括：

1. 购买后关联商品推荐。
2. 任务执行过程中的广告展示。
3. 基于用户偏好的个性化推荐。
4. 推荐算法中嵌入广告主提供的商品信息。
5. 与商家协商形成的专属折扣和直接报价。
6. 基于用户潜在需求的预测式广告。

• 例如，用户购买健身鞋时，智能体可进一步推荐瑜伽鞋、维生素或其他健康产品。
• 此类内容可能不会以传统“广告”标签和展示形式出现，而是融入智能体的推荐结果。

c) Google是否会对Agent促成的交易收取佣金？

• Google已取消对相关交易收取佣金的初始计划。
• 取消佣金的目的包括：

1. 吸引更多品牌和商家加入Google平台。
2. 推动商家绕过Amazon、Rakuten、Shopee、Lazada及AliExpress等平台。
3. 让消费者直接与供应商交易。
4. 扩大Google可变现的广告空间。

• Google计划邀请约100万个品牌入驻，商家通过Google平台销售商品无需支付交易佣金。
• 平台将主要通过广告、推荐信息收费及直接报价等方式变现。

d) 还有哪些新增广告入口？

• Google Maps将逐步成为直接销售渠道，用户可在地图中一键购买商品。
• AR眼镜和智能眼镜将形成新的广告展示和交易入口。
• 用户可通过注视商品直接触发购买。
• Google Workspace还可销售数据库及其他企业解决方案。
• Google预计从2028年开始，将出现更多新型广告形式。

7. Google与Meta在个人Agent领域的竞争格局

a) Google的主要优势是什么？

• Google可整合Search、Gmail、Android、Chrome、YouTube、Maps、Photos和Gemini等产品。
• Google拥有约20项日活跃用户规模超过10亿的服务。
• Google正在构建统一数据湖，并为用户形成数字化分身，包括：

1. 心理测量数据，用于理解用户想法。
2. 行为指标，用于预测用户行为及动机。

• Google虽然不掌握完整的人际关系网络，但能够通过邮件、日历、位置、搜索和视频等信息了解用户正在做什么。
• Google的更多产品界面都可成为Agent及广告触达入口。

b) Meta的主要优势是什么？

• Meta在Instagram、Facebook及WhatsApp等社交产品中具备关系链优势。
• Meta更擅长分析：

1. 人际关系如何影响用户。
2. 兴趣群体和同伴压力如何影响消费决策。
3. 社交媒体内容如何改变用户行为。

• Meta在直播电商方面也具备优势，可将直播电商与智能体结合，推动商品销售。

c) 决定个人Agent长期胜负的关键因素是什么？

• 全球客户入口和分发能力，权重约占1/3：

1. 是否拥有操作系统。
2. 是否拥有全球在线平台。
3. 能否在用户需要时直接触达用户。

• 包含商家生态的端到端Agent流程，权重约占1/4。
• 信任、安全和运行时模型能力，权重约占25%。
• 规模经济和成本效率构成剩余关键因素：

1. 是否拥有自研软硬件。
2. 是否能降低推理、推荐和交易成本。
3. 是否能以更低成本快速扩张。

• Google表示，依靠TPU，其边缘侧、推荐及交易环节的成本比Meta低约68%。

8. Gemini收入增速、开放节奏与竞争力

a) Gemini各业务板块的收入增速如何？

• API业务：

1. Q2增长约38%。
2. 受开放延迟影响，Q3预计增长约24%。
3. 随开放范围扩大，Q4预计增长约45%。

• 企业客户业务：

1. Q2增长约40%。
2. Q3预计增长约30%。
3. Q4预计增长约58%。

• 订阅业务：

1. Q2增长约12.5%。
2. Q3预计增长约10%。
3. Q4预计增长约15%。

b) Gemini 4何时全面开放？

• 当前主要面向政府机构和网络安全合作伙伴。
• 11月计划向API客户开放。
• B2C客户中，初期仅AI Ultra用户可以使用。
• 预计到2027年1月向所有用户开放。
• 从本季度末开始，Google将逐步向更多用户开放。

c) Gemini 4为何推迟发布？

• 主要原因是计算能力不足。
• Google计划从xAI借用价值约10亿美元的算力。
• 随xAI算力到位以及台积电生产的TPU V8开始安装，算力约束将逐步缓解。

d) Gemini 4的能力和市场定位如何？

• 企业Agent和任务执行能力预计达到顶尖水平，略优于Astra。
• 数学、金融和科学分析能力预计达到顶尖水平，与Astra大体相当。
• 编程能力仍相对落后：

1. 若Opus 5.5或Astra为100%，Gemini 4约为90%。
2. 仍存在约10%的差距。

• B2C、生产力及任务执行场景将出现明显提升。
• B2B市场预计仍由Claude和ChatGPT占据主导地位。

e) Gemini企业市场份额将如何变化？

• 当前企业市场份额约为12%。
• Gemini 4进入企业市场后，预计提升至约18%。
• 增长主要来自与Google Workspace、Google Ads、数据库、GCP及大模型能力的整合。
• 编程能力仍是Gemini 4.1需要重点改善的方向。

9. Gemini付费订阅用户规模

a) 当前Gemini付费用户有多少？

• 仅统计B2C市场，受访者披露当前付费订阅用户约为2,240万。
• 同时披露Q3付费订阅用户为2,310万，并表示“未实现增长”。
• 当前值与Q3值及增长表述之间存在一定矛盾，需进一步确认统计时点和口径。

b) Gemini 4开放后用户规模将如何变化？

• Gemini 4向大多数用户开放后，预计付费订阅用户将明显增长。
• 预计明年Q1付费订阅用户达到约3,000万。
• 当前付费转化率约为2%。
• 按2%的转化率估算，非付费用户规模约为付费用户的50倍。

10. Google资本开支与算力扩张计划

a) 未来资本开支将如何增长？

• 今年资本开支预计约为1,950亿—2,050亿美元。
• 2027年保守批准的最低资本开支预算约为2,550亿美元。
• 若年中达到特定KPI，可能额外追加约400亿美元。
• 潜在资本开支上限可达到约2,950亿美元。
• 追加支出将主要用于：

1. 应对存储价格上涨。
2. 扩大TPU及相关算力部署。
3. 通过博通、联发科和Marvell获取更多台积电产能。

• 2027—2028年资本开支预计仍将持续增长，不会收缩。

b) Google如何缓解先进芯片和封装产能限制？

• V8和V9将引入更多合作伙伴。
• V9将与博通合作采用EMIBT先进封装。
• 该方案有助于缓解台积电产能限制，提高芯片按时交付能力。
• Google预计到2028年，整体算力限制将有所缓解。

11. TPU V8交付、分配与收入贡献

a) TPU V8交付节奏如何？

• Q3获得约5.5万颗V8。
• Q4预计获得约30万—35万颗。
• 明年上半年预计获得约300万颗。
• 明年下半年预计获得约400万颗。
• 今年V8供应量占整体规划比例较小，主要供应增量将在明年释放。

b) TPU V8性能相比V7提升多少？

• V8峰值计算能力约为V7的5倍。
• 随V8大规模部署，Google的训练和推理算力压力将明显缓解。

c) Q3和Q4的V8如何分配？

• Q3约5.5万颗V8中：

1. 约3.8万颗用于Google内部。
2. 约1.2万颗承诺提供给Anthropic。
3. Anthropic部分采用芯片和系统直接交付的租赁形式，不部署在GCP内。

• Q4约30万—35万颗V8中：

1. 约25万颗预留用于内部。
2. 其余约10万颗提供给Anthropic和Meta。

d) 内部V8主要用于训练还是推理？

• 当前内部V8约22%用于训练，78%用于推理。
• 后续训练占比预计稳定在15%—18%。
• 推理占比将逐步提升至约85%。

e) TPU销售与租赁收入贡献如何？

• Q3 TPU销售收入约为10亿美元。
• Q4预计达到30亿美元或以上。
• 客户付款通常存在60—90天账期，因此收入确认会存在时间差。
• 今年Q4由V8带来的云收入增量最多约为3亿美元，主要原因是芯片安装到服务器需要约60天。
• V8的大部分收入和利润贡献将在2027年体现。
• 若2027年获得约400万—600万颗V8，预计可贡献约180亿美元收入。

12. TPU V8与NVIDIA Rubin定价及利润率

a) TPU V8的租赁价格如何？

• V7按需租赁价格约为每小时12美元。
• V8按需租赁价格计划为每小时21美元。
• V8承诺使用协议折扣为：

1. 1年期折扣约30%。
2. 3年期折扣约55%。
3. 5年期折扣约72%。

• 多数客户采用3年期合同。

b) TPU V8的成本和利润率如何？

• V8实际运行成本低于每小时3.7美元。
• 按3年期合同计算，价格约为每小时9.45美元。
• 即使按约3.8美元成本测算，利润率仍较可观。
• Google预计V8将持续提升GCP利润率。

c) NVIDIA Rubin的价格与折扣如何？

• Rubin按需租赁价格约为每小时13美元。
• 由于Rubin为外部采购产品，成本高于自研TPU。
• Rubin长期合同折扣为：

1. 1年期约20%。
2. 3年期约35%。
3. 5年及以上约50%。

• Rubin折扣幅度明显小于TPU。

d) 按需租赁和长期协议的结构如何？

• 按需使用约占12%。
• CUD长期承诺使用协议约占88%。
• 在CUD中，约58%—60%为5年期合同。
• 3年期和1年期合同合计约占CUD的40%。
• 5年期合同多采用照付不议模式，即使客户放弃使用，也需履行付款义务。
• 该结构反映客户对算力需求较为迫切，希望锁定长期供应。

13. Google Cloud与Workspace业务表现

a) Google Cloud整体及非AI业务增速如何？

• Q3云业务整体增长约97.8%。
• 传统、非AI GCP业务增速约为25%。
• AI数据中心业务不仅直接贡献收入，也带动了传统CPU和非AI云平台需求。
• 因此，非AI GCP业务仍保持较快增长。

b) GCP积压订单情况如何？

• 受访者披露，Google Cloud的RCP主要对应客户订单和积压订单。
• 访谈中先后出现“超过5,500亿美元”和“超过5.5亿美元”两种口径，差异较大，需确认准确数值。
• V8投入使用后，Google预计将加快消化相关积压订单，并推动收入确认。

c) Workspace收入增速如何？

• Q3 Workspace收入增长约15%，是Google各主要产品领域中增速较慢的业务。
• 上季度收入较快增长，主要由于：

1. 提高每席位价格。
2. 要求用户采用整合AI能力的产品。

• 本季度价格上涨影响已逐步体现，主要增长动力转向企业付费席位数量增加。
• 大型企业和普通企业席位未出现明显增长。
• 中小企业认为AI捆绑方案更具价值，因此成为Workspace席位增长的主要驱动力。

14. 关键数据与口径待确认事项

a) 搜索广告增速口径

• 提问采用同比口径，但回答使用“环比”表述。
• 14.2%与16.8%究竟为同比增速、环比增速还是固定汇率口径，需进一步确认。

b) AI Overview后续加载率

• 当前广告加载率已超过8%，但后续预测中出现“明年上半年达到4%”的表述，可能为“14%”或其他口径。

c) AI Mode当前加载率

• 访谈中分别出现2.5%和3%的当前口径，可能与统计时点或全球/区域口径不同。

d) Gemini付费用户

• 当前约2,240万与Q3约2,310万、同时“未实现增长”的表述存在冲突。

e) Agent交易规模计算

• “30%的可代理支出×3%的智能体渗透率”按常规计算应约为总支出的0.9%，与访谈中的9%不一致。

f) Google Cloud RCP

• 访谈中出现超过5,500亿美元和超过5.5亿美元两种说法，需要确认单位及准确金额。

结构化总结表格

新版prompt提炼之后的转译

提问者：在访谈开始前，请确认您了解保密信息的定义，并同意遵守相关公司的保密规定和协议，不会在访谈中透露任何此类信息或其他保密信息，也不会将本次会议披露的信息提供给任何第三方用于商业用途。

回　答　者：我同意。

提问者：我们关注到，Q1谷歌搜索广告的同比增速为19%，Q2的同比增速为17%。Q3的同比增速高于还是低于Q2？背后的主要原因是什么？

回　答　者：关于谷歌AI搜索，目前主要包括AI Overview和AI Mode，其增长率有所下降。AI Overview的环比增速约为8%，AI Mode的环比增速约为2%。整体来看，增长率略有放缓，但仍保持较强增长。

提问者：您刚才提到的是AI搜索的环比增长情况。我的问题是搜索广告整体的增长情况。Q2搜索广告整体同比增长17%，Q3搜索广告整体的同比增速是高于还是低于17%？具体情况如何？背后的主要因素是什么？

回　答　者：搜索广告实际上……

提问者：除汇率影响外，还有哪些因素导致谷歌广告业务增速放缓？此前我们沟通时，您认为谷歌会在Q3加大AI搜索广告的广告加载率，市场也普遍预期谷歌搜索广告将继续增长。这方面目前有何最新进展？

回　答　者：环比来看，搜索和广告业务实际增长14.2%，仍保持强劲势头。增速确实有所放缓，上季度约为16.8%，本季度约为14.2%。这并不是因为搜索和广告的受欢迎程度下降，主要原因是汇率产生了非常负面的影响。整体而言，我们仍保持持续增长态势。

除汇率变化外，另一个因素是我们采取了降低搜索广告成本的策略，尤其是AI概览广告，以及嵌入谷歌搜索的其他AI模式广告，以推动这些产品向全球市场拓展，并提高美国以外地区广告主的采用率。这是一项战略举措，目的是让全球其他地区的广告主能够以可负担的价格尝试我们的广告产品。

我们面向全球首次使用AI广告的用户提供了约15%至20%的广告折扣，并因此更积极地向全球其他市场扩张，而这部分业务也受到了汇率影响。上述2项因素共同导致了增速放缓。

回　答　者：目前广告加载率接近22%。AI Overview的广告加载率已提升至约8%。我们的目标是达到40%，目前仍在逐步推进。AI Mode的广告加载率约为2.5%，目标是达到20%，目前也仍在提升过程中。

我们并不急于立即变现，而是在客户体验与新形式下广告加载率的提升之间寻求平衡。目前，我们也在推进UCP，并将购物广告引入UCP，因此会有新的广告形态。我们正在开发更多广告产品，AI Max的表现也很好，没有必要因急于变现而影响客户体验。我们会逐步扩大覆盖范围。

提问者：上个季度，Google常规搜索、AI Overview和AI Mode的广告加载率分别是多少？

回　答　者：上个季度，Google常规搜索的广告加载率为20%，本季度为22%，目标是逐步达到40%。AI Overview上个季度的广告加载率约为7.2%，本季度已超过8%。AI Mode上个季度的广告加载率约为1.4%，目前为2.5%，整体仍在提升。

提问者：目前，AI Mode查询量占Google搜索总查询量的比例是多少？相比上个季度增长了多少？

回　答　者：目前这一比例仍然较小，因为用户需要主动开启AI Mode后再进行搜索。可以从3个方面来看。以Google搜索的全部用户为基础，目前约有34%的用户每月至少使用1次AI Mode，即按月活跃用户口径计算。对于这些用户，约有12%的搜索最终通过AI Mode完成。因此，34%×12%，意味着目前约有4%的搜索通过AI Mode完成。

这一数字环比大幅增长。由于目前的基数很小，环比约增加了.6。

回　答　者：约为15%。与上季度相比，增幅约为0.6%，上季度约为3.47，因此广告加载率大致改善了0.5%。从具体数值来看，0.5除以3.5，增幅接近14%至15%，这是一个显著提升。

提问者：您预计Q4不同类型AI搜索的广告加载率将达到多少？明年AI搜索的广告加载率可能增长到多少？什么时候或哪个季度会出现较为明显的加速？

回　答　者：CEO已经明确表示，不希望因某个季度提高广告加载率，导致业绩突然增长80%或100%。他希望未来4至5年的广告收入能够在未来4至5年内保持稳定增长，而不是出现大幅跃升。因此，不会有某个季度的广告加载率突然超预期增长，我们不希望采取英伟达式的增长方式。

Q4是重要的购物季，我们计划额外提升2%。2026年Q4的目标是达到24%。2027年每个季度计划提升约1.5%，这意味着到明年年底，Google搜索整体的广告加载率将达到约30%。

AI Overview的情况有所不同，我们会加快提升速度。明年上半年，AI Overview的广告加载率将达到4%；明年下半年，我们计划将其提升至15%。由于AI Overview目前的广告加载率较低，我们会更积极地推动增长。

AI Mode方面，上季度的广告加载率将继续维持在约3.5%；2027年年底的目标约为5.5%。整体而言，不会出现剧烈变化，而是每个季度逐步改善，以避免令投资者失望，确保每个季度都保持小幅、持续的提升。

提问者：关于AI Max，Q3是否已经扩展至更多模糊搜索关键词，包括AI Mode中的相关关键词？未来几个季度，AI Max对Google广告业务的推动是否会加速？

提问者：AI Max 是否正在扩展至更多模糊搜索结果或查询，并更多地应用于 AI 模式？此外，您认为下季度 AI Max 对 Google 搜索广告总收入的贡献会达到多少？

回　答　者：从查询类型的覆盖范围来看，AI Max 的采用率约为 7%，与去年同期相比已提升至 13%，接近翻倍。此外，在选择广告主未选取的长尾搜索词方面，目前覆盖率提升了约 15%，即 AI Max 对未选关键词的覆盖范围有所扩大。

从预期来看，本季度 AI Max 预计将推动广告收入增长 1.5%。明年，保守估计将推动广告收入增长约 2.3%，乐观预期为 2.8%。因此，我们预计明年 AI Max 对广告收入的提升幅度约为 2.3%至2.8%。

提问者：我们关注到 Meta 的 Mews 个人 AI Agent 最近非常火爆。Google 是否计划推出自己的个人 AI Agent 产品？如果有，预计何时推出？此外，Mews 这类个人 AI Agent 产品会对整个搜索广告市场产生怎样的影响？目前，用户连接互联网的首要入口依然是 Google 搜索框。如果未来用户更多地利用 AI Agent 执行电商、旅游、住宿等具有商业意图的查询或完成相关任务，是否会给 Google 的高价值查询带来较大的竞争压力？

回　答　者：是的，我们正在开发一款与 Mews 类似的产品。不过，我们计划将其整合到现有的分布式应用中，因为我们已经具备优势。

提问者：Gemini应用目前在不同设备上的安装量是多少？

回　答　者：我们正在集成Gemini，相关营销名称仍在确定中。我们的个人智能体将嵌入Android和Chrome，并进一步嵌入我们所有操作系统的内核。因此，用户无须安装，只需选择开启或关闭即可使用。我们将充分利用现有的强大分发网络。

从产品能力来看，我们的Pro 4表现强劲，相关产品由Pro 4提供支持，同时我们也在并行开发4.5和5。我们的智能体在与Muse的基准测试中表现出色。因此，在B2C领域，我们具备较强的竞争力，有望从Muse手中赢回客户；这与B2B领域不同，我们在编码方面的表现较差。此外，我们还具备Gemini iOS应用带来的便利性，以及Android操作系统内核级集成的优势。

如果将整个广告市场视为100%，未来5年，智能体能够代替用户完成的购买将主要局限于旅行、本地购买和重复购买，这些支出约占个人总支出的28%。由于人们通常不愿尝试新事物，市场增长会较为缓慢。我们预计，在这28%的交易中，2027年约有3%将由智能体完成，2028年约为8%，2029年约为14%。

如果将28%取整为30%，那么2027年约有9%的总支出可能由智能体完成。我们认为Muse可能占据主导地位，但有信心保留约一半的市场。按一半计算，智能体购买可能给我们带来的广告收入损失风险上限为4.5%。

不过，我们认为，由于智能体和UCP的发展，整体广告市场也在增长。因此，2027年潜在的4%至5%损失，将被其他广告形式的大幅增长充分抵消，整体收入不会受到影响。2027年，确实可能有这9%中的一半对我们的收入产生影响，但假设我们在智能体购买市场仅占50%的份额，是非常保守的。

到2028年和2029年，我们认为这一市场的发展会类似于Google搜索，相信可以将市场份额重新提升至78%。在这种情况下，我们损失的是最高14%中的22%，即不到3%至3%。因此，我们收入面临的最大风险约为3%。

回　答　者：对现有搜索广告收入的影响最高为3%至4%，这还不包括未来将出现的、扩展后的不同广告市场。因此，对整体收入没有影响，对原有搜索广告收入的影响最高为3%至4%。

提问者：随着个人AI智能体的加入，您认为整体搜索广告市场或广告空间会扩大还是缩小？过去，谷歌的广告收入主要来自搜索竞价排名，但这种机制在个人AI智能体的搜索和决策链路中是否依然有效，目前尚不明确。此外，Meta目前尚未对其AI智能体产品进行太多直接变现，可能仅收取部分订阅费。如果谷歌推出自己的个人AI智能体，主要会通过哪些方式实现变现？

回　答　者：广告市场将演变为不同的形式。当用户使用智能体时，可能会完全跳过搜索和交易流程中的传统漏斗环节，直接购买智能体推荐的产品。

在这种情况下，谷歌智能体会在用户购买某种产品时，展示可以与该产品一同购买的其他商品；也会在为用户执行任务的过程中展示广告。我们正在开发超过20种适用于个人AI智能体的新型广告。这些广告形式将取代以往的广告形式，并帮助我们继续保持优势。

未来还会出现更加个性化的广告。例如，当用户购买健身鞋时，智能体可能会推荐瑜伽鞋；如果我们了解到用户希望购买更健康的产品，也可能推荐有助于增强肌肉的维生素等。

回　答　者：广告将由与智能体合作销售产品的商家提供支持。用户在线等待时，商家会付费向其展示广告，这些收入将用于弥补智能体带来的损失。

我们认为，客户会要求我们提供产品购买建议。如果我们提供咨询并促成某款具体产品的购买，就会在推荐算法中嵌入广告。此类内容不会以广告形式呈现，但推荐结果中会包含广告主提供的信息。将产品信息、产品优势等内容嵌入助手后，我们会向广告主收费。这是我们目前正在探索的一种新型广告形式。

现阶段变化很大，很难作出预测。不过，我们已经看到广告主对这2种广告形式表现出浓厚兴趣，并愿意投入大量资金。我们认为，这些收入足以弥补跳过中间流程、展示广告及相关环节所造成的损失。

此外，我们正在将 Google Maps 打造成销售渠道，用户可以直接通过 Google Maps 订购产品。佩戴我们的 AR 眼镜时，用户只需注视某件商品即可直接下单。AR 眼镜、智能眼镜等新型 AR 设备也将提供新的广告空间。Google Workspaces 将销售数据库或其他解决方案；用户还可以通过 Google Maps 一键购买商品，同时查看所有广告。我们认为，这些新型广告形式足以弥补传统广告形式造成的任何损失。

在商业化方面，Muse 目前采用订阅模式。Muse 团队中有一些来自 Meta 的人员加入了 Google Gemini 团队，而且每周都有几位。根据我们了解到的信息，Muse 在 2027 年年底前并不打算加入广告，但计划从 2028 年开始大幅拓展广告业务。Meta 本身是一家广告公司，Instagram 和 Facebook 中随处可见广告，因此广告业务必然会到来。

我们也会采取同样的方式，为个人智能体开发新型广告，包括推荐式直接报价。例如，智能体可以向用户表示：“这里有一款正在打折的商品，您是否愿意购买？”当我们能够预判用户需求时，就可以与卖家协商，为特定产品提供折扣价格，促成销售并从中收取佣金。因此，我们认为，个人智能体将拥有更先进的对话式广告、基于推荐算法的广告，以及通过与卖家协商形成的直接报价。

回　答　者：对方是买方。我们正在进行的一项重大调整是，不再对任何交易收取佣金。收取佣金原本是我们的初始计划，但我们取消了这一安排，以吸引更多商家加入并使用Google平台，避开其他电商平台。消费者可以跳过Amazon以及Rakuten、Shopee、Lazada、AliExpress等电商平台，直接与供应商交易，并在我们的平台上一键购买商品。这意味着我们有很大的空间加入不同类型的广告。可以预期，从2028年起，将出现许多新的广告形式。

提问者：如果Google和Meta开发个人Agent产品，这2家公司各自具备哪些优势？从长期来看，个人Agent产品最重要的竞争壁垒或决定最终胜负的关键因素可能是什么？为什么？

回　答　者：竞争壁垒具体是指什么？

提问者：也就是说，从长期来看，个人Agent产品未来的胜负手可能是什么？为什么这可能成为整个Agent领域最重要的因素？

回　答　者：对于Google而言，我们将充分利用Google Search、Gmail、Android、Chrome、YouTube、Google Maps和Google Photos。我们拥有20项服务，其日活跃用户超过10 billion。我们正在这些服务背后构建一个整合的数据湖，并以此为每位用户建立数字化分身，形成心理测量数据和行为指标。前者用于了解用户的想法，后者用于预测用户的行为及其动机。

我认为，这种优势大于Meta旗下Instagram或Facebook的优势。Meta的优势主要集中在社交领域，而随着成年用户年龄增长，他们会逐渐减少对社交产品的使用。

Meta在基于人际关系的兴趣分析方面更具优势，因此能够预测他人如何影响某位用户，以及用户通过社交媒体看到和体验的内容会如何影响其行为。Meta举办的直播电商也会对用户产生影响。基于这些理解，Meta可以进一步优化算法，推送和销售特定广告及产品，这是其主要优势。

Google虽然不掌握用户的人际关系，但了解每个人正在做什么。我们可以查看用户的日历、邮件以及Google Maps中的位置信息，并据此利用这些信息。

回　答　者：我们可以通过算法预测用户需求，并提供相应的商品。Meta目前不具备而我们具备的能力，是我们已经建立了GCP，并计划在取消佣金后邀请约1,000,000个品牌加入。商家在我们的平台上销售商品将完全不需要支付佣金。

如果Google智能体覆盖Search、Gmail、Android、Chrome、YouTube、Google Maps、Google Photos和Gemini，那么我们的每个产品界面都可以成为广告展示入口，在用户需要推荐的特定时刻展示广告。因此，我们能够广泛触达用户。Meta可以在用户使用WhatsApp、Instagram或Facebook时触达他们，但我们覆盖的产品界面更多。

我们还可以利用个人心理测量和行为预测，在恰当的时间、地点进行推荐，推动销售和广告业务。我们认为，这将成为Google的独特优势。我们拥有商家和客户，了解客户，也知道应在何时进行推荐。商家则可以绕过整个购物平台，直接在Google平台上销售商品，这将成为Google的一项突出优势。

Meta在所有关系驱动型模式上更具优势，也会利用同伴压力和兴趣群体推动销售。此外，Meta在直播电商方面也具有优势，可以将直播电商与智能体整合，促进商品销售。我们认为，这些将分别构成双方的独特优势。

不过，最终决定胜负的关键之一，是谁能够触达客户，即是否通过操作系统或其他在线平台在全球范围内拥有面向客户的入口。我们认为，这一因素约占1/3。

第2个因素是，是否拥有包含商家生态系统的端到端智能体流程。我们认为，这一因素约占1/4。

另外25%取决于基于信任与安全的运行时模型，即智能体是否足够可信、安全，能够投入使用。我们认为，这是另一个占比25%的主要因素。

最后一个因素是规模经济。要构建如此庞大的系统，让每个人都能使用智能体、看到广告并购买商品，能否凭借自有软硬件降低服务成本至关重要。我们拥有TPU，按边缘侧、推荐和交易环节计算，目前成本比Meta低约68%。

这将成为系统和平台持续扩张过程中的关键因素。在资金有限的情况下，如果成本更低，就能更快扩张并获得更多收入，进而将收入继续投入平台扩张。因此，成本效益将成为决定扩张速度的关键因素。

提问者：我们关注到Gemini 4已经发布。请问Q3谷歌Gemini的收入、ARR及其同比和环比增速大概如何？目前Gemini 4仅向政府机构和网络安全合作伙伴开放，尚未全面向订阅用户开放。您预计何时会正式开放？全面开放后，Gemini的收入是否会明显加速增长？

回　答　者：从ARR来看，我们主要分为API、企业客户和订阅业务3个板块。

API方面，由于我们推迟了向API开发者开放，预计Q3的API收入增长约24%。随着开放范围扩大，预计Q4将增长45%，增速将进一步加快。Q2的API收入增长约38%。鉴于Gemini 4的性能表现，预计将吸引更多用户使用。

企业客户方面，Q2增长40%。受延迟发布影响，预计Q3增长30%，Q4则有望增长约58%。

订阅业务方面，Q2增长12.5%。受延迟发布影响，Q3预计增长约10%；随着新功能推出，Q4预计增长约15%。

目前，我们主要与政府机构、网络安全合作伙伴等开展合作。11月，我们将向API客户开放Gemini 4；B2C客户中，只有AI Ultra客户可以使用。到2027年1月，我们将向所有用户开放。

推迟发布的主要原因是计算能力限制。我们正在向埃隆·马斯克旗下的xAI借用计算能力。由于Grok并未使用全部计算能力，我们将从xAI借用价值约1 billion美元的计算能力。随着这些计算单元陆续到位，台积电生产的TPU v8也将开始安装。

回　答　者：从本季度末开始，我们将能够向更多用户开放。不过，我想适当管理市场预期。我已经试用过该产品，也参与了相关基准影响分析。在服务企业的智能体能力方面，例如任务执行能力，将达到顶尖水平，实际上略优于 Astra。在数学、金融、科学等领域的分析能力方面，也将达到顶尖水平，整体表现与 Astra 相当。

但需要坦率说明，我们在编程能力方面仍然落后。虽然已经取得显著进步，但与 Opus 5.5 或 Astra 相比，如果它们的水平是 100%，我们目前最多达到其 90%，仍有约 10% 的差距。因此，编程能力仍是需要大幅改进的领域。

在其他领域，尤其是 B2C、任务执行和生产力相关领域，用户将看到明显提升。在 B2B 市场，Claude 和 ChatGPT 将继续占据主导地位。我们目前的市场份额约为 12%，预计这一情况不会发生剧烈变化。不过，凭借与 Google Workspaces 的集成，我们仍将获得更多市场份额。现在，我们正将 Google Ads、Google 数据库、GCP 和 LLM 等 Google 产品整合到统一方案中，以提供更具成本效益的解决方案。

Gemini 4 进入企业市场后，我们预计市场份额将从 12%提升至约 18%。不过，我们仍处于落后位置，因此这仍是 Gemini 4.1 的重点改进方向。

提问者：Google Gemini 目前的付费订阅用户总数是多少？本季度和上季度的数据分别是多少？

回　答　者：仅看 B2C 市场，如果只统计 Gemini 的付费订阅用户，这也是最具实际意义的指标，目前约有 22.4百万付费订阅用户。Q3 的付费订阅用户为 23.1百万，未实现增长。

一旦 Gemini 4 向大多数用户开放，我们预计付费订阅用户数将大幅增长。正如发布 3.1 和 Gemini Pro 3 时出现的显著增长一样，预计明年 Q1 Gemini 4 上市后，付费订阅用户数将达到约 30百万。

目前的转化率约为 2%。因此，基本上将我刚才提供的数字乘以 50，就是非付费用户规模，这一规模要大得多。

提问者：您刚才提到，谷歌目前非常缺乏算力。今年谷歌的资本开支预计约为1950亿至2050亿美元。您预计2027年和2028年谷歌的资本开支规模及增幅将分别达到多少？

回　答　者：我们正在制定年度计划，预计于11月最后1周最终确定。2027年，我们将增加贷款，保守批准预算的最低支出为255 billion美元。此外，我们正在申请一项与特定KPI挂钩的条件预算。如果年中达到相关KPI，我们将获得更多预算，以扩大资本开支。

考虑到存储价格上涨，以及我们通过博通、联发科和Marvell从台积电获得的额外产能，我们正在申请于年中追加约40 billion美元投资。因此，我们的资本开支不会收缩，而是将从约205 billion美元增加至255 billion美元。如果年内达到业绩KPI，还可能在此基础上额外支出40 billion美元。2027年下半年，我们将与创始人和投资者进一步讨论，决定是否启用这笔额外的40 billion美元预算。

积极的一面是，我们将继续增加支出。对于V8和V9，我们能够与更多合作伙伴开展合作。针对V9，我们还将与博通合作采用EMIBT，这是另一种先进封装形式，有助于缓解台积电产能限制，确保更多芯片按时生产。到2028年，算力限制将有所缓解，2027年至2028年的资本开支仍将持续增长。

提问者：关于TPU，2025年Q3已经交付及Q4预计交付的TPU V8分别有多少张？其中，多少张用于内部训练和推理，多少张用于算力租赁，多少张用于直接对外销售？

回　答　者：Q3，我们获得了约55,000颗V8；Q4预计将获得约350,000颗。明年上半年将获得约3,000,000颗，下半年将获得约4,000,000颗。因此，今年V8的供应量占比非常小。

不过，V8的性能远强于V7。就峰值性能而言，V8的计算能力可以达到V7的5倍，因此将显著缓解我们的算力压力。

具体来看，Q3获得的V8中，约38,000颗将用于内部，另有12,000颗已承诺提供给Anthropic。这部分将采用租赁形式，我们会将芯片及其系统直接交付给Anthropic，而不是部署在GCP中。

Q4获得约300,000至350,000颗V8后，我们已预留250,000颗用于内部，其余100,000颗将提供给Anthropic和Meta，因此销售规模会相当可观。相关销售收入预计将从Q3和Q4开始体现，其中Q3的TPU销售收入约为10亿美元，Q4预计约为30亿美元或更多。我们从合作伙伴处收到款项通常存在60天至90天的账期，因此仍会存在一定的时间差，但相关收入将开始体现在账面上。

对于保留在内部、未对外销售的V8，目前约22%用于训练，78%用于推理。明年这一比例将逐步趋于稳定。最终，训练预计占总分配量的15%至18%，推理占比将继续增长至约85%。

提问者：刚才提到，部分TPU V8会放在公有云上进行租赁。V8的租赁价格大约是每小时多少美元？如果对外销售，销售价格大约是多少？能够为谷歌贡献多少利润？

回　答　者：这里涉及2个方面。目前，我们的RCP已超过5,500亿美元，主要是谷歌云为客户提供服务所对应的订单和积压订单。V8投入使用后，我们认为可以大幅消化这些积压订单。

回　答　者：超过550 million美元的RCP待确认金额可以开始流入，并成为我们收入的一部分。假设我们在2027年收到6 million台，即约4 million至6 million台，V8将为我们贡献约18 billion美元。

今年，即使我们在Q3收到首批V8，将其安装到服务器中仍需要约60天。这意味着，Q3收到的设备只能在Q4完成安装，之后可能还需要运行约1个月。因此，今年Q4由V8带来的增量效应较为有限。根据我们的估算，最多约为300 million美元。V8的大部分影响将在2027年体现。

V7的按需租赁价格为每小时12美元，V8的按需租赁价格计划为每小时21美元。大多数客户采用CUD，即承诺使用协议，合同期限为1年、3年或5年。1年期合同提供30%的折扣，3年期合同提供55%的折扣，5年期合同提供72%的折扣。可以将这些折扣率应用于21美元的价格，以计算实际报价。

我们运行V8的实际成本约为每小时不到3.7美元。与每小时21美元的按需价格相比，利润率仍然较好。大多数客户采用3年期合同。即使按照3.8美元的成本和9.45美元的价格计算，利润率仍然相当可观。因此，我们预计V8能够持续提升GCP的利润率。

提问者：下半年，NVIDIA的Rubin也在持续交付。预计Q3和Q4分别会交付多少张卡？Rubin的对外租赁价格大约是多少？此外，Q3谷歌云上有多少张卡采用按需租赁，有多少张卡与客户签订了长期价格协议？按需租赁部分的价格环比上涨了多少？

回　答　者：Rubin的每小时租赁价格需要再核实。Rubin是新产品，我目前还没有记住相关价格。

回　答　者：按需使用与 CUD 的构成方面，按需需求存在波动，平均约占 12%，CUD 占 88%。CUD 中占比最大的是 5年期合同。在其余 88% 中，接近 58% 为 5年期合同，并采用照付不议模式。这意味着，即使客户放弃合同，仍须履行全部义务。由此可以看出，客户迫切希望锁定算力，因此更多客户选择签订 5年期合同，其次是 3年期合同。

Rubin 的按需价格约为每小时 13美元。由于 Rubin 是从外部采购，而非自有产品，因此成本较高。对于 1年期合同，我们仅提供约 20% 的折扣；3年期合同提供 35% 的折扣；5年期及以上合同提供 50% 的折扣。与 TPU 相比，Rubin 的折扣幅度要小得多。

提问者：谷歌非 AI 部分的云收入增速如何？此外，整个 Workspace 的收入增速情况如何？

回　答　者：关于不包含 AI 部分的收入，您指的是仅基于传统 CPU 的 GCP 收入，对吗？

提问者：是的。

回　答　者：非 AI GCP 业务仍在增长。Q3 云业务整体增长约 97.8%，其中仅非 AI GCP 业务的增速为 25%。AI 数据中心带来的益处在于，AI 也提升了市场对非 AI GCP 平台的需求，因此该业务仍保持增长。请再重复一下第2个问题。

提问者：还有 Workspace 的收入增长情况。

回　答　者：Workspace 的增速并不是很快。上个季度，我们提高了价格，并要求所有用户采用整合 AI 的产品，因此 Workspace 收入出现了较大幅度的增长。但进入 Q3 后，本季度 Workspace 收入增幅约为 15%，是我们各产品领域中增长最慢的业务。

回　答　者：目前，每个席位的价格上涨已经有所体现，但本季度的主要驱动力来自企业客户付费席位数量的增加。普通企业和大型企业并未出现显著增长，主要是中小企业从我们的捆绑方案中看到了更大价值。因此，中小企业一直是 Workspace 席位增长的主要驱动力。

提问者：刚才您提到，部分芯片用于按需市场，部分芯片则与客户签订了长期协议。请您再说明一下按需使用和 CUD 的占比。

回　答　者：通常来说，按需使用的占比为 12%，CUD 的占比约为 88%。在 CUD 中，接近 2/3、但尚未达到 2/3，我认为大约 60% 来自5年期合同。因此，与3年期和1年期合同相比，CUD 明显偏向5年期合同。3年期和1年期合同合计约占 40%，5年期合同约占这 88%中的 60%。

提问者：关于 AI Overview，我们想了解目前 AI Overview 广告的整体渗透率，以及 AI Overview 内部广告位的渗透率。

回　答　者：我不太确定第2个问题中“内部广告位渗透率”的具体含义，能否进一步说明？我可以先回答第1个问题。目前，AI Overview 广告的整体渗透率约为 8%，基本维持在 8%至9%之间。美国市场已经达到 12%，世界其他地区基本为 7%，因此综合平均水平略高于 8%。下季度，我们计划将其提高至 9%，整体渗透率正在上升。

此外，还需要考虑 AI Mode。目前 AI Mode 的广告渗透率约为 3%。正如之前讨论的，到今年年底，这一比例将从 3%提高至3.5%至4%。我们之所以采取渐进式提升的方式，是因为需要持续监测客户体验和客户反馈，并据此对照计划进行调整。

提问者：您刚才提到的 AI Overview 广告渗透率是否……

提问者：AI Overview 的广告渗透率是否包含位于 AI Overview 上方和下方的广告？

回　答　者：是的，AI Overview 上方和下方的广告均计入广告渗透率。

提问者：AI Mode 在全球和美国市场的 DAU 渗透率分别是多少？美国市场 AI Mode 的人均每日搜索量是多少？

回　答　者：从全球市场来看，AI Overview 的用户规模约为 2.68 billion 人，AI Mode 的用户规模约为 1.2 billion 人。

美国仍是 AI Overview 的主要用户市场。美国约有 280 million 人使用 AI Overview，约有 114 million 人使用 AI Mode，美国用户占比约为 10%至12%。

从人均每日使用次数来看，全球 AI Overview 的 DAU 用户约为 2.1 次，美国约为 3.6 次；全球 AI Mode 用户约为 0.554 次，美国用户约为 0.88 次。AI Mode 的使用频次仍明显低于 AI Overview，主要原因是 AI Overview 会直接展示，而 AI Mode 需要用户主动触发并进入该模式。

原文转译

电话会议，仅服务于九千咨询客户，接受邀客户，仅供研究交流，不构成投资建议。相关人员应自主做出投资决策并自行承担投资风险。九千咨询不因使用本次内容所导致的任何损失承担任何责任。专家发言内容仅代表专家个人观点，不代表本公司观点。本次会议内容不得涉及国家保密信息、内幕信息、未公开重大信息、商业秘密、个人隐私，不得涉及可能引发不当炒作或股价异常波动的敏感信息，不得涉及对社会或资本市场稳定的言论。未经九千咨询事前书面许可，任何机构或个人不得以任何形式复制、录音、刊载、转载、转发、引用本次会议内容，否则由此造成的一切后果及法律责任由该机构或个人承担。本公司保留追究其法律责任的权利。感谢您的理解与支持。

好的，专家咱们马上开始今天的访谈。在访谈前请您就以下陈述表明我同意或我不同意。就是您知道保密信息的定义，并同意不会在访谈中透露任何此类信息及其他的保密信息。

Okay, before our interview starts, please confirm that you agree that any confidentiality of firms and protocols you will abide by and also none of the information that you disclose here in this meeting will be transferred to any third parties for any commercial uses.

Yeah, I agree.

Okay, 好的，好的。我们想先问一下就是谷歌广告这边的情况。然后我们有关注到就是Q1的时候谷歌的搜索广告它的同比增速是19%，Q2的话同比增速是17%。我想问一下Q3的一个同比增速的话，它比Q2高了还是低了？然后背后的一个主要原因的话是什么？对。

专家刚才说AI search的那个广告。

Yeah, yeah, yes. 我就想问一下那个search, search ads, yes.

So the first question is on the Google search ads. So we understand that for the Q1 the growth rate is around 19% and for Q2 it's 17%. So do you know the figures for Q3? Is it higher than Q2 or lower and what is the main reason for it?

Okay, so Google AI search, right?

Yeah.

Yeah. So, so basically for Google AI search there is actually AI overview and AI mode. And then the growth rate has been actually, um, it has actually come down. So AI overview we have been growing roughly at 8% quarter over quarter. And then for AI mode it has been growing roughly about 2%. So yeah, I would say, uh, the growth rate has slightly slowed down, but, um, but it is still strong growth. Yes.

呃，想问一下就是专家说的一个增长是环比的增长吗？对，然后我刚才的问题其实问的是搜索广告总体的一个增长。对，因为Q2的时候搜索广告总体增长是17%，想问一下Q3搜索广告总体的增长的话是高于17还是低于17？还是什么样的一个情况？背后的主要因素是什么？对。

Okay, 刚才专家回答的是AI search的一个情况，然后他说的是环比的数据。然后我再追问一下那个search广告总体的一个情况。

Okay, so actually the expert, our customer is more related to the search ads in general because the figures that we got...

Search ads.

Yeah, yeah, yeah.

Ah, okay, okay, okay. Ah, search ads actually... So compared to quarter over quarter, we have actually grown at 14.2% for search and ads. So it is actually still maintaining a strong momentum. So yes, we are actually coming down because last quarter we did about 16.8%, but this quarter we did about 14.2%. Main reason for deceleration is it's not because of the less popular search and ads. The main reason is because of the currency, currency exchange rate is actually playing a very negative role for us. Yeah, that's the main one. So yeah, think it as we are still maintaining the continuum. And also, can I do this? Let me, because the call quality is so bad, let me try to rejoin. Let me leave and rejoin right away. Just give me one seconds. I hope that that improves the call quality. Yeah, that improves the call quality. Let me rejoin.

Yeah, okay.

Okay, I'm back. Yep, yep. Please continue.

Mhm.

好,可以繼續了。

好的,明白。想問一下就是除了匯率的一個影響之外,還有沒有什麼什麼因素就是讓谷歌的廣告沒有再進一步的增長呢?因為之前的話,市場普遍預期的話就是,就是我們就是之前的話跟您溝通,然後的話您認為就是說在Q3的時候,谷歌的話會加大整個AI搜索廣告的一個廣告加載率,那想問一下這件事情的話,它後面有沒有新的進展呢?對。

On top of the currency exchange rate changes, are there any other reasons for the decelerations? And secondly, because in previous talks you mentioned that, or the overall market sentiment is that Google would keep expanding on its search ad growth in quarter three. So I'm wondering what what is the reasons here? And also, we know that Google has been working on its AI search ad load, like increasing the ad load here. I want to know the latest progress here on this part.

Yeah. I think actually the the fact that yeah, besides the currency exchange, I think the the other fact is that in order to actually, so we we have taken the strategy to lower the the search ads and its cost for especially for AI overview ads and then also the other AI mode ads embedded which is actually covered inside Google search in order to push it to global regions and then increase the adoption by the regional advertisers outside of the US. So yes, it is actually a strategic approach where we want to make our advertisement product affordable enough for rest of the world advertisers to try it. And so there's actually an artificial factor of significant, we are actually extending roughly about 15 to 20% discount on advertisement for the first time AI ads user basis around the world. That's actually one. And then as a result, we are actually expanding more aggressively toward the rest of the world, which kind of has impacted, which was impacted by the currency. So it is actually those two that combines that actually created the slowing down. Now, in terms of ad load, I think actually the percentage we we are now actually for normal advertisement, we are actually al- Almost 22% ad load. And then for AI overview, we have actually increased it to roughly about 8%. What we are actually targeting is we believe we can get up to 40%, but we are still on the way there. And then for AI mode, um, we actually are doing roughly about 2.5%. So which is actually for AI mode, our target is we want to get up to all the way to 20%. And so we are still on the way up there. So for this one, what we are actually doing is we are we are balancing because we are not in a hurry for immediate cash. What we are actually doing is we are doing balancing between customer experience and then the ad load percentage improvement on our new format. And one other thing is now we are also doing UCP and we are also introducing um basically like the, you know, the shopping ads into UCP. So we know that there's new ad factor coming in. So what we are doing is more ad product, uh AI Max is also doing well. So there's no need to hurry and ruin the customer experience. We are not doing that. We are doing slowly expanding the coverages.

Okay. 好的，想问一下，那上个季度的话，那个AI搜索和AI overview和AI mode里面的广告加载率大概分别是多少？

So, uh, speaking of ad load, we'd like to know the figures from last quarters. Uh, we'd like to know the ad load breakdown for different modes.

Mm. Last quarter versus this quarter, right? And then what was the second question? Sorry.

There's only one question. The breakdown of the ad load for different uh ad mode in last quarter.

Ah, okay. Uh, so uh ad load for normal Google search is last um last quarter we actually did 20%. This quarter we did 22%. And then we are on the way to 40%. AI overview, we did roughly about 7.2%. Now we are actually over 8% this quarter. And then for AI mode, we did about 1.4%, but now we are doing 2.5%. So we are doing a gaining. Yeah.

Mm-hmm. 明白。好的，那想问一下就是AI mode它的一个query的话，现在的话会占到整个Google搜，Google搜索query的比例大概是多，大概是多少呢？相比于上个季度可能增长了多少？

Okay, now we we want to understand the total proportion of AI mode queries in the uh in the total uh Google search queries and what is the uh changes compared with last quarter?

Ah, okay. So it it is still quite small because the user has to intentionally turn it on and then do the search. Um, so roughly, um, okay, so there are three ways to look at it, okay? So in terms of all of our user bases for Google search, about 34% of the users now use AI mode at least on a monthly basis, uh, MAU basis. Now, um, for these people, um, they use about 12% of their search ends up in AI mode. So basically 34% times 12% is basically what you see is roughly about 4% of our total search is on AI mode nowadays. Uh, that number has increased dramatically quarter over quarter. So, uh, for Q-o-Q increase, um, what I can say is because this number is so small, um, what we actually did is we actually increased it about .6 15% so like okay so roughly about 0.6% increase compared to last quarter because last quarter we did about 3.47 so yeah we we roughly did about 0.5% improvement on on the on the rate. So that's actually on the number 0.5 divided by 3.5 that's almost 14% to 15% increase so that's a that's a significant increase.

Mhm. Okay, 明白。好的，然后我想问一下就是，专家认为下个季度，就是Q4的话，那个整个的一个AI搜索，相关的搜索它里面的广告加载率，它可能增长到多少呢？包括明年的话，专家认为就是整个AI搜索的一个广告加载率它可能会增长多到，到到到多少呢？它可能是什么时候会有，或者说在哪个季度会有一个比较明显的一个加速呢？

Okay. The next question, what do you expect the ad load for the different type of AI search would become in quarter four and also in next year? And also when do you think there will be a huge acceleration or huge increase of the ad load for AI search?

Mm. Okay, so okay, so we have a directive from direction from our CEO uh who says that he doesn't want a a sudden 80 or 100% jump on our performances because of we increase the ad load um in a specific quarter. What he wants is for the next four to five years, he wants our ad revenue increase steadily over the next four to five years rather than a big jump. Given that, what you will see is there is actually not going to be a a specific quarter where the ad load suddenly increase like with a surprise. We we don't want to go like the Nvidia style. Uh Q4, what we expect is we are actually going to push because this is a big season for um for shopping. We are actually going to push for additional 2%. Our target for 2026 Q4 is 24%. End of 2027, uh actually next year for each quarter we are going to push for roughly about 1.5%, which means that end of next year, you'll see we reaching about 30% ad load uh for uh for basically um for our roughly overall like Google search. For AI overview, it's a little different story so we are going to push a little bit faster. Um for AI overview we'll hit 4% um like I mean this for next year first half um and then for next year second half we are actually trying to push 15%. So that one AI overview because it's so low we are actually going to try to push it harder. For AI mode, uh this last quarter it's actually going to be roughly about we are going to maintain roughly about 3.5% but for 2027 our end goal for end of the year is roughly about 5.5%. So what you see is please don't expect any dramatic changes. It's actually going to be incremental improvement each quarter uh and so that we don't let the uh investors disappointed. There is a continuous small improvement every quarter.

Mhm, 明白。好的，然后我想问一下就是AI Max的一个进展，对，然后我想问一下就是AI Max的话是否在三季度的时候扩张到了更多的模糊搜索的一些关键词，包括在那个AI mode的里面，对，然后的话，整个AI Max的话专家认为对于谷歌的广告的话，它在未来的几个季度是否会有一个加速呢？对。

Okay. So the next question is on AI Max. We are wondering uh in Q3 are there is AI Max expanding to more blurry search results or queries and also maybe used more in the AI mode and also for the next quarter what do you think the uh contributions of AI Max to the total Google search search ad?

Uh okay. Uh so so okay so I think about the blurry search um so basically um uh okay so so basically yes. So basically uh in terms of the in terms of coverages of the the query types what we actually did is AI Max was adoption was roughly about 7% but it is actually improving to 13% uh when you compare to the same quarter last year. So what you see is it actually almost doubled. Um and then also in terms of choosing um basically choosing the long tail search search words that the advertiser didn't choose now what we are actually seeing is it it is actually increasing roughly about 15% so basically the coverage coverages of the unselected words by AI Max has actually increased. Now in terms of the expectation what we actually think is is that uh this quarter we expect AI Max improving about our advertisement revenue by 1.5%. Uh next year it is actually going to improve our um our ads revenue by roughly about 2.3% is conservative assumption 2.8% is our optimistic optimistic uh view. So roughly about somewhere between 2.3 to 2.8 is our expectation of how much it's going to help us next year.

Mhm. Okay, 明白。好的，然后我想问一下就是个人AI agent的一个情况，因为我们最近有关注到Meta的Mews个人AI agent其实在这段时间特别的火爆，然后想问一下谷歌的话有没有在准备自己的个人AI agent产品，然后这个产品的话预计什么时候会推出？然后对，然后想问一下就是个人AI agent的这个产品就是像Mews这样的产品它会对整个搜索的搜索广告的一个市场可能产生什么样的一个影响？因为现在的话大家和互联网链接的第一入口依然是谷歌的搜索框，那如果说以后大家的话利用更多的用AI agent去执行像电商啊包括很多的像旅游啊住宿啊这些的订阅的话，或者不是说订阅是直接进行上面的那种商业意图的查询或者说完成相关任务的话，是否会对谷歌高价值的那部分query它产生一些比较大的竞争的压力，然后我不知道专家是怎么看的？

Uh next question on personal agent. We understand that the Mews from Meta uh has been quite popular recently. So are there any plans Google uh release a a personal agent of its own and if so when do you expect the release would be? And secondly for a personal agent like Mews what kind of impact does it have on the search ads because uh so far uh the entry uh the the the first entry uh for internet is still the search box of Google. So if we have more uses of the personal agents like Mews uh do you think that will have a huge competitive impact or the pressures on the high value queries uh concerning the e-commerce traveling and accommodations?

Yeah. So um so yes we are uh we are working on a similar product like the Mews. However what we are going to do is we are going to integrate it into the existing distributed apps because we have already upper hands in terms of how how many Gemini apps has been installed across the different devices. So also we are integrating the Gemini uh so we we are still coming up with the marketing name for it. But basically our personal agent will be embedded inside Android and Chrome and then it's going to embed into the kernel of every operating system that we do. So therefore you don't need to actually install it, it's actually going to be there. Um just a matter of you turn it on, turn it off um and and then use it. So uh so that's coming and then we are going to leverage our already strong distribution network. Now in terms of the impact, uh we actually uh our Pro 4 is quite strong and then it is actually powered by Pro 4 and then we are already working on 4.5 and 5 in parallel. So the agent that we actually serve is actually doing great benchmark in against Muse. So we are very confident that in terms of B2C um not like the B2B where we we actually are terrible at coding. Um this one we have a lot of com uh competitiveness to kind of win back the customer from Muse. And then also on top of the conveniences that we afford on Gemini app for iOS and then the kernel level integration at Android operating system. Now saying that, when you look at the edge market and let's say it's 100%, um the ones that agent can actually buy for you, we actually see that in the next five years we see it as more limited to travels and local buy and repeat purchases and those are actually roughly about 28% of people's spending. And then when you actually look at it, people actually are afraid of like trying new things. So they they are slow growth. So we think in 2027 roughly about 3%, 2028 roughly about 8%, 2029 roughly about 14% of the total transaction within this 28% will be done by the agent. So if you multiply it, so 3% of that uh let me just round up 28% to 30%, okay? So 2027 roughly about 9% uh of the total spend could be done by the agent. Now, uh what we think is Muse will actually kind of dominate but we are very confident that we can actually um basically retain um roughly about half of the market. So if you limit it to half of it, um and then then then basically then the impact of it is actually going to be 4.5% is actually going to be the maximum exposure of loss for the agent uh agent based purchases which actually reduces the total advertisement. However, however, we don't we don't think that um the overall edge market because of the agent and then UCP is growing. So that 4% to 5% loss, potential loss in 2027 is more than made up by a large increase in other form of advertisement. So overall revenue will not suffer. But in 2027, yes, you're right that there is actually a potential that half of that 9% is going to impact our revenue. But uh 50% of the market share in the agentic purchases is our very, very conservative. In 2028 and 2029, our view is just like Google Search, uh we believe that we can push it all the way back to 78% market share. So given that scenario, uh we lose 22% of maximum 14%, which is actually going to be less than three to 3%. So maximum exposure to our revenue is actually going to be around 3% 3 to 4% on on the overall current search ads which is not including the new expanded different set of advertisement market that we are actually going to see. So no impact to revenue maximum 3 to 4% impact to our previous search ads revenue. Yes.

明白明白。OK，然后想问一下，就是专家认为的话，就是个人的AI agent的话，它会让广告的市场变得以后会变得更大还是更小呢？因为过去的话，谷歌的它的广告收入其实是来源于这个搜索的一个竞价排名的一个广告。对，然后这种广告的话在AI agent的一个它做决策的链路里面的话，我不知道它是否依然依然依然是有效的。对，然后的话，因为现在的话，Meta的话它其实是没有去对于这个AI agent产品去做任何的一个直接的一个变现的，对吧，它可能只会收一些订阅费，对。然后那对于谷歌来讲的话，它如果说是推出了自己的个人AI agent，它要怎么去对这个AI agent产品去做变现呢？对。

Okay. So with the joint of AI personal agents, do you think the overall ad market or ad space would grow larger or smaller? Because in the past we know for Google your ad revenue mainly comes from the ad bidding mechanisms and for I'm not sure if that kind of mechanism will still be used for personal agent in their searching. And the second part of the question is so far we found that Meta has not monetized much for its new products. Maybe it has has some subscription fees but overall it does not have any monetizations for this product. So if Google is to release its own personal agent, what is the major ways for monetization?

Okay. Actually the interpreter I actually I didn't understand the first question. I understood the second answer. Could you elaborate the first question just one more time? Yeah.

Yes, do you think that with the joint of personal agent for ad search for for for the ads market, yeah, do you think the overall search ads market would grow larger or smaller? Because so far the Google ad revenue mainly comes from ad bidding while this mechanism is not widely used in personal agent.

Yeah, yeah. Um so so great question. So um the advertisement market however is actually going to evolve into different form. And so what's actually going to happen is um when people actually engage the agent, they entirely skip the search and then the transaction and then like basically they skip the funnel and then just go and purchase it whatever is being recommended. Now because of that there are two things that two things happening. Now the agent when as they actually buy something, uh our Google agent is actually going to show other products that they can buy alongside the product they are buying or they will also display advertisement while they are actually carrying out the task for the user. So we are actually actually we are doing um you'll be surprised we are doing actually more than um 20 different types of new type of advertisement to go with the personal agent. So those formats is actually going to replace the previous formats of advertisement and then um retain the edge going forward. Like there'll be more even more personalized advertisement such as like you're buying something for a fitness shoes, we'll be actually recommending hey what about you probably need a yoga shoes and also we know that you probably want to buy something more healthy so therefore here is some vitamins to build up your muscles or whatever. So new form ad- advertisement is happening and then also that is being supported by the merchant which is actually working with the agent to sell something. While waiting online, the merchant is actually going to pay us to show their advertisement. So those are actually going to be something to offset the loss from the agent. Uh and one other thing is um now we believe that the customer will ask us and ask us to advise what product to purchase. If we actually provide consultation and then it actually leads to purchase of a specific product, then um in that process we are going to in the recommendation algorithm, we are also going to embed the advertisement so that it doesn't show as ads but the recommendation actually contains information from the advertisers. Embedding that information into our assistant, um the product information and the you know what's good for it and so on is actually going to be chargeable to the advertiser. So that's a new form of advertisement we are exploring right now. So I think to be honest at this stage things are changing so much but it's hard to predict but what we are actually doing is these two are we have received so much interest and strong spending from our advertisers that we think that this is actually going to be enough to offset the loss from skipping the mid processes, the display ads and all those processes uh by these two new set of um advertisement. Um and and on top of it I think the other one is um uh so basically the so so and and the last thing is actually this let let me actually say that um that we are integrating Google Maps into doing um sales channel so you can order product directly from Google Maps um and also when you wear our AR um AR glasses then you can order something directly by just looking at something and then also AR new form factors such as AR glasses and smart glasses which is coming up they have new advertisement spaces uh which is actually something that and Google Maps um and then Google Workspaces is actually going to sell databases or other solutions and then also Google Maps you will be able to one click and purchase something and also see all the advertisement at the same time. So I think this new form factor of ads is going to be enough to offset any loss from the traditional advertisement format. Um so that uh that one actually now in terms of monetizing it Muse is actually based on subscription there however because we have um few people from the Muse team from uh Meta who actually joined our Google Gemini team um actually there are a few every week um what we actually learned is they do not actually intend to add advertisement until end of 2027 however they have significant plan to ramp up on advertisement from 2028. Meta is advertisement company like Instagram and Facebook you see ads all the time so that's actually coming. Now we are actually going to do same thing we are actually going to invent new types of advertisement for for personal agent um which is going to be the recommendation direct offer which is say like hey here's something that's on discount do you want to buy it so basically if we think that user we are at the stage where we anticipate what the user needs and then we can actually negotiate with the sellers and then basically offer a discounted price for a specific product to make the sales and then receive the commission for doing that. So so what we actually think is our personal agent is actually going to have a more evolved form of conversational advertisement um advertisement based on recommendation algorithm and also direct offers with um negotiation based uh offers with the sellers again is the buyer. Uh and and and one big change we are doing is we are not actually going to um take commission on any of the transaction, which was actually our initial original plan, but we remove it so that we can take on much more number of merchant uh to join us and then use Google platform as avoiding any marketplaces, but basically use all of the we are actually going to allow customer to skip the entire Amazon and all the marketplaces, Rakuten, Shopee, Lazada, um AliExpress, but people can directly work with the supplier and then be able to buy things with one click on our platform. That means that there's so much room for us to add different types of advertisement there. So views and us, you can expect from 2028 onwards, there'll be a lot of new form of advertisement coming.

OK，明白。然后想问一下就是，谷歌和Meta做个人agent产品的话，就是各，就是两家公司各自的优势是什么？然后专家认为对于个人agent产品，它最重要的一个竞争的一个壁垒的话可能是什么？对，然后讲一下这个原因的话可能是什么，对。

竞争的壁垒是什么意思？

对对，就是说专家认为的话就是说，个人agent产品的话，未来的话可能的一个胜负手的话可能是什么？对，为什么认为这个的话可能是未来整个agent最重要的一东西。

Okay. Uh so the next question on personal agent, what do you think are the specific advantage of Meta and also Google for developing their own personal advantage, like what are the distinctive advantage of the two companies? And secondly, to the longer term uh vision for personal uh agent, what do you think is the deciding factor for which company to win at the end and why do you think that will be the deciding factor here?

Yeah. Okay. Um so I think actually for us, Google, we are going to leverage our Google Search, Gmail, Android, Chrome, YouTube, Google Maps, Google Photos. Basically we have 20 services where we have more than 10 billion daily users who use it. So what we are actually going to do is we are actually building a integrated data lake behind all of these. So we are building the digital avatar of every user basis, which actually where we develop the psychometry, which is understanding what they're thinking, and then their behavior metrics, which is like predicting what they're actually going to do and what their motivations are. Um so so that's actually on one side that no one I I think actually that's actually uh bigger than uh Meta's Instagram or Facebook, which is mostly on social and then actually for social, adults and as they grow older, they kind of a graduate from it and then use less of it, right? So we are so I think Meta has upper hands and distinct advantages in terms of relationship-based interest and therefore they can predict how others are impacting this person and then what they see and um experience through the social media will actually have an impact on this person and then the live uh commerce they actually host will have impact on this person and then they will actually advance the algorithm on how to leverage this understanding to propose and sell specific ads and product. It's it's actually their main uh advantages. For us, we don't have relationship, but we know what each person is doing. We look at their calendar, we look at their mail, we look at their uh locations on Google Maps, and therefore we use this to algorithm to predict what they need and then basically offer it. Now, what Meta actually doesn't have and we do have is we have built GCP and then we are approaching about 1 million brands to join us after we remove the commission. It's going to be no commission whatsoever to the merchant to sell on our platform. So if you have Google agent on Search, Gmail, Android, Chrome, YouTube, Google Maps, Google Photos, and Gemini, Android, whatever, um, what happens is that every surface that we have becomes the surface where we can show ad ads at the specific moment where the user needs recommendation. So we are omnipresent. Meta is omnipresent when people are actually on WhatsApp or Instagram or Facebook, but we are omnipresent in more surfaces. Uh, and then we actually leverage what you call the personal psychometry and behavior predictor to recommend something right on the, at the right time, at the right moment, uh, right time, right moment, and then right places to push the sales and advertisement. So that's actually what we think is actually going to be the advantage, distinct advantages for Google. We have merchant, we have customer, we know the customer, and then we know when to push it. And then for merchant, they bypass the entire shopping platform and then they can sell direct on Google platform. So that actually is uh going to be a strong forte uh for Google. Meta, they have uh upper hands on all of the relationship driven model. And then also they will actually leverage the peer pressure and then also the interest groups uh to push forward. And then they will also have strengths in what they call live commerce uh to kind of push forward and integrate with the agent to sell stuff. So we think that's actually going to be kind of the distinct advantages for each. However, we think actually who wins eventually is that who has access to the customer is quite critical. So do you have a global presence of interfaces with the customer in terms of operating system or other online platforms? That's actually we think is about one third of it. The second part is, is that um, like do you, do you, do you have actually basically uh end-to-end agentic processes that include the ecosystem of the merchant? We think that's actually about one fourth of the success factor. And then the other 25% is um the trust and security based uh runtime model, basically agent being able to be trusted and safe enough to use. That's actually another 25% that we see as the main one. And I think the last one is the scale of economy actually kicks in because if you want to build such a large system to enable agent for everyone to see and then see the ads and then buy something, uh the scale of economy where you have your own hardware and software to actually lower the prices of serving it is actually going to be critical. Like um we have TPU where per the edge and then per the recommendation and per the transaction, our cost is about right now about 68% cheaper than what Meta does. So I think that's actually going to be a critical factor in terms of scaling up your system and platform uh going forward. So because you are short of money, but if you can you are cheaper, you can expand faster and then you make more money which actually you can invest for your expansion of your platform. So that uh cost effectiveness is actually going to be critical factor in terms of how fast you can actually scale. Uh yeah, these are what we think is actually going to be the critical factors. 好的，然后的话我我后面会问就是专家那个模型还有云的一个情况。对，然后对，然后我们有关注到就是那个Gemini 4的话它已经发发布了，然后想问一下就是在Q3的时候谷歌Gemini的一个收入和它ARR的情况的话，它大概是它总体的一个增速的话可能是一个什么样的情况，它包括它同环比的一个增速。然后专家认为就是Gemini 4发布之后的话，会让整个的Gemini的一个收入会有一个比较大的一个加速吗？因为当然现在的话Gemini的话它其实还只开放给了政府还有一些网络的防预方等等的，还没有全面的给订阅者开放，然后那专家预计什么时候会开放，然后正式开放的话，它会让Gemini的收入会有一个很大的加速吗？对。

Okay, next question on Gemini. We know that Gemini 4 has been released. Uh and in so we first want to know for Gemini in general in Q3, what is the revenue and uh and ARR growth rate uh like year over year and also quarter over quarter. And the second part of the question is uh we know Gemini 4 has only been uh accessible for government agencies and also the internet defenders uh has not yet been uh opened to the subscribers. But when it does open to the public, uh do you think it will have a huge impact on the revenue? Like do you think it will bring up the drive up the revenues uh very uh in a very large scale?

Um, so uh so our expectation is actually for in terms of the ARR, we have three buckets like um we have API, we have enterprises and then we have subscription. Um what we see is um we API is actually going to in terms of revenue API growth once it's uh it's more available, we expect that uh in Q3 because we are delaying the exposure to the uh API developers, it's actually going to grow the API by roughly about 24% but going forward in Q4, we expect it's actually going to grow it by 45%. So it's actually going to grow faster. Uh Q2 our API revenue growth was about 38% but then we expect given its uh performances, it's actually going to drive more people to use it. For enterprises, we actually experienced 40% increase in Q2. Uh Q3 delayed launch we expect 30% growth but what we expect is Q4 we expect roughly about 58% growth from the enterprise perspective. And for subscription, we did 12.5% in Q2 but Q3 about 10% increases because of the delayed launch but Q4 we expect about 15% increase in the subscription uh because of the new new features. Now um in terms of um okay so basically we currently actually mainly actually work with government or cyber security partners and so on. Um in November we are actually going to launch it to our API customers and also for B2C customers only the AI Ultra customers will be able to access it. And then in 2027 in uh January we are going to make it available for everyone. Um so um I think actually the main reason why we are delaying the launch is because of the compute uh capacity. We are actually borrowing xAI uh compute capacity from um Elon Musk company. They don't uh Grok doesn't use all of it so we are going to borrow about 1 billion worth of compute capacity from them. And then then as those units come in and also our TPU v8 from TSMC start to be installed on from this end of the this quarter, then we'll be able to release it to more people. Um, so, but I think actually I want to manage the uh kind of expectation uh because I I have actually tried it and used it and I also participate in the analysis of the benchmark for the impact. Um, in terms of uh in terms of agent capability to serve enterprises like like doing the task, it's actually going to be uh top notch. Uh actually slightly better than Astra. Um, so in terms of doing the task, it's actually and in terms of doing analysis for mathematics, finance, science, and so on, it's actually going to be top notch. Um, only it's a it's going to be on par with Astra, so it's going to be good. But however, uh I'll be upfront that um we are still lagging behind in terms of uh the coding capability. So we are we are much better, but when you compare to Opus 5.5 or Astra, um we are about I would say 10% short of let's say they are 100% then I think at best we are about 90% of where they are. Uh so the coding is still an area where we still need a lot of improvements. Um but you can see that uh the rest of the area especially for B2C or the task related productivity related, you will see it. So um I think actually in terms of the B2B market, you will see that uh Claude and um ChatGPT continuing to dominate. Our selection rate right now is roughly we have about 12% of the market share. We don't think that that's going to be dramatically changing. Um we we still think that we'll be gaining more market share because of the its integration with Google Workspaces uh and other like we are now integrating uh our entire Google offers like Google Ads and uh Google databases GCP and then also LLM into one bucket and provide much more cost effective solution. So we'll still grow, we expect to capture about 18% as versus 12% when we launch the 4 in enterprise market. However, we are still behind. So that's actually still a main focus area of where we are going to improve for uh 4.1.

明白。好的，想问一下就是那 Gemini 现在的话订阅的用户数的话大概是多少啊？然后上个季度大概是多少？

Okay, uh next question. Uh what the total number of users uh subscribers for Google Gemini and uh what is the figures for this quarter and also the figure for last quarter?

Okay, so just the B2C, um, so I think actually let me try to okay, so if you if you look at the only the paid subscription user basis for Gemini, which is actually what really counts, we have roughly about uh 22.4 million paid subscriber for Gemini. But if you look at uh Q3, we actually didn't improve. Uh we we only have 23.1 million paid subscriber for Gemini. Uh but once we actually make this available uh for most users, we expect that um we have a big jump. Like just like when we launched the 3.1 uh the Gemini Pro 3, we had a big jump. Uh we expect that we are actually going to reach about 30 million sub paid subscribers once we hit the market with Gemini 4 in Q1 next year. Um so that's that's actually going to be um a major our current conversion rate is about 2%. So what you can see is um like basically if you multiply 50 to the number that I shared is that's actually non-paid user basis is um is is actually much bigger than um Our paid basis.

明白。好的。然后我想问一下，就是专家刚才也讲到了，就谷歌现在非常的缺算力，然后今年的话我们已经看到谷歌的 CapEx 的话，它差不多是要有 1950 到 2050 亿美金，对吧？想问一下就是明年的话，包括 28 年的话，谷歌的 CapEx 它可能会有多少呢？可能会可能会增长多少呢？对。

Okay. So you just mentioned that for the uh the Google actually actually lacks a lot of computing powers. And we've noticed for this year the CapEx of Google has amounted to around 195 to 205 uh billion dollars. So what do you expect the scale of the CapEx for Google would uh would look like in 2027 and also in 2028?

Yeah. So, uh so we are finalizing our annual plan. Actually, it will be finalized on the last week of November. But what we are actually going to spend in um next year is that we will actually take out more loan and then um the minimum we will be spending is um $255 billion is the conservative approval approved budget. But then we are actually asking for conditional budget when we actually exceed certain KPI. What that means is in the middle of the year, we will be assigned more uh more budget to expand our CapEx. And then what we are asking for is because of the memory price increase and then the uh additional capacity we secure from TSMC through Broadcom, MediaTek, and Marvell, uh we are asking for about um uh $40 billion more to be invested in the middle of the year. So you can see that um we we are not shrinking. We are increasing from roughly about $205 billion to $255 billion. And then potentially, we can optionally spend additional $40 billion on top if we hit the performance KPI as we progress through the year. So which will be released, additional discussion we will have in the second half of the next year with our um founders and investors to see if we can actually exercise $40 billion more. But that that's uh so I think the good news is uh we are spending more and then the the V8 and V9, we are able to um work with our additional partners. And um and then also for V9, we are going to work with our um Broadcom to work with EMIBT, uh which is another other form of advanced advanced packaging to um unlock the constraints from TSMC. So be able to produce more chips on time. Um so yeah, those are like um so I I don't think actually after we reach um 2028, there will be less constraint from the computing power and then there will be actually continuous spending increase uh in 2027 to 2028.

OK，明白，好的。然后我想再问一下就是整个 TPU 的一个情况。想问一下今年的 Q3 和 Q4 的话，然后谷歌的话，它的 TPU 它的 TPU V8 的话大概到到了多少张？然后有多少张的话用于内部的训练包括推理的需求？有多少张的话用于算力租赁？然后有多少张的话是用于直接的对外的一个销售？对。

Okay. Next question on the TPU. Uh how many V8 TPUs have been delivered in Q3 to uh or will be delivered in Q4? And among those arrived or delivered V8 TPU, how much would be used for in-house training and inference, and how much would be used for rental, and how much would be used for uh for sales?

Um. Okay. Um, yeah, okay. So in Q3, uh, we, we got roughly about, um, 55,000 units of V8. Uh, and actually Q4, we expect to receive roughly about, um, 350,000 units. And then the, uh, second, next year, uh, in the first half, we are going to receive about 3 million units, uh, in the first half and then, then 4 million units in the second half of next year. Um, so this year is actually going to be a really tiny fraction of the V8. However, V8 is, um, much more powerful than V7. Like in terms of, uh, peak performances, it can do five times more compute than V7. Um, so it's actually going to relieve a lot of pressure on our compute. Now, in terms of percentages for Q3, whatever we receive, we are actually going to use about, um, 38,000 of it for internal and then we actually committed to give 12,000 of it to Anthropic. Uh, and it's a actually in the form of lease, so we are actually shipping the, the chip and its system directly to Anthropic, um, not in our GCP. And Q4, when you receive about the roughly about 300 to 350,000 units, what we are actually going to do is, uh, we have reserved 250,000 units for internal, but you see that, uh, we are actually going to use, um, basically the remaining 100,000 units to go onto Anthropic and Meta. So the sales is actually going to be, uh, quite, quite big. And then, um, we expect the, the sales revenue, uh, starting to hit us in Q3 and Q4. Um, it's actually going to amount to Q3 roughly about a billion dollar, but Q4 roughly about 3 billion dollars or more is expected revenue from sales of TPU. Um, there is actually a 60 days to 90 days time period until we can actually receive money from our partners. So that gap will still be existing there, but that, uh, the revenue will actually start to show up on our, uh, our book. Now, in terms of training and inferences within, um, the unsold unit, like with the one that we keep inside, currently we reserve about 22% for training and then 78% for inferencing. Um, so yeah, and then that ratio is actually going to kind of stabilize, uh, next year around. Eventually, we think that training will take about anywhere between 15 to 18% and then inferencing will continue to grow and then take over about 85% of the our total allocation.

嗯，明白，明白。然后想问一下就是整个TPU V8的话，它，呃，呃，刚才有讲到有多少张的话是放在公有云上去做一个租赁的，然后它的一个租赁价格的话大概每小时是多少美，是多少美美美金呢？对，然后它如果对外销售的话呢，它的销售价的话大概是多少？然后算到谷歌的一个利润的话，它大概会有多少呢？对。

Okay, so for the V8, uh, you mentioned that some of them will be rented on the public cloud and what is the hourly hourly rate for V8? And the second part of the question is the sale price for V8 and how much revenue would it contribute to Google in general?

Okay. Um, so there are actually, uh, two things. One is that, um, right now we have over 550 billion dollars in RCP, which is, um, basically the order, the backlog order to serve our customers on Google Cloud. So when the V8 actually, uh, kind of joins us, we believe that we can substantially remove move a lot of RCP which is like over 550 million dollars of waiting money can actually start to come in and become our part of it. So um in uh when once we receive let's say the six six million like four of six million units in 2027, the V8 is actually going to contribute about 18 billion dollars to us. This year, even though we receive in Q3, uh the lead time to installing into the server actually takes about 60 days. So what that means is even if we receive that uh the initial batch, we can only receive we can only install what we receive in Q3 in Q4 and then operate it probably a month or so on. So for this year Q4, the uplift effect of V8 is quite limited. It's only going to be um at best we estimated it uh to be around 300 million dollars. So it's going to be limited. Most of this V8 impact is going to be available in 2027. Now that being said, um V7 we actually lease it at 12 dollars uh on demand and then per hour, but for V8 we are planning to lease it for 21 dollars per hour for um you know on demand. Now most our most of our customer as you know use CUD, committed use uh development, which is one year contract, three year contract or five year contract. If uh for customers one year contract we offer 30% discount, three years contract we offer 55% discount, and then the five year contract we actually do offer um basically actually 72% discount. So you can actually kind of apply this discount rate to 21 dollars and see um basically how how much we offer. Now the cost, the actual cost of running V8 per hour for us is actually roughly about less than 3.7 dollars. So you can kind of see 3.7 dollars versus on demand 21 dollars. The margin is still good enough and most of our customers are on three year contract. For our 3.8 dollar cost versus 9.45 dollars, um the margin is still quite significant. So yeah on this one we expect to be able to make um actually increase the margin for GCP uh on ongoing basis.

嗯，明白。好的，然后想问一下就是，那么下半年的话其实整个Rubin的话，它其实就Nvidia的Rubin的话，它其实也在那个持续的到卡嘛，想问一下Q3和Q4可能可能到了多少张卡，然后Rubin的对外的租赁价格大概是多少？对，然后想对然后顺便想问一下，Q3的话就是可能就是谷歌云上面有多少张卡是on demand的，然后有多少的话是和客户签了长期的一个价格协议的。对，然后像on demand的这一部分的一个价格的话，它可能环比上涨了多少？对。

Okay. And we understand that the the Rubin from Nvidia has also been delivered in Q3 and Q4. And what is the rental price for the Rubin series? And second question is in Q3 uh for the for the ones that's been rented on public cloud, how much of them are on demand and how much of them have been uh rented through a long-term contract? And for the on demand uh chips, um what is the rental uh what is the price increase quarter over quarter?

Okay. Um so for Rubin, um basically per hour rate, uh so actually let me just uh make sure that I double check this, just a minute. Okay, Rubin is actually new one, so actually I I I didn't memorize this one yet. Uh it's uh in my in my reference book right now, just let me give you that.

Mhm.

Um, I I think actually let me I think something that's already in my mind is actually that for the breakdown of on-demand versus the CUD, um, basically our on-demand demand is actually fluctuating but on average we see it as roughly about 12% is on-demand. Uh 88% is actually the CUD. Actually the biggest part of the CUD surprisingly is the five-year contract. So most of our customers, uh if you look at the the remaining 88%, uh almost close to 58% of the uh 88% is based on five-year contract and then uh and on take or pay basis, which means that even if they give up on contract, they still have to fulfill the entire obligation. So you can kind of see that the clients are quite desperate to secure their compute capacity. So more clients are actually doing five-year contract and then the second biggest is the three-year contract. Um, now that being said, if you look at Rubin, um, basically Rubin uh per one hour on demand we are actually offering at roughly about $13 per hour. And then for Rubin because we are actually buying it from external parties, not our own, it's expensive. So we only offer for one-year contract roughly about 20% discount. And then for three-year contract 35% discount. And then for uh the five-year and beyond contract we actually offer 50% discount. So the discount rate is much less uh compared to our TPU.

Mhm. 明白，明白。

Yeah. OK，然后的话想问想问一下就是，呃谷歌的话就是非AI那部分的云收入的一个增速情况，包括整个workspace它的一个收入增速的一个情况，对。

Okay. Uh the next question on the uh revenue growth rate for the cloud services excluding the AI part and also for the workspace revenue growth.

Um, uh so revenue excluding AI part, so you mean the only the traditional legacy CPU based um GCP revenue, right?

Mhm. Yes.

Ah okay, okay. Ah. So if you look at the non-AI GCP, it is still growing uh so our cloud is growing roughly about 97.8% for Q3. However, if you look at only the non-AI GCP, it is still growing at 25%. Um so you can see that uh the benefit of AI DC is that AI actually increases the demand for non-AI uh GCP platform as well. So that actually has been growing. Uh now, I think I forgot the second question. Would you remind me what the second question was?

Mhm. 还有还有workspace的一个收入增长的一个情况，对。

And also the revenue growth for workspace.

Ah okay. Uh so for our um so workspace um uh so it's it's it's not actually growing growing that fast. So I think actually you know that last quarter what we did is we forced everybody to uh kind of adopt we increased the price and then adopt the AI into it. So there was a big jump in terms of the workspace uh revenue last quarter. However, when you look at this quarter uh into Q3, we actually basically uh increased it roughly about 15%. So workspace is the slowest growing area of our um our product. Um and and um I I think actually one Something that you can kind of see is the the price increase per seat has already been reflected, but then you can kind of just see that uh the main driver for this quarter is mainly coming from increase of the paid seat by the corporate. We don't see a major increase from normal enterprises, big enterprises, but what you are seeing is the small and medium enterprises see more value in us in in the bundled offer. So they have been actually the main driver for growth for the for our workspace um places.

Mhm. 明白。剛才這之前專家講的就是有大概有多少張 有多少卡的話是 on demand 的市場的,對,然後有多少張卡是和客戶簽了長期協議的,這比例能再說一下嗎?對。

Sorry, we missed the part when you mentioned about the proportion of the chips used for on demand and the uh the proportion for CUD. Could you specify that uh once more?

Yeah, yeah. So um the rule of thumb is that the portion for on demand is 12%, 1 2. And then the CUD is roughly about 88%. But when you look at CUD, almost uh two-third uh not up to two-third, I would say roughly about 60% of the the the CUD comes from a five-year contract. So it's a disproportionately shifted toward five-year contract as versus three-year and one-year. Three-year and one-year all combined only accounts for roughly about 40% of the five-year contract accounts for about five like 60% within the 88%.

Mhm. 明白,明白。

Okay, got it.

OK,好的,然後對然後OK,然後的話我看一下我同事這邊的話有幾個問題,對,然後想問一下就是 AI overview 當前的廣告滲透率以及 AI overview 內部位置廣告的一個滲透率。對。

第二個是什麼?AI 內部 AI overview 內部位置廣告的滲透率,對。

Okay. Uh the next question on AI overview, we would like to know the overall penetration of the ad in AI overview and secondly, the uh the penetration for the internal like the internal spaces or spots for AI overview. I'm not sure if that second part is clear to you.

Uh I'm not sure about the second part. Uh like internal penetration, you mean like the actually can you clarify the second part? But I can give you the first part as the first answer. The overall penetration of the uh advertisement in AI uh overview is now we actually penetrated about about 8%. Um so basically um we are hovering between 8% and 9%. US we are already doing 12%, but rest of the world we are actually doing um basically 7%. So blended average comes out to be roughly about slightly over 8%. Uh next quarter we are going to increase it to 9%. Um so that uh is actually increasing. I think one thing missing there is AI mode. Uh we are doing roughly about 3% uh which which is actually going to uh kind of continue up to like as we discussed earlier, like um 3% going to 3.5 to 4% by end of this year. Um because the reason why we do it is as we increase it incrementally, we continue to monitor the customer experience and customer reaction and then we adjust as uh against our plan. That's why. Uh can you clarify what you mean by uh AI overview penetration for the internal spots? Yeah.

等一下,就是專家剛才提到的這個 AI overview 廣告滲透率是否 包含 AI overview 上方和下方的广告，对。

Okay, uh, so let me put it this way. Uh, so the ad penetration of AI overview, does it include the advertisement placed on above and below the AI overview?

Uh, yes. You mean, uh, yes, the ad placed above and below it is considered as the penetration. Yes, yes.

好的，好的，然后想问一下 AI mode 的全球和美国 DAU 的渗透率，以及美国 AI mode 的人均日搜索量，谢谢。

Uh, the next question on the uh DAU of AI mode globally and also the for the US market. And also for the US market specifically, uh what is the daily uh queries per person or per capita uh for AI mode?

Okay. Um, okay. Um, so I think actually, um, for so roughly for global perspective, we have about for AI overview, we have about uh we reach about 2.68 billion people. And for AI mode, we have about 1.2 billion people who actually kind of use it. Um, uh, and then uh I think in terms of the in terms of the US market, um, what you see is US is still the predominant uh user basis for uh AI overview. So, um, for US, it's uh it's actually roughly about 280 million people who actually use AI overview. And then when you look at AI mode, um, like we have roughly about 100 uh 14 million person who use AI mode uh for from the US perspective. So roughly about 10% to like ranges between 10% to 12% is actually the US proportion. Um, I think in terms of your question about number of queries in AI overview by um, like for global versus the US, US actually definitely ask more question. But um, the roughly, uh right now we have about um, a per person, right? Uh, if you look at per person, um, if my actually for this one actually let me just uh refer to my dashboard. Just give me one seconds.

Yeah, for the users of AI overview, what you see is uh on a for the DAU person who use AI overview, we actually see that they use about 2.1 times uh as versus US, they use it for roughly about 3.6 times. And for AI mode, uh what you actually see is uh AI mode is uh used uh about 0.5 uh 54 times a day for global. For US, it's a 0.88 times a day for the US customers. Yeah, so AI mode still um still is quite lower than AI overview because AI overview we just expose it as versus AI mode, they have to trigger and then enter the mode.

好的。

Okay.

明白，好的，我这边的话暂时没有其他的问题了，然后非常感谢专家的一个时间，后续有需要有新的问题到时候补充借这边再跟专家联系。

Okay, that'll be all for today's interview. Thank you so much for your time. We will reach out to you if we have any further questions.

Sounds good. Thank you. Yep, yep. Bye for now. Yep.

Okay, bye.