2026.09.23 | 9K | ASIC | ASIC Roadmap Check III

音频信息

音频标题: 260923 - Merit ASIC技术路线调研III

相关标的：undefined

议题：

完整音频链接：

音频纪要总结

GPU与ASIC技术趋势及AI半导体竞争格局专家访谈纪要

1. AI算力需求与GPU、ASIC技术演进

a) 未来2至3年，GPU和ASIC技术将如何演进？

• AI硬件需求仍将保持高速增长

1. AGI和AI有望广泛应用于劳动力替代、银行、医疗、日常生活、智能建模等领域，对算力基础设施形成持续需求。

1. 模型训练阶段已经验证，算力越强，模型表现通常越好，因此行业仍在大规模建设AI硬件。

1. 英伟达、博通等主要厂商的AI相关业务均呈现高速增长趋势，受访者认为未来一段时间硬件需求仍然非常强劲。

• 需求结构将逐渐由训练转向推理

1. 当前行业仍在持续构建更强的基础模型，因此近期训练算力需求依然较高。

1. 但可构建的大型模型数量最终会趋于相对饱和，后续训练主要用于吸收新增数据和持续完善模型。

1. 随着用户规模和应用场景扩张，推理需求增速将显著高于训练需求，推理硬件在整体算力中的占比将持续提升。

• GPU与ASIC将长期共存

1. GPU具备较强的通用性和可编程性，能够适应模型和算法的快速变化。

1. ASIC适合工作负载明确、可预先预测的细分场景，可通过定制指令集、数据路径和软件栈获得更高效率。

1. 未来大型AI公司预计会同时使用GPU、定制ASIC和通用CPU，而非由单一架构完全主导。

• 外部变量仍需关注

1. 政府监管、出口管制及AI在军事领域的应用，可能影响AI部署速度。

1. 若没有政策或外部力量干预，未来18个月AI硬件需求大概率延续当前增长轨迹。

b) 为什么认为近期训练需求仍较集中，而长期推理增速更快？

• 近期训练仍是模型能力提升的核心

1. 行业当前仍在竞相构建更大、更强的模型。

1. 模型性能与训练算力之间仍存在较明显的正相关关系，因此训练基础设施建设尚未结束。

• 长期推理需求的市场空间更大

1. 模型数量不会无限增长，训练需求最终会逐步趋于稳定。

1. 推理对应实际用户和应用调用，只要AI应用持续渗透，Token生成和服务调用量就会不断扩大。

1. 因此，未来训练需求仍会增长，但增速预计明显低于推理需求。

2. 英伟达的核心优势与系统级商业模式

a) 英伟达相较AMD、博通和谷歌的主要优势是什么？

• 长期技术积累和行业领导地位

1. 英伟达开发GPU已有近30年，早期应用于游戏，随后拓展至加密货币计算和AI训练。

1. GPU的高度并行架构和矩阵计算能力，使其成为推动本轮AI算力变革的核心硬件。

• 从单芯片供应商升级为整机架解决方案提供商

1. 英伟达不仅提供GPU，还能提供CPU、网络交换机、高速互连芯片及相关系统。

1. 英伟达能够将计算、网络、互连和散热系统集成为完整机架，而AMD和博通目前更多以销售芯片为主。

1. 系统级交付使英伟达能够获取更高的单吉瓦收入，并提高客户黏性。

• 完整的软件生态构成关键壁垒

1. 英伟达通过CUDA、编译器、固件和应用框架持续优化硬件利用率。

1. AI计算并非所有指令都能并行化，需要在算法、指令集和硬件架构之间进行精细拆解。

1. 一次内存未命中可能造成大量计算周期闲置，因此软件优化对实际计算效率至关重要。

1. OpenAI、Anthropic等客户也会与英伟达工程师协作，进一步优化软件工作负载与底层硬件的适配。

b) 英伟达为什么需要销售完整机架，而不只是芯片？

• 单芯片功耗快速上升

1. Hopper的功耗约为700瓦。

1. Rubin预计功耗可能接近2,000至3,000瓦，传统风冷已难以满足要求，必须采用液冷。

• 散热系统成本显著提高

1. 完整机架成本中，约20%至30%可能来自热交换器、冷却板和液冷管路。

1. 传统风扇可能只需10至20美元，而冷却板成本可能达到200至1,000美元，部分情况下甚至接近芯片本身的成本。

• 高速互连和系统集成难度提高

1. 高功耗GPU集群需要同步解决供电、散热、网络和高速互连问题。

1. 英伟达通过整机架交付，将芯片、互连、交换机和散热系统统一设计，形成与单纯芯片供应商的明显差异。

c) 英伟达系统级模式能够带来多大的收入差异？

• 受访者援引行业估算称：

1. 建设1吉瓦级AI算力的综合投入约为600亿美元，包括土地、半导体和系统基础设施等成本。

1. 英伟达每吉瓦可能获得约300亿美元收入。

1. 博通或AMD每吉瓦对应的芯片收入可能约为100亿美元。

• 差异主要不是单颗芯片价格，而是英伟达销售完整机架，覆盖更多系统价值量；采购博通或AMD芯片后，客户仍需自行采购和集成其他系统组件。

3. AI软件团队与软硬件协同能力

a) 英伟达的软件团队规模和主要职责是什么？

• 团队规模估算

1. 受访者估计，英伟达相关软件工程团队可能在2,000至10,000人之间。

1. 具体人数取决于是否纳入底层编译器、固件、系统软件、高层API及应用优化团队。

• 主要工作层级

1. 底层指令集和编译器优化。

1. 固件、硬件配置、动态电压调节和热感知控制。

1. 指令级并行和计算流水线优化。

1. 训练、数据查看、预处理和清洗等应用工作负载优化。

1. GPU、CPU及多芯片之间的交互优化，包括预填充等流程。

1. 高层API、框架和客户应用适配。

• 生态系统进一步放大软件优势

1. 外部开发者和合作伙伴能够深入接触底层架构，通过寄存器和流水线优化提升效率。

1. 大型终端客户也拥有专门的软件工程团队，并与英伟达联合优化硬件利用率。

b) 谷歌在TPU合作中负责什么，博通负责什么？

• 谷歌的主要职责

1. 软件架构设计。

1. 工作负载定义。

1. 软件与指令集优化。

1. 与银行等企业客户合作，将客户需求转化为TPU优化方向。

1. 围绕每瓦性能、吞吐量和总体拥有成本开展系统优化。

• 博通的主要职责

1. 负责复杂芯片的设计实现、工程执行和量产交付。

1. 提供HBM控制器、SerDes、PCIe和高速网络连接等关键IP。

1. 协助谷歌完成TPU硬件的集成和迭代。

• 团队规模判断

1. 谷歌正在扩大自主TPU硬件团队，但硬件人员仍只是其软件团队的一小部分。

1. 受访者估计，谷歌与AI硬件相关的软件团队规模可能与英伟达接近。

4. GPU、ASIC与不同AI工作负载的适配

a) 为什么ASIC在特定工作负载上可能明显优于GPU？

• 性能优势来自软硬件共同定制

1. ASIC通常可比通用计算方案获得约2至10倍的效率优势。

1. 优势不一定来自更先进的制程或更强的绝对硬件性能，而是来自软件、指令集、工作负载及数据路径的针对性优化。

1. 即使ASIC的通用硬件能力只有前沿GPU的约30%，只要工作负载足够固定，也可能实现更高的整体效率。

• ASIC更适合可预测的任务

1. 当工作负载长期稳定、计算模式明确时，ASIC可以削减不必要的通用功能。

1. 在Decode等细分推理场景中，定制芯片有机会实现更高吞吐量和更低功耗。

• GPU更适合快速变化的模型

1. 软件模型大约每6个月就可能发生变化，而硬件开发周期即使顺利也需要约9个月。

1. GPU具备更强可编程性，能够适应模型、算子和工作负载持续变化。

1. 因此，GPU仍将承担训练、Prefill及尚未稳定的新型工作负载。

b) 未来是否会出现分别针对Prefill、Decode、Attention和FFN的ASIC？

• 细分ASIC会增加，但不会全面替代GPU

1. 随着推理需求扩大，针对Prefill、Decode、Attention或FFN等环节的专用芯片可能逐步增加。

1. 但由于模型结构和软件栈变化快，过度细分的ASIC面临开发完成后工作负载已经变化的风险。

1. 更可行的模式是GPU负责通用、复杂和快速变化的任务，ASIC负责明确且规模足够大的稳定任务。

• 大型AI公司将采用混合算力架构

1. OpenAI、Anthropic和谷歌预计都会同时采用GPU、定制TPU/XPU及CPU。

1. 受访者预计，英伟达GPU可能占大型算力部署的30%至50%，另有约30%可能由定制ASIC承担，其余由CPU或其他芯片承担。

1. 定制ASIC可能来自博通、Marvell、联发科或客户内部研发。

5. Cerebras晶圆级计算的优势与挑战

a) Cerebras的晶圆级计算有哪些优势？

• 配备大容量片上RAM和SRAM。

• 晶圆内部带宽较高，可减少传统多芯片之间的数据搬运。

• 在特定推理场景中，尤其是单用户Token生成速度方面，可能展现较高性能。

• 晶圆级方案已经实现实际部署，说明其技术具备一定工程可行性。

b) Cerebras面临哪些核心挑战？

• 良率和缺陷管理难度高

1. 一整块晶圆不可避免地存在缺陷单元。

1. 启动和测试时必须识别不良区域，并动态调整工作负载以避开缺陷。

1. 每块晶圆的缺陷分布不同，可能导致不同计算机之间的实际性能存在差异。

• 热机械与系统可靠性挑战

1. Cerebras早期曾面临较低良率和较多热机械问题。

1. 晶圆级封装、散热和运行稳定性的工程难度显著高于传统芯片。

• 软件优化复杂度较高

1. 矩阵乘法及其他计算指令需要根据晶圆中可用计算单元重新映射。

1. 受访者估计，其计算架构优化工作量可能比传统CUDA体系高出至少30%，极端情况下可能达到2倍。

• 部署规模仍较有限

1. 受访者判断，Cerebras当前全部产能可能仍不足1吉瓦。

1. 因此，其晶圆级计算虽然已实现商业部署，但整体规模仍明显小于主流GPU生态。

6. 定制芯片市场竞争格局

a) 博通、Marvell、联发科和高通的竞争优势分别是什么？

• 博通

1. 在以太网、PCIe、SerDes及高速连接领域积累超过20年。

1. 深度参与Scale-up基础设施及相关行业标准。

1. 拥有HBM控制器和高速互连等核心IP。

1. 芯片工程执行和量产速度较快，能够处理复杂的机械、电气及系统集成问题。

1. 与谷歌合作开发多代TPU，具备长期客户关系和成熟交付经验。

• Marvell

1. 被受访者视为定制芯片市场第二名，但与博通仍存在较大差距。

1. 在网络和中等性能计算方案方面具有积累。

1. 通过收购Inphi等公司补强互连和数据基础设施能力。

1. 近年来业务组合调整较多，部分以太网和汽车以太网业务已经退出。

• 高通

1. 长期主导调制解调器市场，在3G、4G、5G及无线通信IP方面优势明显。

1. 正尝试进入服务器市场，但目前缺少1千瓦以上高功耗AI芯片的开发经验。

• 联发科

1. 具备大型SoC设计和量产能力，并可能参与部分定制AI芯片项目。

1. 但会议中未对其具体客户和竞争地位作进一步量化说明。

b) 博通为何能在谷歌TPU项目中占据重要位置？

• 谷歌虽然可以使用自有品牌对外宣传TPU，但实际芯片工程和交付高度依赖博通。

• 受访者估计，博通可能承担谷歌TPU硬件90%至99%，甚至更高比例的实现工作。

• 谷歌正在尝试引入Marvell、联发科或加强内部研发，以开发不同TPU SKU并降低单一供应商依赖。

• TPU正在进一步细分为不同类型和不同SKU，以覆盖不同性能、功耗和应用需求。

• 若Anthropic采购谷歌TPU算力，底层硬件价值量仍可能主要由博通实现。

c) 博通主要XPU客户有哪些？

• 会议提及的4家主要客户为谷歌、Anthropic、OpenAI和Meta。

• 对于其余2家主要客户，受访者表示并不了解，即使了解也无法披露。

• 受访者提到Anthropic可能规划约5吉瓦算力采购，对应潜在订单金额约500亿美元，但该信息应以公司正式披露为准。

7. 博通AI半导体收入增长与可预测性

a) 博通XPU收入连续翻倍的指引是否可信？

• 受访者认为目标较为准确

1. 芯片项目在启动阶段就需要向台积电申报晶圆产能和最终客户。

1. 先进芯片产能已经提前分配，因此厂商可以根据晶圆供给和生产节奏预测未来收入。

1. 未来12个月的收入具有较高确定性，未来18个月也具备较强可预测性。

1. 在18个月基础上再外推约6个月，受访者认为当前增长轨迹仍有望延续。

• Hock Tan的指引被认为相对科学

1. 博通对芯片路线、客户需求及台积电产能有清晰规划。

1. 半导体业务实际表现预计将较为接近公司指引。

1. 出口管制是影响产品销售区域和收入兑现的重要外部变量。

b) 收入翻倍主要由销量还是ASP上涨驱动？

• 受访者粗略估计，仅就XPU业务而言：

1. 约50%的增长可能来自ASP上涨。

1. 其余约50%来自销量增加。

• ASP上涨的主要原因

1. XPU不仅包含核心计算芯片，也包括HBM等存储器。

1. 存储器成本可能接近XPU整体ASP的一半。

1. HBM仍处于短缺状态，海力士、美光等厂商正在扩大产量，存储价格对XPU售价影响明显。

1. 晶圆成本上涨也会推升新一代产品价格。

c) 博通XPU业务的利润率如何？

• 核心计算部分利润率较高

1. 博通通常努力使产品达到约80%的利润率。

1. XPU核心计算芯片部分的利润率可能接近这一水平。

• HBM集成会拉低整体利润率

1. 内存部分基本处于盈亏平衡或只有少量利润。

1. 无论集成8层还是12层HBM，博通在存储器部分获得的利润均较有限。

1. HBM价格上涨虽然会推高ASP，但不会带来同比例利润增长。

1. 因此，XPU整体利润率预计低于纯核心计算芯片的利润率。

8. AI硬件估值、资本开支与潜在风险

a) 为什么英伟达和博通收入高速增长，股价却没有同步翻倍？

• 增长预期已经部分反映在估值中

1. 市场此前已知晓收入翻倍和下一年度继续增长的预期。

1. 本季度缺少超出既有预期的新信息，因此业绩增长未能转化为同等幅度的股价上涨。

• 云服务商资本开支可能在2028年前后见顶

1. AI硬件最终需求和资金来源主要是大型云服务商的资本开支。

1. 市场担忧云服务商资本开支增速可能在2028年见顶。

1. 2028年之后算力投入能否继续高速增长，目前缺乏明确判断。

1. 这一不确定性压制了整个AI硬件板块的估值和Beta。

• 盈利能力可能低于收入增长

1. 产品组合变化、HBM成本、良率和系统集成成本可能对利润率形成压力。

1. 收入翻倍并不意味着利润和市值可以同比例翻倍。

b) AI算力客户的商业模式存在哪些风险？

• 部分客户尚未实现盈利

1. Anthropic和OpenAI均属于大型AI算力采购方，但目前并非稳定盈利的上市公司。

1. 受访者认为，约50%的算力客户所使用的资金可能并非真正意义上的自有资金。

1. 这些客户提供的AI服务也未必已经形成可持续利润。

• 大额订单存在客户消失风险

1. 最大风险并不是短期芯片交付问题，而是原计划采购数吉瓦算力的客户突然削减订单或退出市场。

1. 若AI订阅需求下降、融资环境恶化或客户破产，可能对芯片供应商和产能链条形成冲击。

1. 由于先进制程及HBM产能需要提前锁定，大客户订单变化可能造成较大库存和资本风险。

• 政策风险不可忽视

1. 政府可能因AI安全、军事用途或社会影响加强监管。

1. 出口管制也会限制部分高性能产品进入特定市场。

1. 若没有政策干预，未来18个月增长路径较明确；若出现强监管，行业增速可能放缓。

结构化总结表格

新版prompt提炼之后的转译

提问者：首先，请您分享一下对 GPU 和 ASIC 技术趋势的看法。目前已出现 Prefill-Decode 解耦等趋势，也有许多 ASIC 厂商采用不同的路线图、技术和功能。未来 2至3年，GPU 和 ASIC 技术将如何演进？

回　答　者：从整体趋势和近期趋势来看，AGI 和 AI 将拥有广泛的应用，并对劳动力、银行业、日常生活、各类智能与建模以及医疗等领域产生革命性影响。AI 将增长至2倍，市场对 AI 及其应用的需求似乎永无止境。

在生成这些模型的过程中，业界近期发现，用于模型训练的算力越强，模型的表现就越好，而这需要硬件支撑。因此，市场对 AI 硬件建设产生了巨大需求。

例如，英伟达预计其硬件销售额和收入将增长至2倍。博通也呈现相同趋势，其半导体业务同比增长至2倍，今年较去年增长至2倍，并预计将再次增长至2倍。大约1个月或2个月前的趋势和路线图显示，这种增长至2倍的态势还将持续，整体需求规模非常惊人。

回　答　者：硬件需求主要来自AI。过去2周，市场出现了一些反对意见、限制和不确定性，也可能出台政府监管措施，以减缓AI部署。此外，AI还可应用于军事领域。大约2周前或1周前，行业趋势仍是对AI及其应用存在近乎无限的需求，同时对硬件也有近乎无限的需求。

从某种意义上说，真正推动AI变革的是GPU架构。通过并行化以及多重矩阵乘法，可以显著增加运算量。相关需求可进一步分为训练和推理。至少到近期，需求可能更多集中在训练领域。这推动了单个GPU性能的提升，同时也导致功耗进一步增加。不过，未来训练与推理的需求结构将更多转向推理硬件，而训练硬件的占比将有所下降。

从主要参与者来看，美国市场包括Nvidia、AMD、Google和Broadcom等公司。Broadcom也生产大量相关芯片，包括TPU和XPU。

提问者：您提到，近期需求可能更多集中在训练领域。应如何理解这一判断？无论是行业共识、普遍观点，还是资本市场的看法，需求结构似乎都在如您所说的那样转向推理。为什么您认为近期训练领域的集中度更高？

回　答　者：我的意思是，未来训练需求的增速将远低于推理需求。至少到近期，行业一直在努力构建越来越好的模型。但在某个阶段，可针对不同应用构建的模型数量将相对有限。随着新数据不断产生，这些模型仍会得到完善，并需要基于新数据继续构建模型。不过，模型生成最终会趋于饱和，届时用户侧推理需求和相关应用将持续扩展，其增速将远快于训练需求。

提问者：请分享一下您对英伟达、AMD、谷歌、博通和Cerebras等主要参与者优劣势的看法。谷歌可能是博通最大的客户之一，对吗？我们也对Cerebras感兴趣。您能否就上述4家主要参与者分享一下观点？

回　答　者：我先谈谈对英伟达的看法。英伟达是行业中的既有领导者，开发GPU已有近30年。其业务最初涉及游戏主机，也曾尝试进入移动领域。大约在2015年，英伟达的产品开始应用于加密货币和加密货币挖矿。凭借高度并行的指令架构，其计算吞吐量远高于当时的其他产品。

下一个重要里程碑是在2019年左右，英伟达成功利用这些训练模型取得了令人难以置信的成果。以上是对其发展历程的简要概括，如有需要，我也可以进一步回顾其更早期的历史。

目前，英伟达的产品已经发展到Hopper、Vera Rubin和Feynman。英伟达正在构建机架级解决方案。据我所知，目前只有这家公司不仅能够提供GPU，还能提供CPU、网络交换机以及互连系统中的其他芯片，并将这些组件整合到一个机架中。

在散热方面，预计Rubin的功耗将达到2,000瓦或3,000瓦，无法采用风冷，而需要使用液冷解决方案。

提问者：抱歉，我刚才断线了大约1分钟。您能否从Hopper部分重新开始？

回　答　者：目前Hopper的功耗约为700瓦。英伟达解决方案与其他解决方案之间的一项重要差异在于，英伟达销售的是完整机架，而不是单独销售芯片。博通销售芯片，AMD也销售芯片；相关芯片会配备HBM等组件，但英伟达现在销售的是完整机架。

回　答　者：转向 Rubin 后，功耗接近 3,000 瓦，因此需要集成式冷却系统，而这套冷却系统并不简单。目前估计，整个机架成本中约有 20%至30%来自热交换器、冷却板以及液冷管路。作为对比，以前风扇的价格可能是 10美元、15美元或20美元，而冷却板的价格为 200美元至1,000美元，甚至可能与芯片本身一样昂贵。

英伟达由于功耗及系统集成方面的要求，包括高速互连，需要销售整套机架。这是英伟达与 AMD、博通、谷歌之间最大的区别，Cerebras 的情况则有所不同。

另一个角度是，有 1项估算认为，行业正在尝试销售吉瓦级算力，相关成本约为 600亿美元；也就是说，产生 1吉瓦算力大约需要 600亿美元，其中包括土地成本，以及部分半导体和系统成本。据估计，英伟达每吉瓦可获得约 300亿美元收入，而博通或 AMD 每吉瓦产生的收入约为 100亿美元。两者的差异主要在于英伟达销售整套机架解决方案，而购买博通、AMD 芯片时，还需要额外承担其他系统成本。这是其中一个差异。

提问者：您能否快速介绍一下还希望与我们分享的内容？接下来，我们希望重点讨论这些芯片如何工作，例如 GPU 如何与 ASIC 协同，以及它们如何适配不同类型的 LLM 和其他模型。您刚才分享的内容非常有帮助，请继续。

回　答　者：简单谈一下英伟达，其硬件的效用在很大程度上取决于编程模型。不同指令具有不同特性，并非所有指令都可以并行化。

回　答　者：必须对算法和计算指令集架构进行智能拆解，以充分利用硬件。晶体管利用效率的一个重要影响因素是避免停顿。例如，1次内存未命中可能导致超过1,000,000个周期的计算闲置。

因此，计算软件架构，如CUDA及其衍生工具，会由英伟达规模较大的软件团队持续开发，通过新的框架提高硬件利用率，并由此形成生态系统。这正是谷歌认为博通能够提供的重要价值。

提问者：您提到这是谷歌认为博通能够提供的价值，应当如何理解？在谷歌与博通的合作中，双方分别承担什么角色，具体工作由谁完成？

回　答　者：软件架构、工作负载，以及软件和指令集的优化工作由谷歌完成。

提问者：您是否了解英伟达CUDA软件团队，以及谷歌类似团队的规模？这些团队大约有多少名员工或工程师？

回　答　者：我并不在英伟达工作，但估计其团队至少有2至10,000人。团队的扩展能力取决于具体分工。

英伟达内部有工程师负责底层指令集优化和编译器优化，也有工程师负责固件及硬件配置。他们可以根据热感知和优化需求提高电压，进行动态电压调节，这些都属于底层计算控制。

在此之上，还有更高层级的优化，例如前面提到的指令级并行。此外，还包括基于应用的优化，例如处理训练、查看和预清洗等工作负载。

回　答　者：在数据和系统层面，系统软件工程师会优化不同芯片之间的交互。例如，在同时使用GPU和CPU时，他们会优化相关控制和预填充流程，这部分至少需要2,000名工程师。再考虑高层API，相关人数可能是2、2,000，甚至10,000。

此外，英伟达还拥有生态系统，其中一些公司比普通用户更深入地接触底层硬件架构，能够通过优化寄存器和流水线来改善工作负载。再往上是终端客户。OpenAI、Anthropic及其他许多公司都有工程师负责优化硬件，并与英伟达工程师合作，使硬件更好地适配软件工作负载。以上是我对英伟达内部工程师规模的估计。

提问者：我认为值得关注的是，这类软件工程师的数量可能不会大幅增长，而且英伟达正在尝试利用大量AI技术优化许多软件功能。谷歌的情况如何？

回　答　者：您是想了解谷歌硬件工程师与软件工程师的人员构成吗？

提问者：是的。我想了解与英伟达类似的软件工程师团队，以及谷歌的团队规模。

回　答　者：谷歌目前正在进行一些调整。公司内部正推动自主开发TPU，因此也在扩充硬件团队，但其规模仍然只是软件团队的一小部分。

即使在软件团队内部，也有许多应用工程师。他们会与美国银行等传统企业客户合作，了解其工作负载，并将这些需求转化为TPU硬件的优化方向，从而在谷歌云中实现更高的每瓦运算性能、更高的吞吐量，或改善其他总体拥有成本指标。

就团队规模而言，我估计可能与英伟达相同。

提问者：您能否介绍一下，不同GPU或ASIC分别适合哪些工作负载？例如，英伟达刚刚收购了Groq，那么Groq的LPU和Cerebras的WSE是否更适合快速推理，而英伟达和AMD的GPU可能更适合预填充？不同类型的GPU和ASIC与不同工作负载之间存在怎样的关系？

回　答　者：Lucy提到了Cerebras，我先就此谈几点看法。Cerebras是一家初创公司，打造了一种晶圆级的新型计算机。我见证了它的发展，也认识一些早期加入该公司的工程师。Cerebras一度面临良率很低以及大量热机械问题，产品甚至无法正常运行。后来，一些员工并非出于自身意愿离开了公司，转而加入AMD或谷歌。

Cerebras最终实现IPO，这让我非常意外。据我了解，其能够实现IPO的主要原因是OpenAI为Cerebras的生产注入了支持。目前，Cerebras的全部产能甚至不足1吉瓦，因此其晶圆级计算的部署规模其实非常小。令人意外的是，这种晶圆级计算竟然已经实现了实际部署。当然，国家实验室可以对其进行评估，但它至今仍能持续运营并以IPO公司的形式存续，确实令人难以置信。

从良率来看，典型缺陷密度约为0.1。如果其符合台积电的典型缺陷密度，那么一块晶圆上会有大量芯片或部分裸片无法正常工作。因此，在晶圆启动和特性测试过程中，必须标记不良单元及晶圆上的缺陷区域，并根据缺陷情况动态调整工作负载。每块晶圆都会存在缺陷，必须采取相应的规避措施。这也意味着每块晶圆和每台计算机的性能都可能不同，从而带来更多挑战。

回　答　者：在计算优化方面，这比CUDA更困难。例如，编写矩阵乘法软件时，需要对指令进行映射以充分利用硬件。要真正发挥这些计算资源的能力，可能至少需要增加30%的计算机架构优化工作，甚至可能达到2X。因此，从这个角度看，Cerebras能够存续至今本身就堪称奇迹。我对市场对Cerebras表现出的兴趣感到非常意外。

提问者：从我们的角度看，我们也对Cerebras进行了一些研究。我们认为，Cerebras既有优势，也有劣势。其芯片配备容量非常大的片上RAM和SRAM，晶圆内部带宽也很高，但晶圆之间的I/O带宽可能较低。Cerebras披露的一些数据显示，其单用户每秒生成的Token数量相当高。您如何看待我们在讨论开始时提到的问题？未来GPU和“6”的发展趋势是什么？根据我们的理解，Cerebras更专注于Decode，对吗？推理过程可以分为Prefill和Decode，而Cerebras更侧重于Decode。英伟达收购Groq后，在GPU与Groq LPU的组合中，英伟达GPU可能更专注于Prefill和Attention，Groq LPU则更专注于Decode或FFM。您能否分享一下对这一趋势的看法？此外，OpenAI刚刚发布了Colono，但它似乎并未采用PD，即Prefill与Decode聚合的方式。未来是否会出现更多细分用途的ASIC，例如一类ASIC负责Prefill，另一类ASIC负责Decode、Attention或FFM？

提问者：从您的角度来看，您如何看待这一趋势？

回　答　者：Meta的几位副总裁曾对此进行总结，他们的观点与我基本一致：软件模型每6个月就会发生变化。当然，整体上仍会分为训练和推理两大类别，但硬件即使进展顺利，也需要9个月的开发周期。

因此，英伟达销售的系统具备一定的可编程性，类似于带有加速能力的CPU。英伟达通过NVLink将GPU集群连接起来，同时也拥有自己的CPU，不过其性能并不像传统英特尔CPU那样构成系统的核心优势。总体而言，英伟达的硬件具备一定的可编程性。

以Jalapeno为例，从功耗数据来看，其功耗不到1千瓦，约为700瓦，即700瓦。因此，它实际上并没有英伟达当前的解决方案强大，甚至不及目前的GPU，但它是针对特定工作负载定制的。从这个意义上说，其性能可能明显优于采用未充分优化软件的通用处理器。

ASIC的性能通常比通用计算高出2到10倍。这并不一定是因为硬件本身更出色，而是因为软件、工作负载、指令集，甚至数据路径都针对特定工作负载进行了优化。当工作负载可以预先确定时，即使硬件性能可能只有前沿水平的30%，也能够显著提升整体性能。

未来，OpenAI、Anthropic乃至Google等公司都将同时采用GPU和其他芯片。目前的预测是，英伟达GPU所占比例可能为30，即30%或50%，另外可能有30%是ASIC。这些ASIC主要是定制TPU和Halapeno等产品，可能来自博通、Marvell、内部研发，甚至联发科，用于为可预先预测的特定细分工作负载提供定制芯片。

回　答　者：此外，还会有一些通用计算CPU，用于不同的工作负载。

提问者：明白。从博通的角度看，您提到Anthropic未来也会像OpenAI一样拥有自己的ASIC。我的理解正确吗？

回　答　者：我希望这是公开信息。如果不是，我认为Hock会非常自豪地表示，Anthropic预计将向博通预订5吉瓦的计算资源。这实际上是一笔非常大的订单，按此计算，单个客户的订单金额就达到50 billion美元。

提问者：是的，我认为Hock曾经披露过这一信息。他可能在9月初高盛举办的会议上分享过类似内容。不过，我原以为Anthropic采购的是TPU，因此不太确定。这5吉瓦对应的是Anthropic的ASIC，还是谷歌的TPU？

回　答　者：据我所知，谷歌实际上并不完全自行制造硬件。谷歌可以为硬件使用自己的品牌。Hock并不在意谷歌从营销角度宣称“我们自行制造硬件”，但实际上，博通可能交付了谷歌TPU硬件的90、9，至少甚至可能是99%；也就是90%至99%，甚至可能达到100%。

目前有一些新闻稿称，谷歌将与Marvell或联发科合作，或者尝试在内部开发部分TPU及不同SKU。现在，TPU正在分为2种类型，其中1种是E和P。与许多产品类似，每一代TPU都会有不同的SKU。

因此，如果Anthropic采购谷歌的TPU计算资源，这些资源实际上来自博通。

提问者：明白。目前定制芯片市场有多家厂商，包括博通、Marvell、联发科，以及高通。联发科并非美国公司。您认为其中最重要的是……

提问者：这些定制芯片厂商最重要的竞争优势是什么？例如，它们的IP、技术诀窍以及客户关系等方面有哪些主要优势？

回　答　者：高通一直是调制解调器领域的主导厂商，在无线通信技术方面拥有大量IP。高通在5G领域具备优势，此前在5G、4G和3G领域也非常成功，目前正尝试进入服务器市场。但高通并不生产功耗在1千瓦以上的芯片。

我认为，在定制芯片领域，Marvell排名第2位，但与博通相比，可能是差距较大的第2位。Marvell也经历了许多业务调整。例如，我认识一些曾在其以太网业务和汽车以太网业务任职的人员，而Marvell已经退出了这些业务。此后，Marvell收购了Celesti AI，并在几年前收购了Inphi。不过，Marvell通常与中等性能的计算解决方案相关，尤其是在网络领域。

博通深耕网络领域已有20多年，是以太网生态系统中的主导厂商。博通已经采用并深入参与面向新型Scale-up基础设施的多项标准。除以太网外，博通还拥有PCIe技术和大量连接技术。过去，无线业务约占博通半导体收入的50%，其中很大一部分来自为苹果提供的产品。

目前，博通的业务重心正在转向把握人工智能浪潮，尤其是XPU业务。Hawk曾表示，博通在10年前与谷歌启动了一个相关的外围项目，为其开发了数代TPU，但该项目当时的收入很少。我认为这一说法属实。如今，谷歌已经认识到TPU在内部计算资源利用方面的价值，同时也越来越多地将其用于云服务和主要客户。因此，这项业务迅速增长，并在博通半导体业务中占据主导地位。

总体而言，博通的许多优势来自其IP，包括连接HBM的控制器以及SerDes等技术。

回　答　者：此外，其执行速度也很快。实际上，芯片制造在机械、电气和集成层面都极其困难，整个工作流程也非常复杂。谷歌原本可能认为开发芯片并不困难，但现在也逐渐认识到，开发这类复杂的GPU非常不容易。

提问者：关于博通，9月初在分析师电话会上，Hock表示博通拥有6家主要XPU客户，并提到了其中4家，分别是谷歌、Anthropic、OpenAI和Meta。其余2家主要客户是谁？

回　答　者：无论相关信息是否公开，即使我知道，也不能透露。事实上，我并不知道其余2家客户是谁；即使知道，也不能披露。

提问者：能否大致限定这些客户的范围？例如，它们是否属于美国的云服务提供商？任何相关信息都可以。

回　答　者：即使我知道，也不能透露；同时，我确实不知道相关信息。

提问者：正如您在本次通话开始时提到的，Hock还指引称，XPU收入，也就是AI半导体收入，将在本财年增长至2倍，并在下一财年再次增长至2倍。您如何看待这一指引？您认为这一目标是激进还是保守？

回　答　者：我认为这是一个准确的目标。原因在于，我们已经获得台积电的产能分配，因此能够预估晶圆产能及后续生产情况，并在将近18个月的周期内较为准确地预测收入。芯片项目启动时就会与台积电进行接洽，并申报最终客户或最终客户群体。

回　答　者：他们对自身芯片的发展方向有清晰且完善的规划。与此同时，出口管制也带来了一些限制，部分产品无法出口至特定地区。因此，对于未来1年的预测，我认为确定性很高。Hock一直非常透明，整个生态系统可能也掌握了相关情况，包括台积电。我认为，未来18个月的收入具有较高的可预测性。

提问者：明白。从现在起18个月后，考虑到Hock已经给出指引，而博通的财年是从11月到次年10月，其定制芯片业务在第28财年的收入将达到230十亿美元，其中包括来自ASMI的收入。由于这距离现在接近24个月，该目标是否也具有很高的可预测性？

回　答　者：我仍坚持此前的观点。虽然我提到的是18个月，但至少未来12个月具有较高的可预测性，未来18个月也可以预测。在此基础上进行外推，额外6个月并不算太久，因此可以合理认为，这一发展轨迹将会延续。

提问者：您如何看待未来GPU和ASIC的竞争格局？正如之前提到的，我们关注高速推理ASIC，包括Cerebras、英伟达的Grak，以及其他规模较小的参与者。例如，AMD已经收购Talos。您对未来GPU和ASIC细分市场的竞争格局有何看法？

回　答　者：请稍后提醒我回到这个问题。在此之前，我想先谈谈自己对本季度业绩和博通股价的看法，您可能会感兴趣。公司收入实现翻倍，并预计明年收入将再次翻倍。按照通常的理解，股价似乎也应该翻倍，这也是大家原本的预期，但实际上股价反而下跌了。

提问者：我认为，Broadcom股价本季度没有翻倍，甚至没有上涨，主要是因为没有新的消息。上季度，公司表示收入非常强劲，明年将实现翻倍；本季度仍然是同样的表述，因此市场没有看到新的信息。在我看来，这就是Gcom股价没有上涨、表现持平的原因。英伟达的情况可能也类似。英伟达表示，其收入较去年同期翻倍，明年还将再次翻倍。按理说，这样的增长应该令市场震惊，市值似乎也应当翻倍，但实际上并没有发生。您认为原因是什么？我认为可能有2个原因：第1，英伟达的市值已经非常高，整个经济体系中几乎没有足够的资金支持其市值继续翻倍；第2，尽管其市值和收入还会上升，但利润率可能下降。英伟达正在进一步进入PC GPU市场，而该业务的利润率几乎为负，至多持平。此外，其产品良率可能也没有预期那么高。您如何看待英伟达和Broadcom在收入有望翻倍的情况下，股价和市值却没有同步上涨，甚至出现下跌或趋于饱和的现象？

回　答　者：这是另一个值得讨论的话题。我认为，整体而言，AI硬件板块的Beta在一定程度上受到美国主要云服务提供商资本支出增长前景的压制。主要云服务提供商的资本支出增速可能会在2028年见顶，此后的情况目前尚不明确，也较难预测。Broadcom在2028年的收入可能仍会翻倍，这一点确实令人意外。但如果从整个市场，以及作为所有AI硬件最终客户和最终付费方的云服务提供商资本支出来看，其资本支出增速届时可能见顶，而之后会如何发展，目前没有人能够判断。

提问者：市场对2028年后的增长存在疑问，这给所有AI硬件股票带来了很大压力。不过，这是另一个话题。回到今天的讨论，Broadcom的AI半导体收入预计将在明年以及2028财年翻倍，这一增速非常高。因为从我们的角度来看，台积电的CoWoS产能在明年和2028年肯定不会翻倍。Broadcom是否在台积电获得了更多产能？此外，Broadcom产品的平均销售价格，即ASP，是否也会上升？收入连续翻倍意味着非常高的增长，原因是什么？

回　答　者：这里涉及几个因素。首先，受存储器影响，ASP正在上升。XPU不只包括计算芯片，还包括存储器，而存储器成本接近ASP的1/2。此外，HBM仍然短缺，海力士和美光持续希望增加产量，这对ASP有很大影响，同时也会降低XPU的利润率。

提问者：Broadcom在台积电获得的产能情况如何？

回　答　者：这正是其收入具有较高可预测性的原因。对于相关指引，我认为确定性远高于不确定性，至少半导体业务的实际表现会非常接近指引，软件业务可能也是如此。

Hock的指引实际上非常科学，具有较高的可预测性。英伟达未来12个月的情况也类似，因为相关产能已经确定，英伟达正在建设这些机架，并且已经获得预订单。

关于Broadcom收入预计翻倍，但股价涨幅未达到预期，我还有一个推测，其中可能存在风险。例如，Anthropic规划了5吉瓦的算力，但它不是上市公司，也没有盈利。OpenAI也是Broadcom的大客户，但同样没有盈利。

提问者：这种商业模式是什么？如果所有人都放弃 OpenAI，不再购买订阅，或者 OpenAI 最终破产，会带来哪些风险？

回　答　者：我认为，最大的风险可能在于，约 50% 的算力客户所使用的资金并非真正意义上的自有资金，其服务也无法产生实际利润。这可能是一个风险，也是我的个人判断。

真正最大的风险是，原计划需要交付 5 吉瓦算力的客户突然消失。如果一切按照既定轨迹发展，政府也没有对 AI 产生担忧，那么在没有外部力量干预的情况下，这将是未来 18 个月内的预期发展轨迹。

提问者：最后一个问题。能否大致估算一下增长的构成？在博通增长翻倍再翻倍的过程中，多少由销量驱动，多少由平均售价上涨驱动？粗略估算即可。

回　答　者：我认为，平均售价上涨可能贡献约 50%，其余增长来自销量。

这一估算仅针对 XPU。实际上，博通相当一部分收入来自网络业务，与 XPU 无关。其最新产品的平均售价也会上涨，部分原因是晶圆成本上升。

博通通常会努力使产品实现 80% 的利润率，但由于涉及 HBM，XPU 的利润率相对较低。XPU 核心计算部分的利润率可能为 80%，但在集成内存方面基本处于盈亏平衡状态，或者利润非常有限，无论集成的是 8 层还是 12 层内存。

即使内存价格上涨，也不意味着博通会因此获得20。如果内存价格上涨 50%，也不意味着博通会获得 60% 的利润，因为内存成本部分基本没有利润。

原文转译

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， 不得涉及影响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，感谢您的理解和支持。 OK. Uh, expert, um, hello. Uh, so, uh, first of all, uh, could you please uh share with us your view on the technology trends of uh GPU and ASIC? I mean, um, like, uh, there there are uh trends like uh, pre-fielded, pre-fielded decoup disaggregation and there are uh many ASIC players uh uh right now, uh, uh having different different roadmaps, different uh different technologies. They have different like different features. So, um, yeah, could you please uh generally please uh uh share with us a high high level, high high level view on uh how will the technology in in GPU and ASIC uh evolve in the maybe next uh two to three years? Okay. Yeah, let let me um, you know, kind of uh talk a little bit um, about, you know, um, you know, maybe uh, you know, general trends and um, you know, recent trends. Um, you know, the general trend is that the, you know, AGI and AI is going to have, you know, so many applications, um, you know, uh, and and revolutionize the workforce, you know, banking, um, um, you know, uh, daily life, um, intelligence, all sorts of intelligence, modeling, um, that the applications, um, you know, to, you know, medicine and so forth, that that uh, you know, the the general trend is that AI is just going to double. There's a insatiable demand for um AI and its applications. And really that um, to generate these models requires um, you know, they found recently that um, it the the better your compute, um, to to generate the uh training, the uh the model with training, um, the better better it performs. Um, and then that requires hardware. Um, and um, and really that's led to, you know, this, you know, tremendous demand for AI hardware buildup. Um, and really, you know, there, you know, for instance, Nvidia anticipates a uh doubling in uh sales and revenue of of um, uh of of their hardware. Uh, Broadcom follows the same suite of a doubling its semiconductor, um, business, um, year-over-year. It doubled this year from last year, and it anticipates to double again. And the trend, um, and the road map, uh, and a near, you know, maybe a couple of month or two ago, was that it will continue to double. So it's just a unbelievable amount of um, the demand in uh hardware. uh, which is which is AI. Um, but then I think recently, just maybe the last couple of weeks, um there has been some push back and some constraints and you know, uncertainty and maybe some governmental regulations to slow down AI deployment. And also, you know, it has um applications in the military. Um, so this this really has um, um, you know, up to maybe a couple of week weeks ago, um, or a week ago, um, the trend was we we have an insatiable demand for AI and applications and insatiable demand for hardware. So now, you know, really the the uh, you know, what really has um, revolutionized, you know, AI in some ways is the uh the GPU architecture, right? And you're you're now parallelizing um and uh getting a significantly more operations by um, um, uh with these operations and multiple uh multiply multiply uh uh matrix, matrix multiplication. Um, and and then that decomposes into training and and inference. And uh for, you know, probably, you know, till recently, uh there's probably even uh higher concentration in training. Um, and these this has led to more powerful um per uh per GPU, uh and that lets also leads to higher uh power uh higher power consumption. But I think in the trend, the um the mixture of training and inference is going to lean more towards inference hardware and uh less in in in training hardware. So, so so then uh we now dig a little bit deeper and look at, you know, some of the uh, you know, major players, right? We have Nvidia, AMD, uh Google, um, in in the United States, Broadcom, also um making a lot of the um, the uh silicon, um, also TPU, XPUs. Um, and um, you know, so, um, let me pause here, um, if I've given sufficient background. Um, and maybe you can give me a little guidance about uh where you want me to go. Um, how how does it Yeah, sure. Yeah, sure. that makes sense. I understood. got you. Um, yeah, so, um, so I have a a quick follow up here. you you mentioned that um recently um, let me see. you you you said that the the the recently there there's probably even higher concentration in training. How um how do we uh how do we understand that? Uh I mean, maybe the the consensus or a general view uh of the industry or I mean from from maybe capital markets perspective, the the the mix is towards as you mentioned, uh towards um uh inference. So, yeah, so why why why you you you said that uh recently there was higher in training. Right, so I think uh the the growth in um, in training is going to be much slower than inference. Um, is is what I mean. For for um, you know, till recently, um, it's been trying to build better and better models. But um, at some point, they're going to um, kind of there aren't going to there aren't that many models that can that can be built for for different applications. And they will be refined um, requiring with new data coming in to build models. But at some point, um, the the model generation is going to saturate and then the user the inference and then, you know, applications will will, you know, just multiply and grow much faster um, than than training. That's what I mean. Okay, got you. Got it. Okay, so maybe uh... Uh from here, we I I'd like to um uh could you please share with us your maybe uh your your view on the uh advantages advantages and disadvantages of uh of of several major players like uh Nvidia, AMD, um Google, because I think Google maybe Google is one is one of the largest customers of Broadcom, right? Uh Google and uh and maybe we are we are interested in Cerebras. Yeah, so maybe let's uh maybe let's yeah, yeah, could you please share with with us on the the four major players. Yeah, okay, so uh I I know Nvidia, AMD and then you said Google, right? Google and Broadcom. Google, maybe Broadcom and then was there someone else you uh service? Cerebras, Cerebras. Right, Cerebras. Let me let me um, yeah. Let me uh give you my perspective on uh Nvidia, uh uh to start out with. I mean, they're the incumbent and um really they have been building GPUs for almost 30 years. Uh you know, gaming consoles, they try to go into mobile, but really what happened in like about 2015, actually it was, you know, they found applications in cryptocurrency and and crypto uh mining. And they were able to to uh leverage uh um their architecture of uh highly parallel instructions to um, you know, have much a higher uh compute throughput than anything else out there. Um, and really then the, you know, the next milestone really is around 2019 when they were able to um, you know, generate these training models uh with unbelievable success. Um, and really, um, you know, that I mean, I'm just giving you a quick summary and you can, you know, if you want to rewind the the history much any further, uh let me know. But, you know, now, you know, we're we're, you know, looking at, you know, Hopper, you know, um the um Vera Ruben, um and then, you know, uh Fiman. So, uh really what the big um, you know, and what they're doing is is really building rack level solutions. This is a very, I think this is the only company that I know of right now that is building not only the uh the GPU, and the CPU, and the network switch, Um and then there are a couple other uh pieces of chips in the interconnect, right? Um, all together into a rack. So, I mean, just to just to, you know, you know, make sure, you know, it's understood. The the the cooling solutions for, you know, right now they're they anticipate these uh, you know, the uh Reuben is supposed to be like two two or 3,000 watts. That's not going to be uh uh uh air cooled, right? These are going to be liquid cooled um solutions. And and liquid cooling is, you know, the Sorry, expert. Um I was disconnected for maybe a minute. I'm sorry I'm so sorry about that. Um I I could could you just uh start restart from the uh from from Hopper from uh I mean, I was just I was disconnected uh Okay. just now for maybe a minute. Yeah, so right now Hopper Hopper is at around 700 I mean like one of one of the motivations and then like the deltas between Nvidia solutions and other solutions is Nvidia is selling a whole rack. like they don't sell chips. Like uh Broadcom will sell a chip. AMD will sell a chip. Uh, you know, the chip has got HBM and it's got a lot of things. But Nvidia sells the whole rack now. From so really the Delta is now Hopper was around 700 watts. Moving moving to, you know, Ruba uh, Ruben, um, it's almost 3,000 watts. So, so really that requires a integrated uh cooling system. And and the cooling system is is non-trivial. I mean, right right now they estimate that the the ra- uh of the um of the whole rack um cost, about 20, 30% of the uh cost of the rack is due to the uh heat exchange, the um the cooling plate, the the piping involved in liquid cooling. And and just to just to give another um, you know, um, you know, um, you know, perspective is previously fans, maybe like 10, 15, $20, a cooling plate is $200 to $1,000. It's I mean, it can be just as expensive as the Silicon. So, so really now um, the the big delta is that uh Nvidia um because of its the power consumption and the integration required, including the high-speed interconnect, that um, it it's selling a whole rack. So that's the biggest delta um in um in Nvidia to AMD to to um, you know, uh Broadcom to Google. And then um, Cerebras is a little is different. Um, but I so so really that that's the um and and here's another perspective is that the, you know, one one estimate is that it takes, you know, we we talk about giga, you know, we're trying to sell gigawatts of compute. They say estimate that about $60 billion dollars, it takes about $60 billion US dollars to generate one gigawatt of compute. Of that there's land, you know, and then some of it's semiconductor and systems. Um, the estimate is that Nvidia will um will will take about $30 billion per gigawatt. Where whereas Broadcom or AMD um per gigawatt is selling about is is uh generating about $10 billion of revenue. So, so really the delta is because Nvidia is selling a whole rack solution. Whereas um a lot of the revenue um that or or the, you know, system cost um, is is uh, you know, there's an adder in the system cost when you buy chips, um, from Broadcom, AMD, um, um, yeah. So, so that's kind of one delta. Um, AMD is but but it's kind of an interest Yeah, so I think uh AMD, okay, maybe maybe let me pause real quick. Um, yeah, uh, um, yeah, could you please quickly, maybe quickly go through your, I mean your what you want to uh share with us. uh, uh, and next we'll let's focus more on like the how does those um, um chips work, like the how how how how does GPU work with an ASIC, work with uh different, maybe different uh kind of different kinds of uh LLMs, different kinds of models. uh um maybe uh I'm we are more interested in uh in in those areas. Yeah, but but but what you have shared with us is very very helpful. Yeah. And please go, go ahead, go ahead with your uh uh your your your, yeah. Yeah, so then, you know, just uh, you know, touching quickly on Nvidia, it's it's it's utility of of its hardware is actually uh, you know, largely determined by by uh its uh um programming model. Because, you know, typically so there are different, you know, instructions and then um, and and not all instructions are parallelizable. Um, and and you have to intelligently decompose a algorithm uh and a compute uh uh uh you know, ISA um to leverage the hardware. You know, a big thing in um, you know, utility of the, you know, just the transistors is is the um uh um uh you know, preventing stalls. Um and and for instance a a memory miss would could lead to like a over a million cycles of um of, you know, idle compute. So, so just just uh what I'm trying to say is that the uh compute the the software architecture, uh CUDA for instance, um and then it's many derivatives that are always, you know, generated by its um uh you know, actually a pretty large software team in in Nvidia, um will will uh develop um, you know, new um frameworks to utilize its hardware. Um, and that creates an ecosystem. And that's where um a lot of the value that uh Broadcom um perceives for um for I mean uh Google perceives for Broadcom. So, um, um, does that um, is that should I move on to a little bit about broad time? Oh, I have a follow up here. I have a follow up question here. Yeah, how do I how do we understand what you have said the the you said that um um that's the value that Broadcom perceives for for Google. uh sorry, Google perceives Broadcom. How do we understand that? because you you said that maybe I mean, what uh so what's the like the the the the the role uh among between uh Google and Broad? who who who who do the job uh between Google and Broadcom. Google and Broadcom. That would be Broadcom that would be Google. So, so really uh it the the role it it or or the software architecture and then the workloads and then optimizing the the software and instruction sets, Um, uh that would be done by the uh the Google, right? Got you. Yeah. Okay, cool. By the way, I have another follow-up here. So, do you have do you have any idea on uh what's the team size of the like the CUDA the CUDA software team uh at Nvidia and and and similar team at Google? what's the team size? Like how how many employees? How many engineers? Yeah, yeah, I my okay, this is I I don't work at Nvidia, but but I would guess there there's at least two to 10,000. Um, um and um and the the the the scalability is is really um determined by so, um there there are internal engineers um at Nvidia that, you know, build low-level instruction set, um optimization, compiler optimization, um, um you know, there's also firmware and also, you know, um they they can um provision the hardware. You know, they can raise the voltages, you know, dynamic voltage scaling, uh based on thermal awareness, based on um and optimization. Um and then uh that that's low-level um, you know, um you know, compute control. Then there's higher level optimization like I mentioned in instruction set parallelism. And on top of that, there there's also application-based um um, you know, for instance getting getting workloads whether you're you're training, you know, seeing uh and pre pre um pre-cleaning. data, um, uh, that is also, um, you know, at a system level, there there are system, uh, software engineers that will optimize um, the interaction of different chips. Like for instance, if you have a GPU, if you have a CPU, and optimizing the, you know, um, the, you know, the controls for that and the pre-filling and um, that leads to at least 2,000, probably, um, uh, and then there's even high-level API, 2, 2,000, maybe 10,000. And on top of that, they have an ecosystem. So, that that means there are, you know, uh, you know, companies, right? That that, you know, have um, you know, uh, more access than, you know, the the general population, um, to um, the um, low-level, um, architecture, uh, hardware architecture, and are able to optimize, you know, registers and um, you know, pipe, you know, the pipeline, um, to optimize the the workloads. Um, and on top and then even on top of that, then you have your end customer. Um, you know, there's OpenAI, there are engineers within OpenAI, there's engineers within, um, you know, Anthropic and um, you know, uh, many other companies that are, um, optimizing, trying to optimize the the hardware and, you know, working with the uh, you know, NVIDIA engineers to uh, optimize the hardware for the software workloads. Um, so, so, um, that that's that's my estimate of internal NVIDIA, uh, engineers. Um, Got it. And I think what what's interesting is that I think, uh, the number of these engineers, software engineers, is not going to grow as much. And is, and I think they're really trying to leverage actually a lot of uh AI to uh optimize um a lot of uh these software functions. Um, Got it, got it. Okay. How about that in in Google, at a Google? Um, so, Google, I mean, in terms of hard, so you mean the um, the mix of hardware versus so, uh, kind of software engineers? Is that what you're? I mean I yeah I mean uh the the software engineer, I mean the team is similar to to team at NVIDIA. was the team size at Google? So, um, okay, there there are a couple of things that are changing in in Google is that they're they're there's an internal effort to try to uh develop their own TPU. Um, and because of that, they're they're they're building up their hardware team, but it's still a fraction of the uh software team and um, uh you know, who who you know, the so even within the software team, there there are a lot of application, kind of application engineers. So they work with, you know, like a Bank of America, they work with uh, you know, some someone who, you know, some legacy or um, um, kind of um, uh, enterprise company to understand what their workloads are and translate that into um how that can be uh the the hardware TPUs can be optimized um, uh, for to, you know, for, you know, more, you know, ops per per watt, um, higher throughput or the other any other, you know, total cost of ownership, um, metrics, um, in the Google Cloud. Um, so, within, I I would say, my my guess is, um, it it may be the same, maybe the same number. Um, in terms, yeah. Okay, no problem. Okay, got you. Yeah, okay. so, um, yeah, so maybe next could you could you share with us um, like, um, what what kind of work uh what what kind of what what kind of workload is are most suitable uh to to to to like different different GPUs or ASICs like like maybe for maybe maybe fast inference is uh maybe uh like the Nvidia Nvidia just acquired acquired Groq, right? So maybe Groq's uh uh LPU and Cerebras uh uh WSE are most suitable for fast inference. And maybe GPU is Nvidia and AMD's GPU are most suitable for uh uh prefill, uh maybe. So, yeah, so so what what's the like what what's the relationship between uh different kinds of GPU and ASICs and and different kinds of uh workloads? Yeah. Yeah, I think, you know, it was really interesting that I that I heard, um, you know, uh even uh I think it was Lucy mentioned about uh, you know, cerebrus. Um, let me let me just make a couple of comments on cerebrus. Um, it had a, you know, it was is of course a startup. Um, you know, that built a, you know, um, you know, a a new type of uh computer, right? Wafer level. that uh really, I mean, I had seen it grow. I had seen, you know, I I met I know knew some of the early engineers, um that went there and some of them, um, during at one point, you know, it it you know, the yield is very low, there are a lot of thermal mechanical issues, you know, things weren't even working. Basically, and then I and I and I saw some people leave the company, not not on their own accord. Then they go to like AMD or they go to Google. Um and uh really um I was really surprised that um it went IPO. And the the main reason that I understand it when IPO was because OpenAI uh really infused um the production of Cerebra. And what what I understand is it's not even even all its all its capacity is under 1 giga W. So it's actually a very small scale um deployment of of and and it's actually surprising that it has any deployment at all of this uh wafer level um, you know, compute. Uh of course, you know, get, you know, somebody can evaluate it in a national lab and um but it's it's really um it's really unbelievable that it's still it's still surviving. It's still survived as an IPO. And and really if you I mean, the there the yield on that, uh, you know, if you look at typical uh defect density is about 0.1, um, and that that means, you know, if it follows, you know, the typical densities of um, you know, uh TSMC, there there are a lot of chips in, you know, like uh or or parts of parts of the uh die that are not functional. So really at at a um when you do when they do the, you know, a bring up and characterization of the wafer, they they must mark the bad units, the bad parts of the the wafer. And have um, you know, dynamic adjust the workloads based on the defect. Because they're not there's going to be every wafer is going to have a defect. But they're going to have to work around the defect. And and that leads to a an adjustable or uh each each wafer and each computer can be can be uh different in uh performance. Um, so, so that really leads to um, you know, even more challenges. in optimizing uh the the computer than Cuda. Um, in that and, you know, say you you write something and, you know, uh a matrix matrix multiplication, um, uh software. Um, and you have to map the instructions to leverage the the hardware. Um, and that just it just takes, you know, maybe uh, you know, you know, at least 30%, maybe, you know, 2X more um, you know, um, computer architecture, um, optimization to to leverage to to maybe even utilize the the compute. So, so in that sense, you know, really it's it's kind of a miracle, um, that is is even alive at all. Um, so that's why, so so I think uh I was really interested, I was really uh surprised to hear hear the interest in Cerebrus. Um, uh, So, um, so I um, can I can you Yeah, go ahead, go ahead. Yeah, yes, yes. and maybe yeah. Um, yeah, so yeah, I I think so from from uh from our perspective, we we also did some research on Cervas, but we think that as Cerebras has uh its own advantages and disadvantages, like they have like, um, on chip RAM, uh memory, uh very very uh very large uh SRAM memory and uh very high uh bandwidth, bandwidth uh within within the wafer. and maybe they have lower uh IO bandwidth between the wafer. So, yeah, and and they have like they have disclosed um some some numbers, some figures that uh their tokens per second per user uh is quite high. So, yeah, so maybe I mean from uh from your perspective maybe uh could you like uh share with us your view on um, uh what I said uh at the beginning of our our discussion. so maybe so what what's the trend uh within GPU and 6 like in the future will there be like uh because Cerebra is uh from our perspective, from our understanding it's it's more focused on uh decode, right? So so in inference there there there could be uh there are there are uh prefill and decode and and Cerebra is more more focused on decode and then uh Nvidia acquired uh Grock. And so so so Nvidia's GPU is maybe in their in in in the combination of uh GPU and Grock's LPU, uh Nvidia's GPU is more focused on prefill uh and maybe attention. And then uh Grock's LPU is more focused on uh decode or or FFM. Yeah, so so uh so so yeah, so what what what I'm interested what I'm interested or maybe maybe you can share with us, uh so how so because you know, another c player, uh Open AI just released their uh Colono, right? So, so but seems that is uh does not does not adopt the uh the PD the prefill decode aggregation approach. Yeah, so so what's the what's the trend? What's so would there be like the more uh like more more niche ASIC in the future, like the the one type one kind of ASIC doing uh doing prefill, one one kind of ASIC doing uh decode or doing or or doing um attention or FFM. Yeah, so what what's Uh uh, what's that trend? Uh, from your perspective? Yeah, I think uh, you know, this is this is, you know, we we had some VPs from uh, meta, you know, kind of, you know, I kind of the the you know, I think they they summarized kind of my perspective. is that, you know, these software models will change every six months. Um, you know, of course there's still going to be general I think there's still going to be general categories of training and then inference. But I mean hardware takes if they're lucky, nine months. So, so really what ends up happening is Nvidia when they sell a system, it's actually it's there's programmability. It's almost like a CPU with with acceleration. right? They have they have the, you know, GPUs and then they cluster them to with MV link and then they have their own like uh again, CPU is not really like the making the performance of the, you know, like traditional Intel CPU. Um, so so really the Nvidia when they sell the hardware, there there's programmability, some programmability. And and really then, you know, and and really actually Jalapeno just to I mean, I I I look at the power numbers, right? It's about it's less than a kilowatt. It's about 700 um 700 watts. in um uh in in power consumption. So it's not it's not actually as powerful um as as the Nvidia uh uh solution even a GPU today. Um, but but it's tailored for specific workloads. So in that sense, the performance can be much better than a general purpose um processor with maybe general sub optimized uh hardware. uh software, sorry. Um, and in that, in that sense, the the ASICs typically outperform general compute by two to 10x. Not because um necessarily the the hardware is much better, but because the the software and then the workload and the instruction set, um, and even maybe the data path is optimized for specific um workloads. So, so that's that's really and when you know what your workloads are going to be, then you can really um, you know, get, you know, significantly uh significant improvements in um in uh performance um with, you know, you know, maybe uh, you know, hardware that may be, you know, 30% you know, uh um of, you know, the uh bleeding edge. Um, um yeah. So, um, so I think uh in the in the future, I think uh, you know, for instance an open AI, um, it it's going to um, you know, or a anthropic, um is going to or even, you know, um, you know, um, yeah. So though those companies um and and even a Google, they're going to have some GPU and actually, I think right now the projection is that um maybe it's uh, you know, 30 30 or 50% of GPU uh, you know, Nvidia GPU. and then there's going to be probably uh, you know, 30% of, you know, ASICs. Um, you know, really uh custom TPUs, you know, Halapeno, um, you know, and it may, you know, and it might come from, you know, Broadcom, it may come from Marvel, it may come from internal development, it may come even from MediaTek. Um, to, um, to uh, you know, uh have these custom um, you know, silicon, custom silicon uh for specific niche workloads that are uh previously predictable. And then there's going to be some general um you know uh compute, you know, C- CPUs. Um, for, you know, um, you know, different workflows. Okay, got it. Got it. Clear. Okay. And from maybe from the perspective of Broadcom, um you you mentioned that um maybe uh in the future and Anthropic will also have their their ASIC like like like like what uh OpenAI has did. Is my is my understanding correct? So, I I hope it's uh public public information, um and I I bet I mean, um, but I I and I, yeah, if it's not public information, I think Hawk would be very proud to to say that I think uh Anthropic uh uh anticipates to book 5 gigawatts of um of uh compute from Broadcom. So it's actually um it's actually a very significant order. Um, you know, you multiply that that's, I mean, that's $50 billion from one customer. Um Yeah, yeah, yeah. So so Mhmm. Yes, I think, yeah, I think Hawk has said that Oh but um yeah, I I think Hawk has shared um similar information in the um in the in in the in the in the conference uh hosted by Goldman Sachs maybe in early September. Oh but I I don't think but I don't think I thought I thought I thought it was purchasing TPUs, I'm not sure. I mean, the 5 gigawatts is it is is the is is the 5 gigawatt Anthropic's ASICs or Google's TPU? Well we Oh, Um, well, so as far as I know, actually uh Google doesn't really make its own hardware. It can brand its hardware and and and you know, the thing with Hock is, he actually doesn't care from a marketing perspective, you know, if Google says, "oh, we make our hardware", but we actually Broadcom uh delivers I think probably 90, nine, at least maybe even 99%. 90 to 99% of the hardware uh uh of Google's um TPUs. Um, maybe it's even 100%. Um, you know, I think now they they're, you know, I think now there's some, you know, press releases that, you know, they're going to Marvel or MediaTek to try or internally try to build some um TPUs or have some skews of TPUs which are um and and now they're splitting into like the um the two types of um uh TPUs. One one is a E and P, um and and there's going to be, you know, just like many many things, there's going to be skews of TP uh TPUs per generations. uh per generation. But um, you know, if Anthropic is, you know, buying Google you know, compute, it really comes from Broadcom. Um, in terms of TPUs. So, Okay. Okay, got it. Okay, um, May okay. Um so All right, so, um, Okay, so could could you please share with us like uh what's like what's the uh, so there are several um custom silicon players in the US right now, right? So, so there were there there are Broadcom, Marvel, uh uh as you said, uh MediaTek, MediaTek is not in the US, but yeah, but MediaTek and also maybe Qualcomm. So, uh yeah, so so what do you think what what are the what do you think is the is the most important important competitive competitive advantages uh among uh this custom silicon players. So like their like their IP, their uh know-how, their uh their uh relationship with with with with with customers. uh what what do you yeah, what what's what's the uh what's their major competitive advantages? Yeah, well, I so, I mean, Qualcomm, you know, is traditionally um um the dominant player in in modems. Um, and uh they actually have a significant portfolio of IP in wireless technology, five 5G, they had a very successful 4G, um and uh 3G, uh and uh and and now they're they're trying to get into server um space. uh but I think, you know, they they they don't build 1 kilowatt plus chips. Um, Marvell is in I I think it's it's also um, it's it's a a second, maybe a a distant second uh in terms of uh customs uh silicon to um, to Broadcom. Um, they they're they they also have had a lot of, you know, changes, you know, they had for instance, I know some, you know, some people that were uh in the ethernet business, uh automotive either net business and they've exited that and then uh they acquired Celesti AI, then they, you know, they a couple of years ago acquired uh Inphi. Um, but but they're um typically associated with medium um performance um compute solutions. um, you know, especially in the networking space. Um, Broadcom has um had 20 plus years in um in in the networking space. um and is a dominant, you know, uh um player in the ethernet ecosystem. Um, and uh, you know, they're they've, you know, adopted um, you know, and are heavily involved in various standard various standards um for the, you know, new scale up infrastructure. And not only do they have, you know, um, you know, uh Ethernet, they have PCIE. Um, they and they have a lot of connectivity. And previously, you know, a, you know, a significant portion, maybe 50% uh of its uh semiconductor revenue was in wireless. Uh and really in you know, um a lot of uh, you know, things for Apple. So, um, really now the shift for Broadcom is to um, you know, you know, follow this AI wave. right? And really uh this these XPUs, which, you know, Hawk, you know, says and I think it's true that, you know, they started this, you know, side project with Google 10 years ago to build, you know, several generations of TPUs that made very little money. And all of a sudden, you know, Google has realized the value of the TPU for um, you know, internal um, uh, you know, uh uh compute utilization and also um a little bit and more and more for the for the cloud and major customers, um that um, it's, you know, this this business has exploded. really, you know, um in in being a, you know, a dominant uh share of the semiconductor industry. And really, I think a lot of the advantages is is its IP. Um, whether it's the um, the controller uh to the HBM or the uh the Serdes. Um, and and its ability to execute uh very quickly. Actually, you know, these making these chips are are extremely hard. Uh mechanically, electrically, and also, you know, um and uh integration and then the the whole uh workflow that Google although it thinks, oh, you know, if we develop, how hard is it to develop a chip? They're finding out how hard it is to develop um, you know, these complicated GPUs. Um, and immediate so then um, I think do you want me to make a few comments about MediaTek or do you want me to um focus a little bit more about something that I've touched uh Okay. uh uh, all right, so, yeah, so maybe yeah, maybe maybe some more more uh maybe more uh information on on on Broadcom. we So, yeah, yeah, so maybe um uh, so in early September, uh in the uh conference in the in the in the analyst call, um Hawk said that uh Broadcom has six major XPU customers. uh, and and he mentioned the four four customers, like Google, Anthropic, Open AI, Meta. Yeah, so I'm interested, what are the other two major customers of Bro XPU business? Yeah, so, um, to be to be uh, you know, whether it's transparent or non-transparent, uh even if I knew, I couldn't tell you. I wouldn't tell you. Yeah, got it. Okay. Because because I um, and um, really, um, the Maybe some like some some uh yeah, maybe like like like like maybe a a limit on the on the like on the range of the of the of the players of the customers. like is it a like a CSP CSP or in in in the US or yeah, any information would be would be good. Um, you know, even if I knew, um, I I I wouldn't tell you and I couldn't tell you. Um, Yeah, so but I mean that's this is actually, you know, and and and that also that comment is that I don't know. Okay. So, yeah, because I and I don't know and if I did, I couldn't tell you. Uh both both things, right? All right. All right, all right, cool. Yeah. Yeah, yeah. Okay. Yeah, and also, um, also Hawk guided that, yeah, as you mentioned uh at the beginning of of of this call, of our call, um the the XPU revenue, the the AI semi revenue will double this fiscal year and then double next fiscal year. So, uh what's your view on that? I mean, do you think it's it's like a aggressive goal or a conservative conservative goal? I I think it's a, um, it's just it's a exact goal. Because let me and I let me tell you why. Because we we order, we we already uh have the allocation by um TSMC. Um, we we so we actually are very uh, uh we can anticipate the um the wafer capacity and then the build, um, you know, very, you know, for for the we can project, you know, this revenue, you know, pretty very pretty accurately for for almost 18 months. And really, well, and I think uh hopefully you you may know, but a project start uh uh Silicon project starts by engaging with TSMC and we declare our end customer or end customers for for our our output. So so they're really uh and then so TSMC also has a, you know, a well you know, a large, you know, a a very good model and, you know, outlook of um, where their their Silicon uh, is going to go. So, and and there are export control limitations. You know, you know, there there are things that, you know, you can't ship here and you can't ship there. Um, so, so in terms of the projections for the next year, I think it's going to be uh, very predictable. And I think uh, Hawk has been very transparent. I think maybe even the whole ecosystem, probably, you know, well TSMC knows, you know, and I think uh, there, it's that um revenue is pretty predictable for the next 18 months. Got you. Got it. Okay. Um, yeah, so let me see. So 18 months from from today will be like, okay. Yeah, so because uh, yeah, Hak guided that, So, Brocom's fiscal year is from from from November to to October. So uh, so it's it's like uh 230 billion dollars revenue from custom Silicon uh from ASMI uh uh in fiscal year 28. Yeah, so so would that be also be very, very predictable because it's like it's like almost 24 months from today. So, so again, I I think I I'll at least stick to my point that, um, you know, uh, I I think I said 18 months, but at least 12 months, and I think 18 months is is predictable. Um, you know, and and I think we extrapolate, we could, you know, it, you know, it six months is not is not a, uh, you know, that far along, so I think that's fair, um, to say that to that trajectory will will continue. Um, yeah. So, Okay, got it, got it. Okay. Okay, no problem. Okay. Um, yeah, maybe, let me see. Um, um, do do um, do you have any, do you do you have any like any any view on uh, what would be the competitive landscape uh, of GPU and ASIC in the future? I mean, uh, yeah, because from yeah, because because because as I as I mentioned earlier, we we are actually we are interested in in in in in in in um, in fast inference, in fast fast fast inference uh ASICs. uh, which include which includes players like uh Cerebras and and Nvidia's Grak and also there are some other smaller players like yeah, some other some like like like uh AMD has has acquired Talos. Yeah, so, uh, yeah, so do you have a like do do do you have any idea on or or or or or um, uh, your your view on on the competitive landscape uh in the future of this like of of of the GPU and ASIC uh uh uh segments. Yeah. well, let me let me uh even I'll I'll try to remind me the the the question, but I want to, you know, like um, you know, just kind of, you know, kind of uh, you know, give my opinion about about, you know, uh the the quarter uh the quarter results and uh Broadcom stock. And I think you might you might find it interesting, maybe it's not. So, so really you think, wow, you you you doubled the revenue, right? And then next quarter, you know, next year, you you know, you say you're going to double the revenue again, our stock should double. right? That's what we all thought, but instead it went down. right? So, so I think uh, you know, my maybe maybe you can give me your opinion, right? But but I think really what uh the reason why it didn't double or didn't even go up is there's no news, um, this quarter. Last quarter they said they their their uh revenue was very strong and it's a, you know, it's gonna double, you know, uh next year, right? And then they said the same thing. So they didn't see any news. So in my opi- you know, that I think that that's my opinion why the um the um the price of uh Gcom didn't go up and it's the same. I think it's the same story with Nvidia. Nvidia is also projecting they they also said we doubled last, you know, from year over year, from last year. And next year we're going to double again. And then, you know, people should, you know, you would think, oh, everybody's jaws are going to drop, you know, the that means the market cap is going to double. But it didn't. Right for Nvidia. So, so you know, and that's, you know, you would think what what is going on? What? because I think, I mean there's two reasons, but you can maybe you can give me your opinion, right? And and I think that really, I mean, Nvidia for for the two things, their it's market cap is very high. Um there's almost no more, you know, there's very little money in the economy for it to double. almost um, and and the second is it's it's I think it's margin, um, is going to go down. It's although it's market cap and uh revenue is going to go up, they're entering into like PC more PC GPUs, which is which is almost negative margin. Uh, if if flat margin. Um and there there just isn't um, you know, their their their yields may not be as high as uh as they like. Um so I I think that's in my opinion. And maybe, you know, you can give me your opinion if you like. why why why uh you know, Nvidia and uh Broadcom even though they say they're doubling in revenue. their their stock and market cap has has uh uh not even it's gone down or saturated. Yeah, yeah, yeah. Did so so that yeah, maybe I think that that would be another interesting topic and I would quickly share share with you my my view my my my opinion on that. I I think overall the the AI I think overall the the uh the beta of the AI hardware is like is like like sort of uh um how do I say that? How how do I say that um um pressured by the uh by the um by the um um uh outlook or the uh projected growth of uh capital expenditures of of uh major CSPs in the United States because um maybe the major CSPs uh capex will like the the growth, the growth will will peak maybe in 2028 and then after that maybe we don't know. I mean it's I I I don't think it's maybe currently it's it's pretty predic predictable. Like even, yeah, it it it's definitely surprising to to see that um I I think from my perspective that Broadcom will still double its uh revenue in 2028. Uh but if we look at it look at the whole market, look at the whole like capital expenditures of the CSP of the of the like the final customers, final payers of the all the AI hardware. Their capex growth will like will peak and and nobody can tell whether what's what what what the growth will be after 2028. So, yeah, so so I think that that that that that put a lot of a lot of pressure on all the all the AI hardware uh stocks. but that's that's another totally another topic uh than than today, yeah. Yeah, so, yeah, and also, maybe one last a few questions uh in for for our call our discussion today. Yeah, so, uh I think, yeah, so I I think I think um Broadcom's uh AI semi revenue doubling uh this uh next year and 20 fiscal year 28, uh is really like the growth is really high because I just want to confirm with you like that because from our perspective, the the growth of uh TSMC is like the CoWoS growth will definitely not double. like we're not like we're not double like in in next year and and then uh 2028. So, is is it is it because uh Brocom has acquired uh has secured more capacity uh at TSMC and also maybe the like the the ASP like the the selling average selling price of uh uh Brocom's products also also also uh is is will be increasing. Yeah, so so so so it's I mean, double the revenue double and double is really like really really high growth. why is that? Oh, well, so there's a couple of things involved in here, right? So the ASP is is going is going up because of the memory. So, XPU is not just, you know, just just the the compute, right? It it has it has memory and and that memory is almost half the price of the ASP. um and, you know, there there's, you know, also a shortage in in HBM. right? And and, you know, they keep wanting to make more Hix, Micron and you know, uh that that actually plays a a a big role in in the ASP. Um and that decreases the the margin of um or or profit. Yeah, or margin um in in margin of of the uh XPU. Um but uh let's see. Yeah, how about the uh like the the the capacity that Brom secured at TSMC? So that's that's that that's why the um the revenue is so predictable. We we already it's it's I think uh I think I'm almost I'm, you know, much more certain than uncertain. uh that, you know, the guidance is going to be very very close for at least for semi. And probably for software. Um I I I don't I I think uh Hawks's guidance is um uh it's actually very scientific um and predictable. and similar in with Nvidia for the next uh 12 months because of the capacity and they they they invest in the, you know, Nvidia, they they build all these racks and they have pre-orders. I think, you know, this is this is something that you can give me some feedback, but here's another theory why, you know, the the uh the stock did not go up as much as um uh, you know, you know, as anticipated with these double. Maybe and then, you know, this is my theory, okay? and maybe um and and the risk, there's a risk um of uh you know, like I mentioned 5 gigawatt of uh you know, compute from anthropic, right? Uh which is not an IPO company, right? I mean it's not a public company, it doesn't make money. Open AI is a big customer. It doesn't make money. Right? In fact, it's like you you say, okay, it doesn't what is this business model? You know, what are the risks of, you know, everybody, you know, abandoning opening AI and go, you know, not buying the, you know, subscriptions and um, you know, or it gets, you know, it eventually goes bankrupt, right? Um, I think I think the big risks is that, you know, maybe 50% of the compute, um, customers are are being financed with, you know, you know, you know, not not real money in in a way. Uh, not uh, not services that generate real real uh, profit. Um, I think that could be a risk. I think that could be this is this is my theory. I don't I maybe. that makes sense. So, so really that's the biggest risk is that now your customer that you were going to deliver 5 gigawatts of compute evaporates, right? So, um, those those are that that could be actually the biggest risk. Um, but if everything, you know, um, you know, follows their trajectory and maybe they're, you know, the government isn't, you know, you know, isn't scared about AI, I think these these trajectories or, you know, without external forces, this is this is a very uh this is the anticipated trajectory, um, in the near future. like 18 months. Got it. Got it. Got it. All right, one one last question from me. Um, yes, so would it be possible to like to give us an estimate of the breakdown of the growth like how much is driven, I mean, from Brocom's growth double and double, how much is driven by volume and how much is driven by uh ASP, ASP increase? Mm. A rough a rough estimate would be good enough. How much is it? I I would say probably ASP maybe um, 50%? Um, and then you know, the remaining is from volume. All right. Um, what what uh, so And let me say that for XPUs only. Like actually a pretty significant a portion of the revenue from Broadcom is from its networking uh business, which has nothing to do with. And that is um, that uh, the ASP will also go up. the um, uh for for its, you know, latest products. Um, and and that's part because of the the wafer costs. Um, but they traditionally, uh, Broadcom tries to generate uh 80% uh margin on um, on its on its products. Um, you know, less so for XPUs because of the HBM involved. The the core uh, you know, XPU uh compute is probably 80%, but then it makes almost it it breaks even or is very, you know, makes no very little um, um, you know, per you know, margin for integrating, you know, memory, you know, whether it's eight eight high, 12 high memory. Um, and that price goes up, it doesn't mean Broadcom is now going to make, you know, you know, 20, you know, if if, you know, memory goes up, you know, 50%, it doesn't mean, you know, we're going to make 60%. Um, it's almost that's a break that there's no uh margin from um, you know, cost of memory. Um, as far as it's. Yeah. All right, got it. Okay. Okay, thank you so much. I think that's all for today. Thank you so much and yeah, maybe in the future we we