华泰证券指出，不同于OpenAI严重依赖外部算力（英伟达）和云设施（微软），谷歌正在构建从芯片 （TPU v7p）到模型（Gemini 3.0）再到应用（搜索+Waymo）的完全自给自足的生态闭环。看好谷歌自 给自足"全栈式"AI 生态和能力，抢回主导权正当时。报告指出，这一闭环正在转化为实打实的财务回 报：TPU部署大幅降低推理成本，搜索市场份额企稳回升至90%以上，充沛的广告现金流为高强度的资 本开支（Capex）提供了充足弹药。长江证券也指出，同等算力规模的情况下，ASIC的光模块用量远超 GPU。若只对比谷歌和英伟达算力卡的数字：TPU v7的FP8精度算力估算为4614 TFLOPS，vs Rubin （2die）约16667 TFLOPS，推测TPU v7的1.6T模块配比约1:4.5，vs Rubin（2die）的1:5，换算一下，在 纸面算力相等的情况下TPU v7光模块用量是Rubin（2die）的3.3倍。换言之，若英伟达份额有所下降， 则ASIC或将取得更多份额，光模块板块增速更快，占整体Capex比例或将进一步提升。 数据显示，截至2025年10月31日，上证科创板100指数(0006 ...

华泰证券认为，谷歌凭借TPU v7p自研芯片至Gemini模型的全栈AI生态，正强势反击。TPU推动云业务增速达34%，仅次于Azure；Gemini月活达6.5 亿用户，AI Overviews服务超20亿用户。搜索市占率重回90%以上，充沛广告现金流支撑高强度资本开支。 硬科技护城河：TPU v7p对标B300，云业务增速超越AWS 市场长期低估了谷歌的"全栈式"AI反击能力。 近日，华泰证券在最新研报中称，不同于OpenAI严重依赖外部算力（英伟达）和云设施（微软）， 谷歌正在构建从芯片（TPU v7p）到模型（Gemini 3.0）再到应用（搜索+Waymo）的完全自给自足的生态闭环。 看好谷歌 自给自足"全栈式"AI 生态和能力，抢回主导权正当时 。 报告指出，这一闭环正在转化为实打实的财务回报： TPU部署大幅降低推理成本，搜索市场份额企稳回升至90%以上，充沛的广告现金流为高强度的 资本开支（Capex）提供了充足弹药。 华泰证券认为，谷歌的云端大规模自研 TPU 及配套软件生态，带动云业务增速及份额提升；广告业务在 Gemini 赋能下具备变现弹性，充沛现金流反 哺 AI 投入与应用落地。 反 ...

华泰证券认为，与依赖外部算力资源的竞争对手不同，谷歌早在2016年就启动TPU部署，2017年已具备训练和推理能力。目前正推动TPU部署至第三方云服 务商，如Fluidstack，并采用按用量分成模式，有望开辟新的增长空间。 市场长期低估了谷歌的"全栈式"AI反击能力。 近日，华泰证券在最新研报中称，不同于OpenAI严重依赖外部算力（英伟达）和云设施（微软），谷歌正在构建从芯片（TPU v7p）到模型（Gemini 3.0） 再到应用（搜索+Waymo）的完全自给自足的生态闭环。看好谷歌自给自足"全栈式"AI 生态和能力，抢回主导权正当时。 报告指出，这一闭环正在转化为实打实的财务回报：TPU部署大幅降低推理成本，搜索市场份额企稳回升至90%以上，充沛的广告现金流为高强度的资本开 支（Capex）提供了充足弹药。 华泰证券认为，谷歌的云端大规模自研 TPU 及配套软件生态，带动云业务增速及份额提升；广告业务在 Gemini 赋能下具备变现弹性，充沛现金流反哺 AI 投入与应用落地。反观OpenAI仅具备大模型研发能力。当前用户规模虽多，但C端付费意愿较低，而B端商业化路径也尚未跑通，先发优势能否继续仍有待 ...

Core Viewpoint - The market has long underestimated Google's "full-stack" AI capabilities, which are self-sufficient from chip development (TPU v7p) to model creation (Gemini 3.0) and application deployment (Search + Waymo) [1] Group 1: AI Ecosystem and Financial Performance - Google's self-sufficient "full-stack" AI ecosystem is translating into tangible financial returns, with TPU deployment significantly reducing inference costs and stabilizing search market share above 90% [1][6] - The cloud business is experiencing growth, with Q3 cloud revenue reaching $15.2 billion, a 34% year-over-year increase, and market share rising from 18.6% to 19.3% [4] - The advertising business, empowered by Gemini, shows strong monetization elasticity, providing ample cash flow to support ongoing AI investments [7][10] Group 2: Competitive Positioning - The TPU v7p chip, with an FP8 computing power of 4.5 PF, directly competes with Nvidia's B300 chip, showcasing Google's dominance in computing power [3] - Unlike competitors that rely on external computing resources, Google has been deploying TPU since 2016, now expanding to third-party cloud service providers [3] - Google's AI ecosystem, built on TensorFlow and OpenXLA, has the potential to compete with Nvidia's CUDA [3] Group 3: User Engagement and Product Integration - Gemini 3.0 has improved capabilities, with monthly active users reaching 650 million, and is expected to leverage Google's extensive user traffic through deeper integration with search [6] - The Chrome browser is accelerating the integration of Gemini features, enhancing user experience with personalized search results and content generation [6] Group 4: Future Projections - Based on the comprehensive ecosystem development, revenue forecasts for Google have been raised, with expected revenue of $405.17 billion in 2025 and net profit of $131.51 billion [10] - The target price for Google has been adjusted to $380, indicating over an 18% upside potential based on a 30x PE ratio for 2026 [1][10]

在全球云端AI服务器投资浪潮下，台积电作为主要代工厂商正迎来前所未有的增长机遇。 瑞银最新研报显示，每1GW服务器项目将为台积电带来10-20亿美元的收入机会，相当于其2025年预期销售额的1.0- 1.5%。随着OpenAI和多家超大规模云服务商宣布数十GW级别的服务器建设计划，台积电的收入增长潜力将远超市场 预期。 不同AI平台的产能需求差异显著 研报分析显示，台积电在英伟达新一代AI GPU平台中的实际总收入金额将逐步增长。每1GW服务器建设中，台积电 从Blackwell Ultra/Rubin平台获得的收入约为11亿美元，而到Rubin Ultra/Feynman平台时将增至14-19亿美元。 博 通ASIC方案效率优势明显，ASIC芯片凭借针对特定工作负载的更高效率，可能产生比GPU更多的芯片单元需求。博 通为谷歌设计的TPU v7p，每1GW需要约4.9千片/月的N3产能,远高于英伟达的2-4千片/月。 从收入占机架价值的比例看，ASIC方案为台积电带来的实际总收入金额更高，达到10-11%，而GPU方案为4-6%。具 体到谷歌TPU v7p项目，台积电每1GW的收入机会可达18.95亿美元。 ...

Core Insights - TSMC's 3nm process has officially entered a golden mass production phase, with third-quarter revenue contribution rising to 23%, surpassing the 5nm process and becoming a key driver for overall operations [2] - The demand for AI and cloud applications is driving TSMC's 3nm production lines to operate at full capacity, with utilization rates at the Tainan Fab18 facility nearing maximum [2] - NVIDIA is a major contributor, increasing its monthly wafer orders to 35,000, which is straining the advanced process capacity [2] Group 1 - TSMC's monthly 3nm production capacity has rapidly increased from 100,000 wafers at the end of last year to 100,000-110,000 wafers, with projections to reach 160,000 wafers by 2025, representing a nearly 50% increase [2] - Major cloud service providers (CSPs) are competing for 3nm capacity, with AWS and Google planning to utilize TSMC's 3nm process for their AI chips [2] - The semiconductor industry anticipates challenges in 3nm wafer supply next year, as CSPs like Google seek to secure more wafer allocations [3] Group 2 - TSMC's 3nm process is expected to account for over 30% of its revenue next year, driven primarily by AI and high-performance computing (HPC) [3] - TSMC plans to increase prices for advanced process technology by 3-5% over the next four years, reflecting strong demand for AI chips and indicating a seller's market for the most advanced wafer foundry services [3] - The introduction of improved versions of the 3nm process, such as N3E and N3P, aims to optimize performance, power consumption, and yield [3]

Core Viewpoint - The rapid advancement of AI large models is driving the transition of AI from a supportive tool to a core productivity force, with computing power chips being crucial for training and inference of these models [2]. Group 1: Computing Power Indicators - **Process Technology**: Major overseas companies are utilizing advanced process technologies, with Nvidia's latest Blackwell series using TSMC's 4NP (4nm) technology, while AMD and Intel are at 5nm. Domestic manufacturers are transitioning from TSMC's 7nm to SMIC's 7nm [3][4]. - **Transistor Count and Density**: Nvidia's B200 chip, using Chiplet technology, has a transistor density of 130 million/mm², while Google's TPU Ironwood (TPU v7p) boasts a density of 308 million/mm², significantly higher than competitors [6][7]. - **Performance Metrics**: Nvidia's GB200 achieves FP16 computing power of 5000 TFLOPS, while Google's TPU Ironwood reaches 2307 TFLOPS, showcasing a significant performance gap [10][11]. Group 2: Memory Indicators - **Memory Bandwidth and Capacity**: Most overseas manufacturers are using HBM3e memory, with Nvidia's GB200 achieving a bandwidth of 16TB/s and a capacity of 384GB, significantly surpassing domestic chips that primarily use HBM2e [18][19]. - **Arithmetic Intensity**: Nvidia's H100 has a high arithmetic intensity close to 600 FLOPS/Byte, indicating efficient memory bandwidth usage, while domestic chips exhibit lower arithmetic intensity due to their lower performance levels [20][21]. Group 3: Interconnect Bandwidth - **Interconnect Capabilities**: Overseas companies have developed proprietary protocols with interconnect bandwidth generally exceeding 500GB/s, with Nvidia's NVLink5 reaching 1800GB/s. In contrast, domestic chips typically have bandwidth below 400GB/s, with Huawei's 910C achieving 700GB/s [26][27].