华泰证券 Google集群拆解 2025年11月 谢春生 SAC NO. S0570519080006 | SFC NO. BQZ938 目 录 1. Google集群的Scale up: 3D结构 2. Google集群的Scale up光互联：光路交换机 3. TPU集群内，光路交换机和光模块占比 4. Google集群的Scale out 1 Google: Scale up与Scale out Ironwood Superpod & Max-scale Cluster Traffic Type Interface Intra-Superpod Computation ICI(Inter-Chip Interconnect) 1.5x-10x Scaling across all the critical vectors: FLOPs, HBM BW, ICI NIC/s Inter-Superpod Computation BW NIC/s Data Ingestion and Storage <-> Superpod Data Center Scale-Out Network Aux Compute ...

Core Viewpoint - Amazon has announced a long-term strategic partnership with OpenAI, involving a financial commitment of $38 billion, which is expected to enhance AI processing capabilities through AWS infrastructure [3][4]. Group 1: Partnership Details - OpenAI will utilize Amazon EC2 UltraServers, accessing hundreds of thousands of NVIDIA GPUs, with the potential to scale to tens of millions of CPUs [4]. - The partnership's value of $38 billion is projected to grow over the next seven years [4][5]. - OpenAI is expected to start using AWS computing services immediately, with full deployment of computing capabilities anticipated by the end of 2026 [5]. Group 2: Competitive Landscape - OpenAI is focusing on GPU usage for its computational needs, contrasting with Anthropic, which has opted for Amazon's proprietary AI chips [5]. - Recent collaborations between OpenAI and major GPU manufacturers, including NVIDIA and AMD, indicate a trend of significant investments in AI infrastructure [6]. Group 3: Financial Performance - Amazon reported a 12% increase in net sales to $180.2 billion for Q3 2025, with a net profit of $21.2 billion, reflecting a 38.6% year-over-year growth [7]. - AWS has experienced its highest growth rate since 2022, driven by strong demand for AI and core infrastructure [7]. Group 4: Market Sentiment - There is ongoing debate in the market regarding the potential for an AI bubble, with experts suggesting that the return on investment from massive AI expenditures may not be clear for at least a year [7].

Core Insights - The global semiconductor equipment sales are projected to reach a record high of $125.5 billion by 2025, driven by capacity expansion for advanced logic and memory to support AI applications and technological iterations across various market segments [1][2] Group 1: Semiconductor Equipment Market - The demand for semiconductor equipment continues to grow, with advanced logic and memory technology iterations leading to upgraded equipment requirements [2] - China is expected to maintain its lead in global 300mm equipment spending, with total investments projected to reach $94 billion from 2026 to 2028 [1][2] - The evolution towards more advanced processes in logic will further drive demand for etching, thin film deposition, bonding, and coating equipment [2] Group 2: AI and Semiconductor Demand - AI demand is significantly boosting the growth of the logic and memory markets, with data center investments expected to rise from $17 billion in 2020 to $45.3 billion by 2028 [3] - OpenAI plans to deploy a total of 6 GW of AMD GPU computing power, estimating that $50 billion will be needed to bring 1 GW of AI computing online, with two-thirds allocated for chips and supporting data center infrastructure [3] - The overall revenue for AI semiconductors is projected to grow from $8.2 billion in 2020 to $41.3 billion by 2030, with a compound annual growth rate of 24.4% [3] Group 3: Supply Chain and Domestic Production - There are ongoing supply chain challenges, with limitations on acquiring key semiconductor equipment, particularly for advanced chip manufacturing [4] - The U.S. Congress is proposing broader bans on the sale of chip manufacturing tools to China, which could impact the availability of advanced technology [4] - Domestic demand for AI chips is expected to increase as China progresses with its computing infrastructure, following NVIDIA's complete exit from the Chinese market [5]

Group 1: Core Views - The "15th Five-Year Plan" (2026-2030) is crucial for establishing a foundation for achieving the 2035 modernization goals, focusing on high-quality development driven by new productive forces [2][3] - The core development line emphasizes technological innovation, with key areas including artificial intelligence, quantum technology, and low-altitude economy becoming strategic focuses in international competition [2][3] Group 2: Strategic Goals - Economic growth should maintain an average of approximately 4.73% annually to achieve the long-term goal of doubling GDP or per capita income by 2035 [3] - Key quality indicators include a continuous increase in labor productivity, R&D investment intensity reaching 3%, and a rise in the share of non-fossil energy consumption to around 25% [3] Group 3: Policy Recommendations - A systematic policy framework is proposed, including building a modern industrial system centered on new productive forces, deepening reform and opening up, and promoting green low-carbon transformation [4] - Emphasis on enhancing governance efficiency through reforms and digital governance to modernize national governance [4] Group 4: AI Industry Insights - OpenAI has signed a 10 GW computing power order with Broadcom, planning to launch custom data center chips by 2026, indicating a strategic move towards building foundational hardware capabilities [7] - OpenAI is exploring new business models, including partnerships with Walmart for direct purchasing through ChatGPT and plans to launch consumer hardware products, aiming for diversified monetization [8] - TSMC has observed stronger AI demand compared to three months ago, with optimistic revenue guidance for Q4, reflecting confidence in the AI growth trajectory [9]

Group 1 - AMD's stock price rose over 3% in pre-market trading following the announcement of increased adoption of its products by cloud computing companies as alternatives to NVIDIA [2] - OCI's executives expressed confidence in customer adoption of AMD chips, particularly in the AI inference sector, highlighting a partnership aimed at providing cost-effective and scalable cloud infrastructure [2][3] - Oracle Cloud Infrastructure (OCI) plans to deploy 50,000 AMD GPUs starting in Q3 2026, with further expansions planned for 2027 and beyond [3] Group 2 - OpenAI has reached a significant agreement with AMD to deploy up to 6GW of AMD GPUs over the coming years, with the first 1GW set to be operational next year [5] - OpenAI's shift towards AMD comes as it seeks to diversify its sources of computing power, moving away from its historical reliance on NVIDIA, which holds over 90% of the data center GPU market [5] - Oracle's executives have praised AMD's performance, comparing it favorably to NVIDIA, and emphasized the importance of leveraging data resources and enterprise capabilities to add value in the AI landscape [5]

Group 1: Pharmaceutical and Biotechnology Industry - The pharmaceutical and biotechnology sector experienced an overall decline of 1.20% in the week from October 6 to October 10, ranking 25th among 31 industries, underperforming the CSI 300 index by 0.69 percentage points [5] - Year-to-date, the sector has risen by 21.87%, ranking 12th among 31 industries and outperforming the CSI 300 index by 4.54 percentage points [5] - The current PE valuation for the pharmaceutical and biotechnology sector is 30.87 times, at the historical median level, with a 129% premium over the CSI 300 index [5] - Notable sub-sectors that increased include Traditional Chinese Medicine II and Pharmaceutical Commerce, with gains of 1.51% and 0.64% respectively [5] - A total of 244 stocks (50.94%) rose while 215 stocks (44.89%) fell during the week, with the top five gainers being Zhendemedical (21.00%), Wanbangde (12.01%), Changshan Pharmaceutical (8.69%), Zhenbaodao (8.56%), and Pumen Technology (6.66%) [5] Group 2: Regulatory Developments - On October 10, the State Council released regulations for the management of clinical research and clinical application of biomedical new technologies, effective from May 1, 2026, establishing a clear framework for the regulated development of these technologies [6] - The regulations emphasize the establishment of a comprehensive safety supervision system, protection of ethics and rights of subjects, and strict legal responsibilities [6] - On October 9, the National Medical Products Administration announced a draft for public consultation regarding the acceleration of post-marketing research and evaluation of traditional Chinese medicine injections, mandating companies to conduct studies based on fixed production processes [6] Group 3: Mergers and Acquisitions - Novo Nordisk announced an agreement to acquire Akero Therapeutics for $4.7 billion, gaining access to Akero's FGF21 analog Efruxifermin, a potential best-in-class therapy for metabolic dysfunction-related fatty liver disease (MASH) [7] - The acquisition could reach a total value of $5.2 billion upon FDA approval of Efruxifermin for treating MASH-related decompensated cirrhosis [7] Group 4: Electronic Industry Insights - The electronic sector is facing increased tensions due to U.S.-China trade frictions, with the U.S. imposing an additional 100% tariff on Chinese imports, exacerbating global semiconductor supply chain issues [11][12] - OpenAI has established a partnership with AMD to deploy a total of 6GW of AMD GPU computing power, marking a shift towards a "computing alliance" era in AI competition [11][13] - The electronic industry is currently in a phase of moderate recovery, with recommendations to focus on AI server supply chains, AIOT, device materials, and domestic automotive electronics [11][15] Group 5: Market Performance - The Shanghai Composite Index closed at 3889 points, down 0.19%, with major indices showing mixed performance [17] - The electronic sector underperformed the broader market, with the electronic index declining by 2.63% [14] - The market saw a significant outflow of large funds, with net outflows exceeding 128 billion yuan [17]

Investment Rating - Industry Rating: Recommended [5] Core Insights - The humanoid robot industry is expected to accelerate production due to breakthroughs in AI technology and domestic companies' efforts in component localization, creating significant market opportunities [1][12][13] - The domestic new energy vehicle market is experiencing a strong upward trend in deliveries and sales, with major companies like NIO and Xiaopeng achieving record monthly performances, indicating a robust demand for power batteries [2][16][18] - The European energy storage market is poised for explosive growth driven by policy and economic factors, with significant orders being captured by leading domestic companies [3][27] - The green hydrogen and methanol production projects by Goldwind Technology are gaining traction, with successful verification of gasification processes, indicating a promising market for green fuels [4][28][31] Humanoid Robots - The signing of a cooperation framework agreement between SIRIS and Volcano Engine marks a significant step in the development of humanoid robots, with strong demand for domestic component substitution [1][12] - The industry is witnessing a surge in participation from major tech companies, which is expected to enhance the overall market dynamics and opportunities for component suppliers [13][15] New Energy Vehicles - The new energy vehicle sector is in a rapid growth phase, with continuous introduction of high-quality models and advancements in battery technology, which are expected to enhance performance and reduce costs [2][16][18] - The demand for lithium batteries and upstream materials is expected to remain strong, driven by the upcoming traditional sales peak in Q4 and the introduction of competitive overseas models [2][16][19] New Energy - The European energy storage market is projected to see significant growth, with new installations expected to reach 8.8GWh and 16.3GWh in 2024 and 2025, respectively, indicating a strong demand for storage solutions [3][27] - Domestic leading companies are becoming key players in the European market, benefiting from their technological advantages and local presence [3][27] Power Equipment & AIDC - The AIDC industry is experiencing a high degree of synergy, with domestic liquid cooling equipment manufacturers expected to benefit from the industry's growth [7]

Group 1: OpenAI Developments - OpenAI's video AI application Sora achieved over 1 million downloads within 5 days of launch, surpassing the download speed of ChatGPT [1] - Sora is currently limited to iOS devices and operates on an invitation-only basis, which contrasts with ChatGPT's initial public release [1] - OpenAI is enhancing ChatGPT into a "super app" that allows users to directly access third-party applications within conversations [6] Group 2: Strategic Partnerships and Investments - AMD has issued a warrant to OpenAI allowing the purchase of up to 160 million shares at $0.01 each, as part of a strategic partnership to deploy 6 gigawatts of AMD GPU capacity [2] - NVIDIA is also collaborating with OpenAI, with plans to invest up to $100 billion, contingent on the construction of data centers [2] - NVIDIA confirmed its investment in Elon Musk's xAI, contributing $2 billion in a recent funding round, as part of its broader strategy to support AI applications [4] Group 3: Industry Trends and Performance - Hon Hai Precision Industry (Foxconn) reported a 38.01% quarter-over-quarter revenue increase, driven by a surge in AI server shipments [3] - SoftBank is acquiring ABB's robotics unit for $5.375 billion, indicating a strategic shift towards physical AI and automation technologies [7] - Tencent's Hunyuan-Vision-1.5-Thinking model ranked third globally in the latest visual model rankings, highlighting the competitive landscape in AI model development [9] Group 4: Talent Movement in AI Sector - Yao Shunyu, a notable physicist, transitioned from Anthropic to Google DeepMind, citing a desire for more opportunities in the rapidly evolving AI field [10][11]

Investment Rating - The industry investment rating is "Overweight (Maintain)" [1] Core Insights - The report emphasizes the rapid commercialization of AI software and hardware, highlighting significant collaborations between OpenAI and major tech companies like AMD and Samsung, which are expected to drive growth in the AI and storage sectors [3][5][6] Summary by Sections Background - During the National Day holiday, there was a surge in AI-related news, with OpenAI launching Sora 2 and forming a strategic partnership with AMD for 6 GW of computing power, indicating a strong acceleration in AI software commercialization [3][4] Software Developments - OpenAI's Sora 2 was launched, achieving breakthroughs in audio-visual synchronization and introducing a refined IP control scheme, marking a significant step towards commercial viability [4][14] - The model supports diverse video styles and incorporates a multi-layered governance framework to ensure compliance and safety [4][14] Hardware Collaborations - OpenAI and AMD's partnership involves a "compute + equity" model, where AMD provides 6 GW of GPU resources in exchange for equity, potentially giving OpenAI about 10% of AMD's shares if fully exercised [5][16] - This innovative financial model alleviates cash flow pressures on AI companies while tightly coupling supply chain interests [5][16] Storage Sector Growth - OpenAI's collaboration with Samsung and SK Hynix aims to establish the "Stargate" AI infrastructure project, planning to procure approximately 900,000 semiconductor wafers by 2029 and build data centers in Korea [6][17] - This partnership is expected to unlock new growth opportunities in the high-end storage chip market, given the dominance of these companies in the DRAM and HBM markets [6][17] Investment Recommendations - The report suggests focusing on companies in the computing chip sector such as Cambrian and Haiguang Information, and in the storage chip sector, companies like Demingli and Zhaoyi Innovation [7][17]

Core Insights - AMD has signed a 6GW computing power chip cooperation agreement with OpenAI, which is expected to accelerate the ROCm ecosystem development [1] - The Haiguang DTK software stack is fully compatible with the ROCm ecosystem, potentially reducing the adaptation costs for large models in China [2] - Industrial and commercial banks, along with China Unicom, are procuring Haiguang chips worth several billion yuan, indicating a rapid penetration of Haiguang chips in the domestic market [3] Group 1 - AMD and OpenAI's agreement includes the deployment of a total of 6GW of AMD GPU computing power, starting with 1GW based on the MI450 series GPU in the second half of 2026 [1] - The collaboration is seen as a response to the significant gap in AI computing power supply and aims to enhance the synergy between AMD and OpenAI in both hardware and software [1] - ROCm 7 is expected to be available by Q3 2025, offering over 3.5 times the inference capability and three times the training capability compared to ROCm 6 [1] Group 2 - The Haiguang DCU features a complete software stack, including DTK, development toolchain, and model repository, which supports mainstream deep learning frameworks like TensorFlow and Pytorch [2] - This compatibility is anticipated to effectively lower the costs associated with adapting large models domestically [2] Group 3 - Industrial and Commercial Bank of China has announced a procurement project for Haiguang chip servers, with a potential total procurement amount reaching 3 billion yuan [3] - China Unicom's procurement project for Haiguang servers has a budget of 2.782 billion yuan, indicating strong demand for domestic chip solutions [3] - The financial and telecommunications sectors are expected to lead the way in achieving comprehensive domestic production, thereby accelerating the penetration of Haiguang chips [3]