2025年08月24日 证券研究报告 | 数据中心互联技术专题四： CSP云厂AI军备竞赛加速，智算中心架构快速发展 国信通信·行业专题报告 行业研究 · 行业专题 通信 投资评级：优于大市（维持评级） 证券分析师：袁文翀 联系人：赵屿 021-60375411 021-61761068 yuanwenchong@guosen.com.cn zhaoyu6@guosen.com.cn S0980523110003 请务必阅读正文之后的免责声明及其项下所有内容 投资摘要 CSP互联网云厂AI军备竞赛进入2.0时代，智算中心互联技术发展快速迭代。自2023年，ChatGPT3.5 点燃 "大模型革命"起，AI发展万众 瞩目，各大科技公司纷纷投入大模型研发并加大智算中心建设。根据CSP厂商的Capex指引，预计2025年，海外亚马逊、谷歌 、微软 、 Meta四家厂商合计Capex增至3610亿美元，同比增幅超58%；国内字节、腾讯、阿里Capex有望超过3600亿元。本轮AI浪潮前期，英伟达作为 AI芯片领军企业，其AI芯片供不应求；随着CSP云厂持续加大智算中心投入，具备更高性价比的自研ASIC算力芯片成为AI ...

Core Viewpoint - Google has raised its annual capital expenditure plan to approximately $85 billion, driven by strong demand for its cloud computing services, and anticipates further increases in investment next year [2][5]. Group 1: Financial Performance - For the second quarter ending June 30, Google's total revenue reached $96.43 billion, exceeding analysts' average expectation of $94 billion [2]. - Earnings per share were $2.31, surpassing the expected $2.18 [2]. - Google's cloud business sales surged nearly 32% year-over-year, significantly exceeding the anticipated growth of 26.5% [2]. Group 2: Cloud Business Growth - The rise of AI technology has led to a surge in demand for cloud computing services [7]. - Although Google Cloud still lags behind Amazon AWS and Microsoft Azure in total sales, it is catching up by promoting AI solutions, including its self-developed TPU chips [7]. - The number of customers for Google Cloud increased by 28% quarter-over-quarter [7]. Group 3: Capital Expenditure Insights - The significant increase in capital expenditure surprised the market, with a $10 billion rise that offset the positive financial performance [5]. - CFO Anat Ashkenazi indicated that capital expenditure would further expand in 2026 due to market demand and growth opportunities [5]. Group 4: Competitive Landscape - Google faces competitive pressure from OpenAI, prompting the need for substantial investments in AI infrastructure and applications [5][8]. - OpenAI's decision to include Google Cloud in its list of cloud computing providers is seen as a significant victory for Google [8]. Group 5: Search Business and AI Integration - New AI features are enhancing user engagement in Google's search business, helping to counter competition from ChatGPT [9]. - The AI model Gemini has surpassed 450 million monthly users, indicating rapid integration into Google's product line [9]. - Advertising revenue for the second quarter grew by 10.4% to $71.34 billion, exceeding expectations [9]. Group 6: Other Business Segments - YouTube's advertising revenue for the quarter was $9.8 billion, surpassing the expected $9.56 billion [10]. - The "Other Bets" segment, which includes the autonomous driving project Waymo, generated $373 million, falling short of the expected $429.1 million [10]. - Waymo is expanding its service area but has not yet met investor expectations for commercialization [10].

Core Viewpoint - The 2025 World Artificial Intelligence Conference (WAIC) will take place from July 26-29 in Shanghai, showcasing domestic self-developed TPU chips and AI integrated hardware and software solutions by Zhonghao Xinying [1] Company Highlights - Zhonghao Xinying is the only domestic company that has mastered the research and development technology of high-performance TPU architecture AI dedicated computing chips and achieved mass production [1] - The company will present a full-stack TPU solution at WAIC 2025 through a specially designed main exhibition area and the Future Tech exhibition area [1] - Zhonghao Xinying will focus on the technical insights of domestically developed TPU chips and the practical implementation of intelligent computing solutions through dual presentations at the conference [1] Industry Collaboration - Zhonghao Xinying will collaborate with partners such as Tianjin Mobile to hold a lighting ceremony for the TPU Intelligent Computing Center, showcasing the technological strength of innovative autonomous computing infrastructure [1]

Group 1: AI ASIC Market Insights - The AI ASIC market is experiencing rapid growth, with significant price and power consumption advantages over GPUs. The average price of GPUs is projected to be $8001 in 2024, while AI ASICs are expected to average $5236, highlighting a clear price advantage for AI ASICs [1][14][17] - The market size for AI ASICs is expected to grow from $14.8 billion in 2024 to $83.8 billion by 2030, with a CAGR of 33.5% from 2024 to 2030. In comparison, the GPU market is projected to grow from $70.1 billion to $326.3 billion during the same period, with a CAGR of 29.2% [1][20][18] - AI ASICs are anticipated to capture a larger market share in the training and inference sectors, with their growth rates outpacing those of GPUs [1][20] Group 2: Google TPU Development Trends - The development of Google's TPU has revealed three major trends: increasing specialization, enhanced computational power, and improved energy efficiency. The TPU v5 series includes TPU v5e for cost-effective training and inference, and TPU v5p focused on large model training [2][26][81] - The TPU architecture has evolved to support more complex tasks, with TPU v4 and v5 series demonstrating significant improvements in performance and energy efficiency, with TPU v5e achieving a 2.5 times increase in cost-effectiveness for inference tasks [2][57][76] - The latest TPU v7 (Ironwood) has shown a peak performance increase of 10 times compared to TPU v5p, with significant enhancements in HBM capacity and inter-chip bandwidth [2][76][78] Group 3: Stablecoin Regulatory Developments - The introduction of the Stablecoin Ordinance in Hong Kong aims to enhance transparency and reduce redemption risks in the stablecoin industry, providing a clear regulatory framework for compliant institutions [3][84] - Stablecoins are expected to improve cross-border payment efficiency, offering advantages over traditional systems by bypassing the inefficiencies of SWIFT [3][84] - The regulatory framework is anticipated to activate digital financial innovation, paving the way for the integration of stablecoins in various financial applications, including RWA (Real World Assets) [3][84]

Group 1 - Tesla CEO Elon Musk predicts that AI intelligence will surpass individual human intelligence in less than two years and exceed the total human intelligence in about five years [2] - Musk emphasizes the current AI capabilities have surpassed most humans but not the top individuals or specialized teams, indicating a trajectory of "accelerating returns" driven by improvements in computing power, algorithms, and data [2] - The AI industry is rapidly transforming the world, with breakthroughs in large models enabling machines to possess language, vision, and reasoning capabilities, leading to trillion-dollar applications in areas like autonomous driving and smart manufacturing [3] Group 2 - The US and China are leading the global AI race, holding over 80% of AI patents and 90% of unicorn companies, with the US excelling in foundational research and hardware ecosystems, while China focuses on application-driven innovation [3] - As of Q1 2024, China's AI core industry scale is nearing 600 billion RMB, with a total of 478 large AI models released, ranking second globally after the US [6] - Experts suggest that AI technologies, particularly large models, are crucial for driving high-quality economic development in China, advocating for increased investment in foundational research to create a virtuous cycle between AI research and application [6]

Core Viewpoint - The AI boom continues to reshape the global economic landscape, with Nvidia's market capitalization reaching $4 trillion, marking a historic milestone in the tech industry since the AI wave began in late 2022 with the launch of ChatGPT [1][3]. Company Performance - Nvidia's stock price rose by 2.5% to $163.9 per share at the opening on July 9, 2023, closing at $162.9, with a total market cap of $3.97 trillion, nearing the $4 trillion mark [3]. - Nvidia's revenue for Q1 of the 2026 fiscal year reached $44.06 billion, a 69% year-over-year increase, with net profit of $18.78 billion, up 26% [4]. - Nvidia's stock has seen a cumulative increase of over 1000% since the beginning of 2023, with a 20% rise since April 2025 [3][4]. Market Position - Nvidia remains the leading supplier of AI chips for AI servers, with its data center business generating $39.1 billion in revenue, accounting for 89% of total revenue, and growing 73% year-over-year [6]. - The global data center market was valued at $250 billion in the previous year, growing at an annual rate of 20% to 25%, supporting Nvidia's valuation [5]. Future Outlook - Analysts predict Nvidia's market cap could reach $6 trillion, with a target price increase from $175 to $250 per share, driven by the anticipated "golden wave" of generative AI adoption [4]. - Nvidia is entering a "decade-long AI infrastructure buildout period," with significant growth opportunities in AI and robotics, representing a multi-trillion dollar market [6]. Challenges - Nvidia faces competition from self-developed chips by major tech companies like Amazon, Microsoft, and Meta Platforms, which could reduce reliance on Nvidia's products [8]. - The company is also challenged by the rise of low-cost AI models and restrictions on sales to China, which previously accounted for 25% of its global sales [7][8]. - Concerns about quantum computing advancements pose a potential risk, with some experts suggesting breakthroughs could occur within 5 to 7 years, potentially surpassing Nvidia's current technology [8].

Market Overview - The US stock market experienced a collective rise on July 3, with the S&P 500 index up by 0.83%, the Nasdaq up by 1.02%, and the Dow Jones up by 0.77%, reaching historical closing highs for both the S&P 500 and Nasdaq [1][3] - For the week, the S&P 500 index increased by 1.7%, the Nasdaq by 1.6%, and the Dow by 2.3% [1] Employment Data - The US Labor Department reported an increase of 147,000 non-farm jobs in June, surpassing the expected 110,000, with the unemployment rate dropping to 4.1% [3] - Average hourly earnings rose by 0.2% month-over-month and 3.7% year-over-year, indicating reduced inflationary pressures [3] - Government employment saw a significant increase of 73,000 jobs, primarily in state and local sectors, particularly in education [3] - The initial jobless claims decreased by 4,000 to 233,000, marking a six-week low, which was below economists' expectations of 240,000 [4] Federal Reserve Outlook - Following the employment data release, the probability of a rate cut by the Federal Reserve in July dropped from 23.8% to 4.7%, with expectations shifting towards a potential rate cut in September instead [5] Technology Sector Performance - Major tech stocks saw gains, with Nvidia rising by 1.33% to reach a historical high, and its market capitalization briefly exceeding $3.9 trillion [7][8] - OpenAI's decision to continue using Nvidia GPUs and AMD AI accelerators for model training and inference supports Nvidia's market position [8] Chinese Stocks - Chinese stocks showed mixed performance, with the Nasdaq Golden Dragon China Index up by 0.4% [9] - Brain Rejuvenation Technology surged nearly 122%, with intraday gains exceeding 170%, marking a year-to-date increase of nearly 18,000% [9][10]

Core Insights - OpenAI's shift to Google TPU chips marks a significant turning point in AI infrastructure, providing Google with a strong endorsement of its capabilities and potentially accelerating growth in its cloud business [1][2] - The collaboration allows OpenAI to reduce reliance on Microsoft's data centers while challenging NVIDIA's dominance in the GPU market [2][3] - Morgan Stanley projects substantial spending on NVIDIA GPUs, with estimates of $243 billion in 2027 and $258 billion in 2028, compared to approximately $21 billion and $24 billion for TPU [2] Group 1 - OpenAI's large-scale adoption of Google TPU chips represents its first significant move away from NVIDIA, indicating a strategic shift in its computing resources [2] - The partnership is expected to drive Google Cloud revenue growth, which has not yet been reflected in GOOGL's stock price [2][3] - The increasing familiarity of developers with TPU technology may lead to further adoption by companies outside of Google, providing additional growth opportunities for Google Cloud [3] Group 2 - NVIDIA is facing capacity constraints but is still projected to see revenue from Google customers grow over threefold this year, exceeding $20 billion [4] - The demand for alternative architectures is driven by a shortage in inference capabilities, highlighting Google's competitive advantage in the market [5] - Amazon AWS's absence from OpenAI's partner list raises concerns about its capacity constraints and the competitiveness of its Trainium chips [6][7]

Core Insights - OpenAI's shift to Google TPU chips marks a significant turning point in AI infrastructure, providing Google with a strong endorsement of its capabilities and potentially accelerating growth in its cloud business [1][2] - The collaboration allows OpenAI to reduce its reliance on Microsoft data centers while challenging NVIDIA's dominance in the GPU market [2][3] - Morgan Stanley projects substantial spending on NVIDIA GPUs, with estimates of $243 billion in 2027 and $258 billion in 2028, while TPU spending is expected to be around $21 billion and $24 billion in the same years [2] Group 1: Google and OpenAI Collaboration - OpenAI's adoption of Google TPU chips is its first large-scale use of non-NVIDIA hardware, which could lower inference computing costs [2] - This partnership is seen as a major recognition of Google's AI infrastructure capabilities, with OpenAI being the most significant TPU customer to date [2][3] - The collaboration is expected to drive accelerated growth in Google Cloud revenue, which has not yet been reflected in GOOGL's stock price [2] Group 2: NVIDIA's Market Position - Despite facing capacity constraints, NVIDIA is projected to see its revenue from Google clients grow over threefold this year, exceeding $20 billion [4] - NVIDIA's processor market share is expected to approach 65%, indicating strong demand despite current supply issues [4] - The demand for alternative architectures is driven by a shortage in inference capabilities, highlighting Google's differentiated advantage in the market [4] Group 3: Amazon AWS Challenges - OpenAI's absence from AWS indicates potential capacity constraints at Amazon, which may not meet OpenAI's requirements [5] - The choice of OpenAI to use Google's TPU over AWS's Trainium chips suggests competitive disadvantages for Amazon in the custom silicon space [5] - This dynamic is likely to increase investor scrutiny on AWS's growth and expectations for acceleration in the latter half of the year [6]

Group 1 - OpenAI has begun renting Google's TPU chips for the first time on a large scale, reducing its reliance on NVIDIA's GPUs and alleviating pressure on Microsoft's data centers [1][2] - OpenAI's demand for computing power has surged, with paid subscribers for ChatGPT increasing from 15 million at the beginning of the year to over 25 million, alongside hundreds of millions of free users [1] - Companies like Amazon, Microsoft, OpenAI, and Meta are developing their own inference chips to decrease dependence on NVIDIA and lower long-term costs [1][2] Group 2 - OpenAI spent over $4 billion on NVIDIA server chips last year, with training and inference costs each accounting for half, and is projected to spend nearly $14 billion on AI chip servers by 2025 [2] - The shift to Google's TPU was driven by the explosive popularity of ChatGPT's image generation tool, which increased pressure on OpenAI's inference servers [2] - Google has been developing TPU chips for about a decade and has provided this service to cloud customers since 2017, with other companies like Apple and Cohere also renting Google's TPU [2][4] Group 3 - Meta is also considering using TPU chips, indicating a broader trend among major AI chip customers [3] - Google Cloud continues to rent out NVIDIA-supported servers, as they remain the industry standard, generating more revenue than renting TPUs [4] - Google has ordered over $10 billion worth of the latest Blackwell server chips from NVIDIA, starting to provide them to select customers since February [4]