Shah 表示："尽管由于市场规模不断扩大，以及竞争对手超大规模数据中心与博通、迈威尔科技 (MRVL.US)和Alchip等设计公司合作采用自研芯片，预计谷歌的市场份额将在2027年下降至52%，但其 TPU 芯片群仍将是无可争议的行业核心支柱和指路明灯。这一基准线建立在训练和运行下一代Gemini 模型所需的大规模且持续的计算强度之上，而这需要持续、积极地提升自研芯片基础设施。" 此外，Counterpoint预计，到2028年，AI服务器运算ASIC的出货量将超过1500万颗，超过数据中心GPU 的出货量。 Shah指出："排名前十的AI超大规模数据中心运营商在2024年至2028年期间，累计部署的AI服务器运算 ASIC芯片将超过4000万颗。支撑这一空前需求的还有人工智能超大规模数据中心运营商，他们基于自 身技术栈构建了规模庞大的机架级人工智能基础设施，例如谷歌的 TPU Pod和AWS的 Trainium UltraCluster，使它们能够像一台超级计算机一样运行。" 台积电(TSM.US)在人工智能服务器运算ASIC出货量排名前十的公司中占据了近99%的晶圆制造份额。 尽管谷歌和亚马逊在 20 ...

好久没有聊AI芯片的内容了，似乎大家对这块的关注度也低了一些，但行业的暗流涌动一直都在，这篇文章结合产业内的一些资 料，聊一下海外几家大厂目前的格局。 一、TPU的崛起 今年一定要格外关注的就是TPU 的崛起，TPU链的相关公司也是我们重点关注的对象，当然有一些已经由于TPU的概念涨了非常 多了。 在英伟达GPU长期垄断AI训练与推理市场的背景下，Google TPU正凭借对LLM的原生优化优势，成为OpenAI、苹果等科技巨头 的重要选择，逐渐打破GPU一家独大的格局。OpenAI开始慢慢选择Google TPU开展核心推理业务；苹果也基于TPU搭建LLM训 练体系，这些合作其实暗含了复杂的技术适配门槛与隐性成本，并非简单的技术迭代选择。 核心问题在于技术体系的天然差异：OpenAI、苹果等企业的主流模型，最初均基于GPU生态完成训练，而GPU与TPU在数值表示 方式、精度体系上存在本质区别，从低精度到高精度的数据转换过程中，极易出现性能损耗与价值折损。从GPU迁移至TPU，必 须通过专门的软件桥接流程，将训练好的模型导出为TPU适配格式，这一转换过程并非一蹴而就，根据模型参数量的差异，复杂 模型的转换周期 ...

Core Insights - 2025 is identified as a year of significant integration among AI giants, with companies like Google, Meta, OpenAI, and Anthropic expanding their AI capabilities and entering the humanoid robotics space, leading to increased interdependence among them [1] Group 1: Industry Dynamics - OpenAI has expanded its cloud service partnerships beyond Microsoft, signing a $38 billion server deal with Amazon while also increasing its collaboration with Oracle [6] - Google has emerged as a major winner in the AI landscape, securing a $20 billion order from Anthropic for its TPU chips and negotiating a supply agreement with Meta [1][3] - The competitive landscape is shifting as companies aim to control more segments of the supply chain to reduce reliance on key suppliers like Nvidia, leading to complex alliances [1] Group 2: Company Strategies - Google has solidified its AI stack leadership by renting out TPU and cloud servers, while also providing Nvidia servers to OpenAI, positioning itself uniquely in the market [3] - OpenAI is investing in wearable AI devices, acquiring a design team for $6.5 billion, and aims to fill gaps in its AI stack to capture the growing consumer and enterprise AI service markets [6] - Meta has made strides in AI hardware with its Meta glasses but faces challenges in core technology development, prompting it to seek partnerships for chip resources [7] Group 3: Emerging Technologies - The humanoid robotics sector is becoming a new battleground, with major players like Google, Amazon, and OpenAI beginning to develop humanoid robot software and hardware, despite being in early stages [11] - xAI is making progress in language models and training clusters, although it still lags behind leaders like Google and OpenAI [8] - Microsoft is focusing on cloud service adjustments and partnerships, while Nvidia is restructuring to reduce direct competition in the cloud services market [12]

黄仁勋终于得到了他最想要的东西。 本周美国政府正式批准英伟达向中国以及其他"经批准的客户"出售高端的H200 GPU芯片，但需要向美国政府缴纳25%的销售提成。这一提成 比例同样适用于AMD、英特尔等其他美国芯片巨头。不过，英伟达最新的Blackwell和未来的Rubin系列GPU仍然禁止出口。 这标志着黄仁勋长达数月的游说取得成功。过去半年时间，他不断造访佛罗里达与华盛顿，随着特朗普总统一道出访和出席国宴，向白宫宴会厅建设工程 捐款，就是为了这一刻。就在上周，他再一次来到白宫会见总统，终于如愿以偿得到了解锁禁运令。 受这一利好消息推动，英伟达股价盘后应声上涨。受美国政府连续多道芯片加码禁运令限制，过去两年时间，英伟达一步步失去迅猛增长的中国市场，丢 掉了在AI GPU市场原先高达95%的份额。在英伟达最核心的数据中心业务，中国市场的营收占比也从原先的四分之一急剧下滑。 虽然英伟达数据中心业务营收高达1300亿美元(最近财年)，但却存在一个巨大隐患：客户集中度过高，过度依赖于几大AI巨头。其中，前两大客户营收占 比39%，前三大客户营收占比高达53%。 据媒体猜测，黄仁勋的前五大客户正是：微软、谷歌、亚马逊、 ...

Core Viewpoint - Nvidia's stock price plummeted, leading to a market value loss of over 1.4 trillion yuan, raising concerns about the sustainability of the AI boom as CEO Jensen Huang sold $1 billion worth of shares just before the drop [1][3][12] Group 1: Jensen Huang's Stock Sale - Jensen Huang's stock sale was executed through a legally permitted "10b5-1 trading plan," allowing him to sell shares at predetermined times and prices, which is a common practice among executives [3] - The timing of Huang's sale, coinciding with Nvidia's peak stock price, raises questions about whether he perceives the stock as overvalued [3][12] - Historical parallels are drawn to past instances where executives sold shares before market downturns, suggesting a potential warning sign for Nvidia [3][5] Group 2: Financial Performance and Market Reaction - Nvidia's latest quarterly earnings report showed a 126% year-over-year revenue increase and a threefold net profit increase, but the data center revenue fell short of Wall Street's expectations by $200 million [5][6] - The $200 million shortfall, while only 0.4% of the total data center revenue, led to a staggering market value loss of $180 billion, indicating that market expectations for Nvidia were excessively high [5][6] - Nvidia's current price-to-earnings (P/E) ratio exceeds 70, significantly higher than competitors like Apple and Microsoft, suggesting that any slowdown in growth could lead to a sharp decline in stock price [6] Group 3: Competitive Landscape and Market Dynamics - Nvidia's dominance in AI hardware is being challenged as major clients like Google, Amazon, and Microsoft develop their own chips to reduce dependency on Nvidia [8][9] - Competitors such as AMD and Intel are also entering the market with competitive products, further threatening Nvidia's market share [8][9] - The increasing energy demands of AI model training may lead to a slowdown in GPU purchases, questioning the sustainability of Nvidia's growth as an "AI printing machine" [9] Group 4: Industry Outlook and Future Considerations - The recent stock decline is viewed as a "valuation correction" rather than an industry collapse, with AI technology still poised to transform various sectors [11][12] - The AI sector may experience a shakeout where companies lacking technological strength may fail, while those with solid foundations could thrive post-correction [11][12] - Huang's stock sale reflects a cautious approach to market dynamics, emphasizing the importance of not overestimating the company's position and preparing for potential challenges [11][12]

Core Viewpoint - The competition for dominance in the AI custom chip market has intensified, with Broadcom as the leader and Marvell facing challenges but showing signs of potential recovery due to new client developments [1][2]. Market Position and Competition - Broadcom holds the largest market share in the AI workload ASIC market, while Marvell had aimed for a 20% market share but faced setbacks due to increased competition and client issues [1][2]. - Marvell's AI chip business is heavily reliant on two major clients, Amazon AWS and Microsoft Azure, leading to uncertainty in revenue forecasts [2][5]. Revenue Growth and Projections - The overall AI acceleration chip market is experiencing a compound annual growth rate (CAGR) of 50%-60%, with Broadcom's CEO projecting at least 60% growth for their AI business [2]. - Marvell's AI revenue growth is expected to be below 50%, with projected AI-related revenue of approximately $3 to $3.5 billion by fiscal year 2026 [3][5]. Client Developments - A new significant client is expected to increase investment in custom AI chip development, which could positively impact Marvell's 2026 performance outlook [1]. - Microsoft is advancing its self-developed AI chip project, "Maia," which may lead to additional revenue for Marvell if they are involved in the design process [7][8]. Financial Outlook - Marvell's management has shown confidence through substantial share buyback programs and insider buying, indicating optimism about future performance [7]. - Analysts currently estimate Marvell's AI revenue at around $3 billion, contributing to an overall revenue projection of approximately $8.15 billion for fiscal year 2025, reflecting a 41% year-over-year growth [7]. Valuation and Investment Potential - If Marvell secures $500 million to $1 billion in additional revenue from the Microsoft Maia project, total revenue for fiscal year 2026 could approach $10.5 billion, suggesting an attractive forward valuation of 7.4 times sales [8]. - Marvell's stock appears to be appealing within a valuation range of 7-8 times FY26 sales, compared to the current 8.3 times [8].

Core Viewpoint - The article emphasizes the necessity for cloud vendors to continuously invest in computing power, models, chips, and ecosystems to build a "super AI cloud" [2][25]. Group 1: AI Cloud Competition - AI cloud has become a new entry ticket in the cloud computing arena, crucial for vendors to escape price wars and rebuild competitive advantages [2]. - The competition for "AI Cloud No. 1" is intensifying among domestic cloud vendors, with the focus on market leadership becoming a core industry concern [2]. - Globally, only four major players remain in the AI cloud space: AWS, Microsoft, Google, and Alibaba Cloud [2][11]. Group 2: Evaluation Criteria for AI Cloud Leaders - The evaluation of who is the "AI Cloud No. 1" depends on various standards, with models being a key factor for some [5][6]. - The article outlines four critical questions to assess the capabilities of AI cloud vendors: 1. Annual infrastructure investment of at least 100 billion [6]. 2. Possession of million-level large-scale computing clusters and cloud scheduling capabilities [8]. 3. Availability of top-tier large model capabilities that perform across various scenarios [9]. 4. Strategic layout of AI chip computing power [10]. Group 3: Capital Expenditure Insights - Major cloud vendors like Google, Microsoft, and AWS have significantly increased their capital expenditures to meet the explosive growth in AI infrastructure demand, with Google raising its annual target to $85 billion [6][7]. - Alibaba's capital expenditure for 2024 is projected at 76.7 billion RMB, significantly lower than its competitors, indicating a disparity in financial strength [10]. Group 4: Development Models - Two primary development models are identified: "Cloud + Ecosystem" (AWS and Microsoft) and "Full Stack Self-Research" (Google and Alibaba) [12][19]. - The "Cloud + Ecosystem" model allows vendors to leverage external models, reducing R&D costs and risks while increasing platform attractiveness [14][15]. - The "Full Stack Self-Research" model involves significant upfront investment but can create a strong competitive moat and higher long-term value [19][20]. Group 5: Alibaba Cloud's Position - Alibaba Cloud is positioned as a representative of the "Full Stack Self-Research" model in the Eastern context, competing closely with Google Cloud [25]. - The company plans to invest over 380 billion RMB in cloud and AI hardware infrastructure over the next three years, demonstrating a commitment to enhancing its capabilities [24]. - Alibaba Cloud's strategy includes embracing open-source models, creating a large AI model community, and addressing hardware constraints through software ecosystem development [24][25].

Core Viewpoint - NVIDIA's CEO Jensen Huang asserts that NVIDIA's growth will continue to outpace that of Application-Specific Integrated Circuits (ASICs), citing a high failure rate among ASIC projects and emphasizing NVIDIA's technological advancements and cost optimization [2][3]. Group 1: NVIDIA's Market Position - Huang believes that while many companies are developing ASICs, about 90% will fail, similar to the high failure rate of startups [2]. - NVIDIA is not overly concerned about the competition from ASICs, as they recognize that without NVIDIA, the computing field cannot thrive [3]. - Huang emphasizes that the development of ASICs is not the main challenge; rather, the deployment requires significant investment and expertise, which NVIDIA possesses [4]. Group 2: NVLink Fusion Announcement - NVIDIA introduced NVLink Fusion, a technology aimed at integrating third-party CPUs and accelerators with NVIDIA's ecosystem, allowing for semi-custom designs [5][7]. - NVLink Fusion enables non-NVIDIA CPUs to connect to NVIDIA GPUs via a short-distance chip-to-chip connection, enhancing flexibility for system vendors [9][11]. - The technology is seen as a step towards allowing third-party chip manufacturers to integrate their designs with NVIDIA's high-performance NVLink network [15]. Group 3: Industry Collaboration - Companies like Alchip, AsteraLabs, Marvell, and MediaTek are confirmed to be developing accelerators that will support NVLink Fusion, indicating a growing ecosystem around NVIDIA's technology [15]. - Fujitsu and Qualcomm are also working on new CPUs that will pair with NVIDIA GPUs, aiming to enhance efficiency through NVLink Fusion [15]. - Cadence and Synopsys are participating as technical partners in the NVLink Fusion initiative, providing IP blocks and design services to companies looking to build compatible hardware [16].

Core Viewpoint - NVIDIA's CEO Jensen Huang asserts that NVIDIA's growth will continue to outpace that of Application-Specific Integrated Circuits (ASICs), citing a high failure rate among ASIC projects and emphasizing NVIDIA's rapid technological advancements and cost optimization [1][2][3]. Group 1: NVIDIA's Market Position - NVIDIA is not concerned about being marginalized in the AI market, recognizing its essential role in the computing field [2]. - Huang believes that most ASIC projects will be canceled if they do not outperform existing chips, indicating a competitive landscape where NVIDIA's technology remains superior [2][3]. Group 2: NVLink Technology - NVIDIA has introduced NVLink Fusion, a new technology aimed at integrating third-party CPUs and accelerators with NVIDIA's ecosystem, enhancing flexibility for system suppliers [5][7]. - NVLink has evolved since its introduction in 2016, significantly increasing bandwidth and enabling faster interconnects between GPUs [6][9]. Group 3: Future Developments - The NVLink Fusion initiative allows for semi-custom designs, enabling third-party chips to connect with NVIDIA GPUs, although it remains proprietary [10][14]. - Companies like Fujitsu and Qualcomm are developing CPUs that will support NVLink Fusion, aiming to improve efficiency and performance [16]. Group 4: Industry Collaboration - Cadence and Synopsys are participating as technical partners in the NVLink Fusion program, providing IP blocks and design services to companies looking to build compatible hardware [17].