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 ...

Counterpoint Research says the race to build custom AI silicon is accelerating, with hyperscalers scaling internal chips to meet surging demand. • Broadcom stock is showing upward movement. What’s driving AVGO shares up?Hyperscalers Ramp Custom AI ChipsLeading cloud and AI providers, including Alphabet Inc.’s (NASDAQ:GOOGL) Google; Amazon.com, Inc.’s (NASDAQ:AMZN) Amazon Web Services; Microsoft Corp (NASDAQ:MSFT) ; OpenAI; ByteDance and Apple Inc (NASDAQ:AAPL) , are rapidly expanding deployments of AI serve ...

如果您希望可以时常见面，欢迎标星收藏哦~ 微软也加入了大型科技公司摆脱对英伟达依赖的浪潮，推出了自己的人工智能（AI）芯片。各大 科技公司都在开发定制芯片或寻求供应商多元化，以降低对英伟达的依赖——英伟达占据了AI芯 片市场90%的份额。然而，英伟达以其图形处理器（GPU）为代表，正通过构建AI工厂展开反 击 。 它 不 再 仅 仅 销 售 GPU ， 而 是 通 过 垂 直 整 合 芯 片 、 服 务 器 、 软 件 和 模 型 ， 转 型 为 一 家 " 全 栈 AI"基础设施公司，决心不放弃其在AI市场的领导地位。预计英伟达今年将成为台积电最大的客 户。尽管一年前中国市场曾因DeepSeek芯片强调性价比而引发"冲击"，但英伟达的股价和销售额 依然大幅增长。 加速摆脱英伟达 由于价格高昂、供应短缺以及封闭的生态系统（CUDA），大型科技公司正在加速摆脱对英伟达 GPU的依赖。 NVIDIA GPU的高昂成本是关键驱动因素。它们不仅价格昂贵，而且供应常常无法满足需求，导 致及时采购困难重重。此外，尽管NVIDIA芯片用途广泛，但它们并未针对特定公司的特定AI任务 进行优化。因此，大型科技公司正在开发专为自 ...

Core Viewpoint - Jensen Huang has established a $4.6 trillion empire through Nvidia, focusing on AI infrastructure, but there are three significant threats to the company's future that are not addressed in earnings calls [1] Group 1: Threats to Nvidia - **Threat 1: Major Customers Developing In-House Chips** Microsoft, Meta, Amazon, and Alphabet account for 40-50% of Nvidia's revenue and are all creating custom AI chips, which could replace Nvidia's offerings. Inference workloads, which represent 80% of long-term AI compute, are at risk if these companies build their own chips [2][3] - **Threat 2: AMD as a Competitive Alternative** AMD's MI300X chips have gained traction, offering competitive performance at 20-30% lower costs compared to Nvidia. Microsoft Azure and Oracle Cloud are adopting AMD technology, and OpenAI is reportedly testing AMD chips to reduce dependency on Nvidia [4][5][6] - **Threat 3: Geopolitical Risks from China** China's approval of H200 chips may seem positive, but it poses a risk as the country has a history of extracting technology and then developing domestic alternatives. If Nvidia becomes too reliant on the Chinese market, future bans could severely impact revenue [7][8] Group 2: Nvidia's Strategic Omissions - **Lack of Discussion on Customer Developments** Jensen Huang focuses on AI demand and partnerships in earnings calls but avoids discussing customer chip development, AMD's market share, and the implications of inference versus training margins [9][10] - **Market Realities Ignored** The optimistic view assumes AI growth benefits all players, while the pessimistic view recognizes that customers are building their own solutions, AMD is providing cheaper options, and geopolitical tensions could threaten Nvidia's market position [10]

Core Viewpoint - The news highlights a significant market reaction to the rumor that Intel will manufacture Apple's M-series chips, indicating a potential shift in the semiconductor landscape and a re-evaluation of Intel's market position [1][3]. Group 1: Apple's Endorsement - Apple ships 20 million "standard" M-series chips annually, and transferring production to Intel would significantly impact Intel's business [4]. - Apple's role as a stringent quality inspector adds credibility to Intel's manufacturing capabilities, especially for the simpler M-series chips [4]. - Intel has entered a substantive collaboration phase with Apple, having signed a confidentiality agreement and received advanced process design kits (PDK) [6][4]. Group 2: Cook's Strategy - Apple's decision to support Intel, despite TSMC's strong performance, serves as a political statement and aligns with U.S. manufacturing policies [7]. - By outsourcing the production of lower-end M-series chips to Intel, Apple aims to diversify its supply chain and reduce dependency on TSMC [8][10]. - Establishing a dual-supplier system with Intel and TSMC is crucial for Apple to mitigate capacity risks and enhance bargaining power [9]. Group 3: Valuation Reconstruction - The market's reaction reflects a potential breaking of Intel's "IDM curse," as major tech companies show interest in Intel's manufacturing capabilities [11][16]. - Intel's previous struggles with its IDM model have led to significant capital expenditures with minimal returns, but the prospect of securing high-profile clients could change this narrative [14][15]. - The involvement of top-tier clients like Apple, Google, and Meta increases the likelihood of Intel's success in its foundry business, potentially leading to a substantial increase in its market valuation [17][18].

Core Insights - Nvidia's GPUs dominate the AI chip market with a 90% share, but competition is increasing as tech giants develop custom ASICs, threatening Nvidia's leadership [1][3] - The shift from "training" to "inference" in AI development favors more energy-efficient chips like TPUs and NPUs over traditional GPUs [5][6] Group 1: Nvidia's Market Position - Nvidia's GPUs are priced between $30,000 to $40,000, making them expensive and contributing to Nvidia becoming the highest-valued company globally [1] - Major tech companies are moving towards developing their own chips, indicating a potential decline in Nvidia's dominance in the AI sector [1][3] Group 2: Custom AI Chips - Google's TPU, designed specifically for AI, outperforms GPUs in certain tasks and is more energy-efficient, leading to lower operational costs [3][5] - Companies like OpenAI and Meta are investing in custom chips, with OpenAI planning to produce its own chips in collaboration with Broadcom [3][5] Group 3: Economic Factors - The cost of installing Nvidia's latest GPUs is significantly higher than that of Google's TPUs, with estimates of $852 million for 24,000 Nvidia GPUs compared to $99 million for the same number of TPUs [5] - The emergence of cheaper custom chips is expected to alleviate concerns about an AI investment bubble [5] Group 4: AI Ecosystem Changes - The AI ecosystem centered around Nvidia is likely to change as large tech companies collaborate with chip design firms, creating new competitors [6] - The current manufacturing landscape, dominated by TSMC for Nvidia chips, may shift as companies develop their own semiconductor solutions [6] Group 5: Chip Types - CPUs serve as the main processing units but are slower compared to GPUs, which can handle multiple tasks simultaneously [8] - TPUs are specialized for AI tasks, while NPUs are designed to mimic brain functions, offering high efficiency for mobile and home devices [8]

Core Viewpoint - The main debate on Wall Street regarding Nvidia Corp centers on the demand for AI, but the more critical question is how long Nvidia can maintain high margins of over 70% before hyperscalers seek alternatives [1] Group 1: Nvidia's Market Position - Nvidia's primary threat is not from competing GPUs but from Google's TPUs, which signify a shift where hyperscalers may stop outsourcing the most profitable aspects of AI [1] - Google is scaling TPUs not to compete in hardware but to reduce its dependency on Nvidia, allowing it to run AI on its own terms and infrastructure [2] - TPUs only need to be "good enough" for large in-house workloads, which allows hyperscalers to erode Nvidia's pricing power gradually [3] Group 2: Industry Trends - The risk for Nvidia arises when hyperscalers realize that custom silicon can significantly improve their gross margins, leading them to seek alternatives to Nvidia [4] - Major companies like Amazon, Meta, and Microsoft are already developing their own alternatives, indicating a trend away from reliance on Nvidia [4] - Nvidia does not need to lose compute share to lose its margin leadership; it only requires hyperscalers to create credible alternatives that set a price ceiling [5] Group 3: Investor Insights - While the demand for AI remains strong, the pricing power of Nvidia is in jeopardy, as the company may face negotiations rather than obsolescence [6] - Once hyperscalers gain real leverage, the notion of maintaining "70% margins forever" will become a thing of the past [6]

Core Insights - The demand for AI HPC (High-Performance Computing) is driving the need for advanced packaging solutions, particularly TSMC's CoWoS technology, but some cloud service providers (CSPs) are considering Intel's EMIB technology due to increasing chip integration requirements [1][2] Group 1: CoWoS Technology - TSMC's CoWoS solution connects different functional chips using an interposer, with various versions like CoWoS-S, CoWoS-R, and CoWoS-L developed [1] - The market demand is shifting towards CoWoS-L, especially with NVIDIA's upcoming Blackwell platform set for mass production in 2025 [1] Group 2: EMIB Technology - EMIB offers several advantages over CoWoS, including a simplified structure that eliminates the expensive interposer, leading to higher yield rates [2] - EMIB has a smaller thermal expansion coefficient issue due to its design, which reduces the risk of packaging warping and reliability challenges [2] - EMIB can achieve larger packaging sizes, with EMIB-M already supporting 6 times the mask size, and projections for 8 to 12 times by 2027 [3] Group 3: Market Dynamics - The demand for CoWoS is facing challenges such as capacity shortages and high costs, prompting CSPs like Google and Meta to explore EMIB solutions [2] - Intel's EMIB technology has been in development since 2021 and is already applied in its server CPU platforms, with Google planning to implement it in TPUv9 by 2027 [3] - NVIDIA and AMD, which require high bandwidth and low latency, are expected to continue using CoWoS as their primary packaging solution [3]

Key Takeaways from J.P. Morgan's Expert Discussion on AI Datacenters Industry Overview - The discussion focused on the **AI Datacenter** industry, particularly the scaling up of AI Datacenters and the evolving architecture for hyperscale AI workloads. Core Insights 1. **Shift in Compute Capex**: - There is a rapid shift in compute capital expenditures (capex) towards inference workloads, with techniques like distillation and multi-step optimization yielding significant near-term gains. By approximately **2027**, the share of compute dedicated to inference is expected to surpass that of training workloads [3][4][5]. 2. **Preference for Smaller Models**: - Enterprises are increasingly adopting smaller, fine-tuned models over larger ones, accepting slight quality trade-offs for reduced costs in inference workloads. This trend is exemplified by Cursor's new coding model [3][4]. 3. **Standardization in Hardware**: - The industry is witnessing a move towards standardization in inference-related networking hardware, with expectations for more rack-level standardization in the coming year. White-box solutions are gaining traction through Open Compute Project (OCP) initiatives [3][4]. 4. **Training Constraints**: - Training workloads are facing constraints primarily due to power supply issues, while inference workloads are less affected. The power demands for training are significantly higher, estimated at **5-10 times** that of inference [4][5]. 5. **Longer GPU Lifespan**: - Buyers are now planning for a useful life of **five to six years** for GPUs, an increase from the previous **four years**. This shift reflects a strategic move to repurpose GPUs from training to inference tasks [5]. 6. **Storage Solutions**: - The storage landscape remains hybrid, with HDDs maintaining cost leadership while Flash/NAND is preferred for high-performance needs. Advances in HDD technology, such as HAMR, are helping HDDs remain competitive [5]. 7. **Beneficiaries of Capex Shift**: - Companies like **Broadcom**, **Marvel**, and **Celestica** are expected to benefit from the shift towards inference workloads. Broadcom's work with custom ASICs for major players like Google and Amazon positions it favorably in this evolving market [5]. Additional Important Points - The discussion highlighted the growing comfort among operators in mixing branded and white box solutions, indicating a trend towards flexibility and cost-effectiveness in hardware choices [1][3]. - The preference for Ethernet and PCIe for inference workloads is driven by cost considerations and the ease of capacity expansion, contrasting with the continued use of InfiniBand for training clusters [3][4]. - The call emphasized the importance of co-packaged optics for high bandwidth requirements, particularly for workloads exceeding **1.6T** [3][4]. This comprehensive analysis provides insights into the current trends and future expectations within the AI Datacenter industry, highlighting key shifts in technology, investment strategies, and market dynamics.

Core Insights - AI datacenters are entering a new phase focused on inference rather than training, which is expected to reshape the competitive landscape and spending patterns in the industry [1][2][8] Shift from Training to Inference - The focus is shifting from training large models to optimizing inference processes, with techniques like distillation and quantization making inference cheaper and more efficient [2][3] - By 2027, inference is expected to dominate incremental compute spending, with a notable shift already occurring in 2025-2026 [3] Beneficiaries of the Shift - Broadcom is highlighted as a key beneficiary due to its custom ASICs that support inference for major companies like Google, Amazon, and Meta [4] - Marvell Technology is also positioned to benefit as inference workloads increasingly rely on Ethernet and PCIe, moving away from expensive training-oriented technologies [5] Hardware and Networking Trends - Celestica is well-positioned as the industry moves towards standardized, cost-effective inference hardware, allowing operators to source from multiple vendors [6] - Arista Networks continues to support high-performance training networks, but the shift towards Ethernet in inference may create new opportunities for networking companies [6] Power Efficiency and Deployment - Inference is significantly less power-hungry than training, often requiring 5-10 times less power, making it easier to deploy in datacenters with limited grid capacity [7] - The trend towards making AI cheaper, faster, and easier to run is expected to drive spending towards companies like Broadcom and Marvell [8]