根据摩根士丹利Joseph Moore团队25日发布的报告， 这种强劲的需求主要驱动因素在于token生成量的 增长，自年初以来，token生成量增长了5倍以上 ，这给生态系统带来了巨大压力，并推动了对处理这些 工作负载的投资激增。 点击 上方 硬AI 关注我们 大摩指出，受益于大型语言模型对推理芯片的巨大需求，英伟达面临GPU供不应求局面。但在持续的供应限制、毛利率 压力等负面影响下，大摩轻微下调英伟达目标价至160美元。长期来看，公司增长轨迹依然强劲。 硬·AI 作者 | 张雅琦 编辑 | 硬 AI 过去四周，投资者情绪因宏观经济和供应链风险而恶化，但与此同时，对英伟达GPU核心的需求却因主要 大型语言模型（LLM）对推理芯片的巨大需求而飙升，且这种需求遍及所有地区。 多家AI公司报告用户数量呈爆炸式增长，例如，Open Router等API公司的数据显示，许多公司为满足推 理软件的巨量需求，被迫争抢GPU资源，甚至出现"最后一块GB200"在2025年仅剩一块的状况。 摩根士丹利认为， 这种对推理的需求是关键。 这是由使用模型并产生收入的部分驱动的，证明了推理模 型的扩展是真实存在的，这与仅依赖于风险投 ...

Group 1 - Investor sentiment has deteriorated due to macroeconomic and supply chain risks, but demand for NVIDIA's GPUs has surged due to the significant need for inference chips driven by large language models (LLMs) [1] - Token generation has increased over five times since the beginning of the year, creating immense pressure on the ecosystem and driving a surge in investment to handle these workloads [1] - AI companies are experiencing explosive user growth, with many forced to compete for GPU resources to meet the massive demand for inference software [1] Group 2 - Morgan Stanley has lowered its target price for NVIDIA to $160 from $162, reflecting overall valuation declines in the peer group rather than changes in the company's fundamentals [2] - Despite strong demand, supply constraints for NVIDIA's Blackwell chips, particularly the GB200/300 models, are limiting the ability to meet the explosive growth in demand [2][4] - Morgan Stanley has raised its revenue forecast for fiscal year 2026 by 10.7% and adjusted earnings per share up by 11.9%, indicating that these figures may still be conservative [5]

点击 上方 硬AI 关注我们 测试结果显示，相比前代英伟达Hopper GPU，GB200 NVL72服务器能帮助Cohere在1000亿参数模型的训练实现高达3 倍的性能提升，此外，IBM和Mistral AI也已成为CoreWeave GB200云服务的首批用户。 "世界各地的企业和组织正在竞相将推理模型转化为代理型人工智能应用，这将改变人们的工作和娱 乐方式。" 硬·AI 作者 | 李笑寅 编辑 | 硬 AI CoreWeave再度抢占先机，率先部署英伟达GB200系统，AI巨头争相入局。 英伟达今日在其博客上宣布， AI云计算提供商CoreWeave已成为首批大规模部署英伟达GB200 NVL72 系统的云服务提供商之一。Cohere、IBM和Mistral AI已成为首批用户。 根据最新MLPerf基准测试，这些系统提供了前代H100芯片2-3倍的性能提升，将显著加速大模型训练和推 理能力。 CoreWeave首席执行官Michael Intrator表示，这一成就既展示了公司的工程实力和执行速度，也体现了 其对下一代AI发展的专注： "CoreWeave的设计就是为了更快速地行动——我们一次又一次 ...

Core Insights - The event highlighted the transformative shift in data centers towards AI-driven computing, with projected capital expenditures exceeding $1 trillion by 2028 for data center construction, primarily focused on accelerated computing chips [2][12][13] - NVIDIA's Blackwell architecture is fully operational, showcasing significant performance improvements and a roadmap for future products like Rubin and Feynman, which promise substantial enhancements in computational power and efficiency [3][42][45] - The introduction of the Quantum-X CPO switch and Spectrum-X technology aims to revolutionize networking capabilities, reducing energy consumption and increasing deployment efficiency [5][46] - The advancements in AI applications, particularly in autonomous driving and robotics, are supported by NVIDIA's new systems and frameworks, enhancing the development and training processes [6][26][24] Capital Expenditure and AI Infrastructure - Data center capital expenditures are expected to reach $1 trillion by 2028, with a significant portion allocated to accelerated computing chips [2][12] - NVIDIA plans to deliver 1.3 million Hopper GPUs to major cloud service providers in 2024, with an increase to 3.6 million Blackwell GPUs in 2025 [2][3] AI Model Training and Inference - The demand for computational power for AI training and inference has surged, with estimates suggesting a 100-fold increase in required computing resources compared to the previous year [10][11] - NVIDIA outlines three levels of AI: Generative AI, Agentic AI, and Physical AI, each representing a different stage of AI development and application [8][10] Product Development and Future Roadmap - Blackwell has been fully launched, with significant customer demand and performance improvements, including a 40-fold increase in inference performance compared to previous models [3][42] - Future products like Vera Rubin and Rubin Ultra are set to enhance computational capabilities further, with expected performance increases of up to 15 times [45][42] Networking Innovations - The Quantum-X CPO switch is anticipated to launch in late 2025, offering substantial energy savings and improved network efficiency [5][46] - Spectrum-X technology will provide high bandwidth and low latency, integrating seamlessly into NVIDIA's computing architecture [5][46] AI Applications in Autonomous Driving and Robotics - NVIDIA's Halos system aims to enhance safety in autonomous vehicles, while the open-source Isaac Groot N1 model supports robotics development [6][24] - The integration of Omniverse and Cosmos platforms accelerates the development of AI for autonomous driving, enabling end-to-end training capabilities [26][24] Data Center Evolution - The transition of data centers into AI factories is underway, focusing on processing, analyzing, and generating AI-driven applications [12][13] - NVIDIA's Dynamo operating system is designed to optimize AI factory operations, enhancing efficiency and performance [35][36]

点击 上方 硬AI 关注我们 摩根大通表示，英伟达、Marvell、亚马逊等客户确实下调了2025年的CoWoS订单预期，降幅约为8-10%。但这并非需 求问题，而是客户此前过度乐观的预期回归理性。2025年CoWoS产能仍将供不应求，英伟达的Blackwell芯片出货量有 望达到600万片。 硬·AI 作者 |赵 颖 CoWoS订单削减并非需求问题，而是客户此前过度乐观的预期回归理性。2025年CoWoS产能仍将 供不应求，英伟达的Blackwell芯片出货量有望达到600万片。 "砍单"疑云：预期调整，而非需求萎缩 关于为何会出现"砍单"的说法？摩根大通的供应链调查显示： 英伟达、Marvell、亚马逊等客户确实下调了2025年的CoWoS订单预期，降幅约为8-10%。其 中，英伟达的产能预期下调了约4-4.5万晶圆。 但这并不意味着需求出了问题。这些调整更多是由于客户最初的预期过于乐观，远超台积电及整个 生态系统的供应能力。随着2025年交付时间临近，台积电开始要求客户提供更准确的预测，促使客 户修正了此前的过度预订。 摩根大通还指出，产品变化和订单优先级也可能导致了供应链的预期调整： 英伟达有多款产品 ...

Investment Rating - The report maintains an "Overweight" rating for TSMC with a price target of NT$1500.0 [3][7]. Core Insights - The report indicates that recent adjustments in CoWoS order projections from customers like NVDA and Marvell/AWS do not reflect a demand issue, as overall CoWoS demand remains higher than TSMC's capacity to fulfill in 2025 [2][4]. - TSMC's CoWoS capacity is expected to remain tight through 2025, despite a more than double capacity expansion, driven by strong demand from AI labs and US CSPs [4][6]. - NVDA's demand for CoWoS wafers is projected to reach approximately 390k in 2025, sufficient for producing around 6 million Blackwell chips [6][7]. Summary by Sections Demand and Supply Dynamics - Adjustments in customer projections stem from overbooking and TSMC's push for accurate forecasts, rather than a decline in demand [4][5]. - Strong demand trends are noted for various products, including NVDA's H200s and H20s, indicating a robust market environment [4][5]. Capacity and Production Forecasts - TSMC's CoWoS capacity is projected to reach 75k wafers per month by the end of 2025, with total capacity for the year expected to be around 725k wafers [4][6]. - NVDA's CoWoS wafer demand is expected to grow by approximately 20% in 2026, supporting the production of 7.5 million Blackwell and Rubin chips [6][7]. Market Sentiment and Stock Outlook - Despite potential short-term negative sentiment due to order adjustments, the underlying demand for AI hardware remains strong [6][7]. - The report expresses a positive outlook on TSMC and its AI ecosystem, emphasizing its critical role in enabling AI technologies across various sectors [6][7].