智通财经获悉，周四盘后交易中，CoreWeave(CRWV.US)股价一度暴跌12%。这家专注于人工智能的云 基础设施供应商公布的第四季度营收虽高于华尔街预期，但未能提振股价。根据公告，CoreWeave第四 季度营收同比增长110%至15.7亿美元，略超出市场平均预期15.5亿美元，每股亏损89美分，逊于市场平 均预期。截至发稿，该股盘后大跌9.56%。 业绩指引方面，公司预计2026年全年营收为120亿至130亿美元，分析师此前预期为120.9亿美元;预计 2026年调整后营业利润为9亿至11亿美元。但第一季度的营收指引为19亿至20亿美元，低于市场共识预 期的22.9亿美元。 本季度CoreWeave与模型开发商Poolside达成合作，并推出对象存储服务，将信贷额度从15亿美元提升 至25亿美元。尽管存储业务将助其与亚马逊AWS等巨头竞争，但公司仍聚焦专业云基础设施领域。 Intrator在博客中写道："2025年CoreWeave成为史上最快突破50亿美元年营收的云平台。" 美股频道更多独家策划、专家专栏，免费查阅>> 责任编辑：栎树 公司计划将2026年资本支出目标定为300亿至350亿美元，远超 ...

CoreWeave首席执行官Mike Intrator在分析师电话会议上表示，其服务核心英伟达图形芯片仍供不应求。 他指出，第四季度英伟达H100处理器均价较年初波动在10%以内，而旧款A100芯片价格在2025年出现 上涨。 公司计划将2026年资本支出目标定为300亿至350亿美元，远超2025年的103.1亿美元。截至去年底，公 司活跃电力容量为850兆瓦，签约电力达3.1吉瓦(市场预期活跃电力约827兆瓦)。CoreWeave预计2026年 底活跃电力将超1.7吉瓦(高于市场预期的1.59吉瓦)，并计划到2030年在现有签约规模基础上新增超5吉 瓦。 Intrator表示："需求正从最初集中于超大规模云和基础模型领域，向全经济领域扩散。如今企业级需求 爆发式增长，主权领域需求兴起，新兴参与者纷纷入场锁定所需基础设施。"公司营收积压订单从第三 季度末的556亿美元激增至668亿美元。 调整后税息折旧及摊销前利润为8.98亿美元，低于市场预期的9.29亿美元。自去年3月上市后，截至12月 31日公司债务达213.7亿美元。 周四盘后交易中，CoreWeave(CRWV.US)股价一度暴跌12%。这家专注于 ...

Core Viewpoint - The U.S. government's issuance of chip manufacturing equipment licenses to Samsung and SK Hynix, while leaving TSMC's situation unresolved, indicates a strategic adjustment in U.S. export control policies rather than a simple policy shift [1][3]. Group 1: U.S. Export Control Policies - The U.S. has replaced the "Verified End User" (VEU) exemption with an annual approval system for companies like TSMC, Samsung, and SK Hynix, citing vulnerabilities in the previous system that could lead to improper technology transfer to China [1]. - As a result of these restrictions, Samsung's production capacity utilization at its Xi'an factory has decreased by 12% [1]. Group 2: Impact on Korean Companies - Korean companies are feeling significant pressure due to the U.S. export controls, despite efforts by the South Korean government to negotiate transitional buffers with the U.S. [3]. - The Chinese government has expressed strong opposition to U.S. actions, viewing them as unilateral suppression that threatens global supply chain stability [3]. Group 3: China's Semiconductor Industry - Despite U.S. pressure, China's semiconductor industry is finding opportunities for advancement, achieving self-sufficiency in mature process technologies, with SMIC's 28nm process yield reaching 95% [3]. - In the high-end AI chip sector, Huawei's Ascend 910B chip is nearing the performance of NVIDIA's A100, indicating significant competitive advancements in certain areas [3]. Group 4: Global Semiconductor Landscape - The ongoing semiconductor competition involves not only technology but also national security and economic interests, with countries like the U.S. and China engaging in complex geopolitical interactions [5]. - The U.S. aims to force companies like Samsung to compete with local firms in China through its annual licensing system, but this strategy may inadvertently weaken U.S. technological barriers and bolster China's semiconductor development [5]. Group 5: Supply Chain Dynamics - Many multinational companies, including Microsoft and Apple, are attempting to mitigate risks by relocating production and supply chains, although the costs associated with these transitions remain significant [5]. - Chinese companies are expanding their operations in regions like Mexico and Southeast Asia, promoting regional and diversified development of the supply chain [7]. Group 6: Future Outlook - The dynamics of technology iteration and market rules will dominate the ongoing semiconductor competition, with Bloomberg's analysis suggesting that reliance on tariffs and export controls may not effectively reshape market dynamics [7]. - The semiconductor industry requires continuous innovation and adaptability to secure advantageous positions in the evolving international landscape [8].

Core Viewpoint - The article emphasizes that electricity has become a critical constraint on the development of artificial intelligence (AI), with the demand for power surging due to the exponential growth of AI models and data centers [2][14]. Group 1: AI Power Consumption - ChatGPT consumes approximately 2.9 watt-hours per response, which is nearly ten times the energy used by a traditional Google search, leading to a daily consumption of over 500,000 kilowatt-hours [3][9]. - The energy required for training AI models has significantly increased, with GPT-3 consuming 1,287 megawatt-hours for a single training session, enough to power 3,000 Tesla vehicles for 200,000 miles [9][10]. - The projected AI computing power in the U.S. is expected to require about 1,269 terawatt-hours by 2030, accounting for 22% of the total electricity consumption [10][12]. Group 2: Electricity Supply Challenges - The U.S. electricity grid is aging, with 70% of transformers exceeding their 25-year design life, leading to a low load reserve margin of only 20% [10][11]. - The average outage duration for U.S. users reached 662.6 minutes in 2024, a year-on-year increase of 80.74%, with states like Virginia and Texas experiencing even longer outages [11]. - The demand for electricity from AI is characterized by "pulse-like" spikes, which poses significant challenges to grid stability [10][11]. Group 3: Renewable Energy Solutions - Solar power combined with energy storage is viewed as the most viable solution to meet the growing electricity demands of AI, given its economic advantages and shorter construction timelines compared to other energy sources [15][17]. - The cost of solar power generation is the lowest among various energy sources, with prices ranging from $0.038 to $0.078 per kWh, making it an attractive option for data centers [17][18]. - Major tech companies like Google, Microsoft, and Amazon have set ambitious goals for achieving 100% renewable energy for their data centers by 2030, indicating a strong commitment to sustainable energy solutions [20][21].

Group 1 - The Trump administration is considering approving the export of NVIDIA's H200 AI chips to China, which have significantly improved performance compared to the previous H100 chips, with H200 estimated to be twice as powerful as H100 [1] - The H200 chip features HBM3e memory, providing a memory speed of 4.8TB per second and a memory capacity that is approximately double that of the A100, with a bandwidth increase of 2.4 times [1] - NVIDIA's H200 NVL, based on the Hopper architecture, offers a 1.5 times increase in memory capacity and a 1.2 times increase in bandwidth compared to H100 NVL, enhancing performance for large language model fine-tuning [1] Group 2 - Google’s TPU is considered the only AI accelerator that can compete with NVIDIA's GPUs, leveraging frameworks like TensorFlow and OpenXLA to build a comprehensive AI ecosystem [2] - Google is increasing its capital expenditure to meet strong demand for AI infrastructure, with a projected Capex of approximately $91-93 billion for 2025 and significant increases expected in 2026 [2] - Google has established a leading position in the industry with top-tier capabilities in reasoning, multimodal abilities, agent tool usage, multilingual performance, and long context handling [2] Group 3 - Zhongji Xuchuang is a main supplier of optical modules for Google, with products like silicon photonics and 1.6T already in mass production, and a 3.2T product currently under development [3] - TeraHop, a subsidiary of Zhongji Xuchuang, has launched the first silicon photonics-based 64x64 OCS switch, which reduces power consumption for AI clusters and aids in network architecture [3] - Dahong Technology has developed spatial intelligence technology similar to Google's nano banana technology, utilizing optimized Gaussian splashing techniques for 3D modeling from multi-angle images [3]

Core Viewpoint - The prevailing concern regarding the depreciation of GPUs in the AI industry is largely unfounded, as the actual depreciation cycle is more favorable than many investors believe [1][2]. GPU Depreciation and Lifespan - Analysts suggest that the profit cycle for GPUs is approximately 6 years, and the depreciation accounting practices of major cloud computing firms are deemed reasonable [2]. - The cost of operating GPUs in AI data centers is significantly lower compared to the GPU rental market, allowing for a high marginal contribution rate when extending the lifespan of older GPUs [3]. - GPUs can have a practical lifespan of 7 to 8 years, with many companies still using GPUs that are over 5 years old and generating substantial profits [5]. Lifecycle Transition of GPUs - GPUs transition from high-performance tasks, such as training advanced AI models, to lower-demand inference workloads, allowing older GPUs to remain in active service [6]. - The variety of AI workloads enables older GPUs to be repurposed effectively, maintaining their profitability [6]. Cost Considerations - AI cloud computing companies often choose GPUs based on user expectations and budget, with older GPUs being utilized for lower-tier services while newer models are reserved for premium offerings [7]. - Many AI services can run on open-source models that require less computational power, further enhancing the utility of older GPUs [8]. Economic Advantages of Older GPUs - Despite higher energy consumption, older GPUs are often preferred due to their lower procurement costs, making them more cost-effective overall [10].

Core Viewpoint - The article discusses the debate surrounding the economic lifespan of GPUs, which is crucial for understanding the profitability of major tech companies and the validity of current AI valuations. Bernstein's report suggests a depreciation period of 6 years for GPUs, arguing that this is economically reasonable, while critics like Michael Burry claim the actual lifespan is only 2-3 years, warning of potential accounting manipulation to inflate profits [1][11]. Group 1: Economic Viability of GPU Depreciation - Bernstein analysts argue that a 6-year depreciation period for GPUs is justified, as the cash costs of operating older GPUs are significantly lower than their rental prices [2][4]. - The report highlights that even 5-year-old NVIDIA A100 chips can still yield "comfortable profits," indicating that the depreciation policies of major cloud service providers are fair and not merely for financial embellishment [2][4]. - The analysis shows that the contribution profit margin for A100 chips can reach up to 70%, with operational costs being substantially lower than rental income, providing strong economic incentives for extending GPU usage [4][5]. Group 2: Market Demand and Old GPUs - The current market environment supports the value of older GPUs, as there is overwhelming demand for computing power, with AI labs willing to pay for any available capacity, even for outdated models [6][7]. - Industry analysts note that the A100's computing capacity remains nearly fully booked, suggesting that as long as demand stays strong, older hardware will continue to hold value [8]. Group 3: Depreciation Policies of Tech Giants - Google has a depreciation period of six years for its servers and network equipment, while Microsoft ranges from two to six years, and Meta plans to extend some assets to 5.5 years starting January 2025 [9][10]. - Notably, Amazon has reduced the expected lifespan of some servers and network equipment from six years to five years, reflecting differing views within the industry on hardware iteration speed [10]. Group 4: Criticism and Concerns - Michael Burry warns that tech giants are artificially inflating profits by extending the effective lifespan of assets, predicting that this accounting practice could lead to a profit inflation of $176 billion from 2026 to 2028 [11][12]. - Burry specifically points out that companies like Oracle and Meta could see their profits overstated by 26.9% and 20.8%, respectively, due to these practices [12]. - Previous warnings from Bank of America and Morgan Stanley indicate that the market may be underestimating the true scale of AI investments and the potential surge in future depreciation costs, which could reveal a lower actual profitability for tech giants than expected [14][15].

Core Insights - Major companies plan to invest $1 trillion in AI data centers over the next five years, with a focus on depreciation as a key financial consideration [2] - The lifespan of AI GPUs is uncertain, with companies like Google, Oracle, and Microsoft estimating a maximum lifespan of six years, but potentially shorter [2][4] - Investors are concerned about the depreciation period, as longer asset lifespans lead to smaller impacts on profits [2] Depreciation Challenges - AI GPUs are relatively new, with NVIDIA's first AI-specific processor launched around 2018, and the current AI boom starting in late 2022 [4] - NVIDIA's data center revenue surged from $15 billion to $115 billion in the fiscal year ending January 2023 [4] - There is no historical reference for the lifespan of GPUs, making it difficult for companies to estimate depreciation accurately [4][5] Market Reactions - CoreWeave has set a six-year depreciation cycle for GPUs, indicating a data-driven approach to asset valuation [4][5] - Despite high demand for NVIDIA's A100 and H100 chips, CoreWeave's stock fell 16% after earnings guidance was affected by third-party data center developer delays [5][6] - The stock of Oracle has also dropped 34% since reaching a historical high in September [6] Skepticism in the Market - Short-seller Michael Burry has expressed doubts about the longevity of AI chips, suggesting that companies may be overstating their lifespan and underestimating depreciation costs [6] - Burry believes that the actual lifespan of server equipment is around two to three years, which could inflate reported earnings [6] Technological Advancements - AI chips may depreciate within six years due to wear and tear or obsolescence from newer models [8] - NVIDIA's CEO has indicated that older chip models will lose significant value as new models are released [8] - Amazon has shortened the expected lifespan of some servers from six years to five years due to rapid technological advancements [8][9] Strategic Procurement - Microsoft is diversifying its AI chip procurement to avoid over-investment in any single generation of processors [9] - The rapid iteration of technology in the AI sector complicates depreciation estimates, requiring careful financial forecasting [9]

Core Viewpoint - The depreciation period of GPUs is a critical issue affecting corporate profits and investment returns, especially as major tech companies plan to invest $1 trillion in AI data centers over the next five years [3][5]. Depreciation Challenges - The actual lifespan of GPUs is under scrutiny, with estimates ranging from two to six years, leading to concerns about inflated earnings by companies like Microsoft, Google, and Oracle [3][6]. - The lack of historical data on GPU usage complicates depreciation assessments, making it difficult for investors and lenders to gauge the value of these assets [5][6]. Market Reactions - Concerns about AI spending have already impacted stock prices, with CoreWeave's shares dropping 57% from their June peak and Oracle's stock falling 34% from its September high last year [3]. - CoreWeave has adopted a six-year depreciation cycle for its infrastructure, but its stock fell 16% following earnings reports due to delays from third-party data center developers [6][3]. Technological Impact - Rapid technological advancements are pressuring the depreciation of AI chips, with new models being released annually, which may render older models obsolete more quickly [7][8]. - Companies like Amazon have shortened the expected lifespan of some servers from six years to five years due to the accelerated pace of technological development in AI and machine learning [7]. Corporate Strategies - Microsoft is diversifying its AI chip procurement to avoid over-investment in any single generation of processors, acknowledging the rapid pace of innovation [8][9]. - Depreciation estimates are influenced by various factors, including technological obsolescence and maintenance, requiring companies to justify their assumptions to auditors [9].

Core Insights - The article discusses the significant financial implications of determining the depreciation period for GPUs as major tech companies plan to invest $1 trillion in AI data centers over the next five years [1] - The depreciation period directly affects financial performance, with longer periods allowing companies to spread costs over more years, thus reducing profit impact [1][4] - Concerns about AI spending are reflected in stock price declines for companies like CoreWeave and Oracle, indicating investor skepticism about over-investment in AI [1] Depreciation Challenges - Estimating GPU depreciation is complicated due to a lack of historical usage data, as the first AI processors from NVIDIA were launched around 2018, and the current AI boom began in late 2022 [4] - CoreWeave has adopted a six-year depreciation cycle for its infrastructure, while its CEO emphasizes a data-driven approach to assess GPU lifespan [5] - Market opinions vary, with some suggesting actual GPU lifespan may be as short as two to three years, leading to concerns about inflated earnings projections by major tech firms [5] Technological Pressure - The rapid pace of technological advancement is a key factor in GPU depreciation, with new models potentially rendering older ones obsolete within a short timeframe [6][7] - NVIDIA has shifted to an annual release cycle for new AI chips, increasing the risk of older models losing value quickly [7] - Amazon has reduced the estimated lifespan of some servers from six years to five due to accelerated technological development in AI and machine learning [7] Strategic Responses from Tech Giants - Microsoft is diversifying its AI chip procurement strategy to avoid over-investment in any single generation of processors, learning from NVIDIA's rapid product cycles [8] - Depreciation estimates in fast-evolving industries like technology require careful consideration of various factors, including technological obsolescence and historical lifespan data [8]