一系列股票抛售及会计违规指控，使英伟达(NVDA.O)成为关于AI及相关股票价值的讨论焦点。针对市场质疑，英伟达近日开始反击。天风国际证券分析师 郭明錤也通过他的分析，尝试为英伟达财务报表"正名"。 英伟达发送7页备忘录回应12项质疑 上周末，英伟达投资者关系团队向华尔街分析师私下发送了一份七页备忘录，直接回应了多位持怀疑态度投资者提出的十二项质疑。 备忘录采用公司标志性的绿色字体排版，开篇即回应了迈克尔·伯里（Michael Burry）上周在社交媒体上的一则帖子，该帖批评公司基于股票的薪酬稀释及 股票回购。伯里因在2008-2009年金融危机前精准押注次贷违约而闻名，其事迹被改编成电影《大空头》（The Big Short）。 "客户根据实际寿命和使用模式，将GPU折旧周期设定为4至6年。诸如2020年发布的A100等老款GPU仍可高效运作并产生可观利润，其经济价值 远超过部分评论者所称的2至3年。" 郭明錤：英伟达的财报结果完全符合财报常识和行业现状 英伟达在备忘录中指出："自2018年以来，公司回购股份总额为910亿美元，而非伯里所称的1125亿美元。伯里先生似乎错误地将受限股票单位（RSU）的税 款 ...

摘要 英伟达 Q3 财报显示收入同比增长 205%，调整后毛利率达 73.6%，预 计 Q4 收入中值为 650 亿美元，同比增长 65%，调整后毛利率中枢为 75%。公司预计 2026 财年毛利率维持在 70%中段，缓解了市场对其毛 利率受存储涨价和 ASIC 影响的担忧。 英伟达正加强供应链韧性，与台积电合作在美国本土生产 Blackwell 晶 圆，并扩大与富士康、安靠等合作伙伴的美国制造能力。安靠股价盘后 上涨 10%，反映市场积极情绪。 英伟达预计 Blackwell 和 Ruby 系列 2026 年合计出货 2000 万颗，目 标收入 500 亿美元，并可能上调。新增沙特采购协议约 100 亿美元， Anthropic 新增合作需求约 300-350 亿美元，使得原定 5,000 亿美元 目标有 10%的上调空间。 英伟达 CEO 黄仁勋回应 AI 泡沫及 GPU 折旧周期问题，认为云厂商给予 GPU 卡 5-6 年折旧周期合理，且随着推理比例增加，周期可延长。客户 测试替代方案后仍选择英伟达，表明其竞争力强劲。 Q&A 英伟达最新的财报表现如何？其未来展望如何？ 英伟达最新发布的 Q3 财报显示 ...

以下文章来源于傅里叶的猫 ，作者猫叔 傅里叶的猫 . 曾任芯片EDA大厂资深工程师，聊技术、聊产业、聊投资 H200放开的消息昨天已经传的沸沸扬扬了，国内的新闻基本都是这样写的： 但这个新闻最早是出自彭博，比路透要早2个多小时。 而彭博的新闻是下面这个写的，也就是说根据彭博的这个描述，目前只是初步讨论，而且完全有可能只 是停留在讨论，永远不会放开。 这事还得回溯到前段时间中美领导层见面，川普说会谈到Blackwell，大家都以为B30A会放开。后来的 事大家也都知道了，川普说没有谈Blackwell。 但又过了两天，WSJ上的消息说是因为川普的高级顾问们都反对，所以才没有谈，我们当时在星球中就 发过这个： 两国领导开会那天上午，有朋友就发我这样的截图： 所以可能高端的Hopper要放开的事也讨论了很久了。 说话正题，这次的说法是H200要放开，先看下H200的性能： | Specification | H100 | H200 | | --- | --- | --- | | GPU Architecture | Hopper | Hopper | | GPU Memory | 80 GB HBM3 | 141 ...

Core Viewpoint - Nvidia has confirmed a strong revenue outlook for its data center business, projecting over $500 billion in revenue for the fiscal year 2025/26, with potential for further upside [1][7]. Financial Performance - Nvidia reported third-quarter revenue of $57 billion, exceeding Wall Street's expectation of $55.4 billion. The fourth-quarter revenue guidance is set at $65 billion, also above market estimates of $62.4 billion [3]. - The company anticipates a recovery in gross margin to 75% in the fourth quarter, aligning with previously set management targets, despite rising costs for HBM memory and other components [3]. Earnings Forecast - Goldman Sachs has raised Nvidia's future earnings per share (EPS) expectations by an average of 12% for the coming years. The firm has also provided EPS forecasts for fiscal years 2028 to 2030, estimating $15.60, $18.65, and $22.10 respectively [4]. Key Issues Addressed - Nvidia's management confirmed the expectation of exceeding $500 billion in data center product demand for the fiscal year 2025/26, with ongoing customer orders suggesting further growth potential [7]. - The next-generation Rubin chip is scheduled for release in mid-2026, with significant revenue contributions expected in the latter half of the same year, alleviating market concerns regarding product roadmap execution [7]. - Management provided evidence of the GPU product lifecycle, noting that the Ampere architecture GPU (A100), launched six years ago, continues to operate under high loads, indicating exceptional durability and longevity beyond customer depreciation expectations [8]. Data Center Business Growth - Nvidia's data center computing business achieved $51.2 billion in revenue for the third quarter, marking a 56% year-over-year increase. The new Blackwell Ultra (GB300) series accounted for two-thirds of total shipments in the Blackwell series [9]. - The data center networking business saw a remarkable 162% year-over-year growth, reaching $8.2 billion, driven by strong demand for NVLink, Spectrum-X, and Infiniband solutions, with significant contributions from major clients like Meta, Microsoft, Oracle, and xAI [10]. - Looking ahead, Nvidia maintains its long-term outlook for the AI infrastructure market, predicting global annual spending to reach $3-4 trillion by 2030, and aims to secure a significant share of this expansive market [10].

Summary of Key Points from Conference Call Industry Overview - The conference call primarily discusses the cloud computing industry, focusing on GPU depreciation policies and their financial implications for major cloud service providers such as Amazon, Meta, Microsoft, and Google [1][2][4]. Core Insights and Arguments - **Depreciation Impact on Profit**: Amazon's depreciation adjustments in 2024 added approximately $600-700 million in quarterly profits, but a subsequent adjustment in 2025 is expected to reduce net profits by $100-300 million, highlighting the significant impact of depreciation policies on financial performance [1][2]. - **Future Profit Growth**: A projected change in accounting standards in 2025 is anticipated to result in a net profit increase of about 5.6%, with cumulative net profit additions of $300-500 billion from 2023 to 2028, totaling $1.46 trillion [1][4]. - **GPU Rental Market Dynamics**: Despite a decline in rental prices for older GPUs like the H100 and A100 (down 25% and 30% respectively since September 2024), there remains a strong market for these older models, with some being rented at 95% of their original price [1][3][11]. - **Product Iteration Speed**: NVIDIA has accelerated its product release cycle from every 2-3 years to annually, with new chips like the Blackwell offering significantly improved performance and cost efficiency, which is expected to drive faster market updates [1][7]. - **GPU Lifespan Concerns**: High utilization rates in data centers are shortening GPU lifespans to 1-3 years, raising concerns about equipment wear and tear as demand and supply dynamics shift [1][10]. Additional Important Insights - **Differing Depreciation Policies**: Major cloud providers have adjusted their GPU depreciation periods, with Amazon changing its policy from 6 years in 2024 to 4 years in 2025, while others like Microsoft and Google have extended theirs to 6 years [2]. - **Market Trends in GPU Pricing**: The rental and second-hand market for GPUs is experiencing rapid depreciation, with significant price drops observed since 2019, indicating a need for companies to adapt to changing market conditions [3][11]. - **Emerging Cloud Providers**: New cloud service providers like NeoCloud are facing more volatile pricing and higher operational costs compared to larger firms, which may impact their competitiveness and financial stability [14][15][17]. - **Financial Implications for New Providers**: Companies like Corwave adjusting their depreciation periods are expected to see substantial savings in depreciation costs, significantly affecting their operating and net profits [18]. Conclusion - The conference call highlights the critical role of GPU depreciation policies in shaping the financial landscape of the cloud computing industry, with implications for both established and emerging players. The rapid pace of technological advancement and market dynamics necessitate ongoing adjustments in strategy to maintain competitiveness and profitability [1][19].

Summary of Key Points from the Conference Call Industry Overview - The discussion centers around the **GPU (Graphics Processing Unit)** market, particularly in the context of **AI (Artificial Intelligence)** applications and data center operations [2][10]. Core Insights and Arguments 1. **GPU Lifespan and Depreciation**: - GPUs can profitably operate for approximately **6 years**, and the depreciation accounting used by major hyperscalars is deemed reasonable [3][11]. - Cash costs for operating GPUs are significantly lower than market rental prices, leading to high contribution margins for older GPUs [3][11]. - Older models, such as the **A100**, can still yield comfortable margins even after **5 years** of use, suggesting a **5-6 year depreciation lifespan** is justifiable [3][11]. 2. **Market Dynamics**: - There is a concern that if compute demand softens, older GPUs may be decommissioned despite being functional, but this would be a broader issue beyond just depreciation accounting [3][11]. - Data center operators often face "burn-in" issues, where older configurations may not be optimal for newer hardware, leading to operational inefficiencies [3][13]. 3. **Contractual Implications**: - Long-term contracts can shift the economic burden of GPU depreciation to end-users, as seen in examples like **OpenAI** signing a **5-year contract** for **Coreweave H100 capacity** [4][15]. - This indicates that even if GPUs depreciate faster than expected, the costs may be absorbed by customers through higher prices [4][15]. 4. **Pricing Trends**: - Unlike memory and storage, accelerated compute does not behave as a commodity; older GPUs often command higher prices than expected based on performance metrics [5][16]. - This suggests that legacy workloads are still prevalent, and cloud vendors may charge a premium for these services [5][16]. 5. **Investment Implications**: - **NVIDIA (NVDA)** is rated **Outperform** with a target price of **$225**, highlighting the significant datacenter opportunity [8]. - **AMD** is rated **Market-Perform** with a target price of **$200**, driven by high AI expectations and potential growth from new deals [8]. - **Broadcom (AVGO)** is also rated **Outperform** with a target price of **$400**, supported by strong margins and cash flow [8]. - Companies pivoting into AI datacenter assets, such as **Iren**, **Riot**, **Corz**, and **Clsk**, are noted for their re-rating potential [9]. Additional Important Observations - The depreciation of GPUs may not follow a linear model, as they tend to lose more value in the first year but retain value better afterward [3][13]. - The overwhelming demand for compute resources means that even older, less efficient hardware remains in use, countering concerns about the need for immediate replacements [6][17]. - The report emphasizes that the assumptions regarding GPU lifespans and depreciation are more favorable than some market participants fear [6][17]. Conclusion - The analysis indicates a robust market for GPUs in AI applications, with significant implications for investment strategies in related companies. The dynamics of depreciation, contractual obligations, and market pricing trends are critical for understanding the future landscape of the GPU industry.

智通财经APP获悉，英伟达(NVDA.US)将在11月19日盘后公布2026财年第三季度财报，预计其盈利将再次超出预期，调整后每股收益预计为1.26美元；市场 还预计该公司本季度营收为营收为552.8亿美元，较去年同期增长超过55%。 过去一年，英伟达的营收增长受到开发生成式AI模型所需芯片的强劲需求推动。英伟达主导着生成式AI芯片市场，这些芯片已被证明在多个行业有用，包 括营销、广告、客户服务、教育、内容创作、医疗保健、汽车、能源与公用事业以及视频游戏开发。 各行业工作流程现代化的需求日益增长，预计将推动对生成式AI应用的需求。根据财富商业洞察的最新报告，全球生成式AI市场规模预计到2032年将达到 9676.5亿美元。预计该市场在2024年至2032年期间的复合年增长率为39.6%。 生成式AI的复杂性需要广泛的知识和巨大的计算能力。这意味着企业将需要显著升级其网络基础设施。英伟达的AI芯片，包括A100、H100、B100、B200、 B300、GB200和GB300，是构建和运行这些强大AI应用的首选，使该公司成为该领域的领导者。随着生成式AI革命的展开，预计英伟达的先进芯片将推动 其营收和市场地位大 ...

Key Takeaways NVIDIA expects Q3 revenues of $54 billion, and the consensus sees 55.6% growth from the year-ago period.Data Center, Gaming, Professional Visualization and Automotive are all projected to post strong gains.Positive Earnings ESP and solid segment trends support expectations for another quarterly beat.NVIDIA Corporation (NVDA) is likely to beat on earnings when it reports third-quarter fiscal 2026 results on Nov. 19, after market close.The company expects revenues of $54 billion (+/-2%) for the ...

Core Viewpoint - OpenAI has emerged as a leading player in the AI sector, heavily investing in data centers and GPU acquisitions, but faces significant challenges due to electricity shortages and inefficiencies in energy usage [5][11][12]. Group 1: AI and Power Consumption - The total electricity consumption of data centers in the U.S. reached 176 terawatt-hours (TWh) in 2023, accounting for 4.4% of the national electricity generation, with projections to double by 2028 [11]. - The average Power Usage Effectiveness (PUE) globally in 2024 is expected to be 1.56, indicating that only two-thirds of electricity is used for GPU computing, while the rest is wasted on cooling and other systems [15]. - The inefficiency of AI systems is highlighted, as they consume significant power while having low utilization rates, exacerbating the electricity crisis [10][12]. Group 2: Challenges in the U.S. Energy System - The aging U.S. power infrastructure is struggling to meet the increasing demand from AI technologies, leading to rising electricity costs for consumers [12][13]. - The shift towards nuclear power and the reduction of renewable energy projects have further complicated the energy landscape, making it difficult to sustain the growing needs of AI companies [16][17]. Group 3: Future of AI Chips - Current AI chips like the H100 and A100 are becoming outdated, with newer models (H200, B200, B300) expected to dominate the market by 2025, potentially rendering older chips obsolete if they remain unused due to power shortages [20][22]. - The stock prices of AI companies are closely tied to their GPU availability, and any delays in utilizing these chips could negatively impact their market valuations [22][24]. Group 4: Strategies for Energy Supply - Companies are exploring various strategies to secure energy, including building new power plants and relocating data centers to countries with more favorable energy conditions, although this presents its own set of challenges [25][27]. - Some companies are even considering space-based data centers powered by solar energy, although this concept is still in experimental stages and poses numerous technical challenges [28][31]. Group 5: Comparison with China - In contrast to the U.S., China's data center electricity consumption is significantly lower at 166 TWh, representing about 2% of total social electricity use, while also focusing on green energy initiatives [33][34]. - The emphasis on sustainable energy practices in China suggests a more stable environment for AI development compared to the energy crisis faced in the U.S. [34][36].

Core Insights - OpenAI has been aggressively investing in AI infrastructure, including a $300 billion partnership with Oracle for data centers and a $100 billion chip purchase from NVIDIA, amidst a growing AI bubble driven by GPU sales [1][3] - Microsoft CEO Satya Nadella highlighted a critical issue: the lack of electricity is hindering AI development, despite the abundance of chips [3][5] Energy Consumption and Efficiency - In 2023, U.S. data centers consumed 176 terawatt-hours (TWh) of electricity, accounting for 4.4% of the national total, with projections to double by 2028 [5][8] - The average Power Usage Effectiveness (PUE) globally in 2024 is 1.56, indicating that only two-thirds of electricity is used for GPU computing, while one-third is wasted on cooling, power systems, and lighting [7][8] Challenges in Power Supply - The aging U.S. power grid is struggling to meet demand, leading to increased electricity costs for consumers, which has risen significantly from 2021 to 2022 [8][10] - The shift in energy policy under the Trump administration, including cuts to renewable energy projects, has exacerbated the situation, making it difficult for tech companies to secure sufficient power for their operations [10][12] Chip Lifecycle and Market Dynamics - Current AI chips like the H100 and A100, released in 2022, may soon be outdated as newer models (H200, B200, B300) are set to dominate the market by 2025, potentially rendering existing inventory obsolete [12][14] - The valuation of AI companies is closely tied to GPU availability and demand, meaning that unutilized chips could negatively impact stock prices [14][16] Strategies for Mitigation - Companies are exploring options to build new power plants, such as OpenAI and Oracle's joint natural gas facility in Texas, but face challenges including supply shortages for necessary equipment [16][18] - Some firms are considering relocating data centers to countries with less developed power infrastructure, which could further strain local resources [18][19] Global Comparison - In contrast to the U.S., China's data centers consumed 166 TWh in 2024, representing about 2% of total electricity usage, with a focus on green energy and carbon reduction [22][24] - The future of high-tech companies may hinge less on chip quantity and more on their ability to secure reliable electricity supply for their operations [24]