OpenAI对英伟达最新的一些人工智能芯片并不满意，并且自去年以来一直在寻找替代方案，这可能会使这两家AI热潮中最受关注的公司之间的关系变得 更加复杂。 据媒体援引多位知情人士表示，OpenAI对英伟达最新的一些人工智能芯片并不满意，并且自去年以来一直在寻找替代方案，这可能会使这两家 AI热潮中最受关注的公司之间的关系变得更加复杂。 OpenAI这一战略转变，源于其对用于执行AI推理中特定环节芯片的重视程度不断提高。所谓推理，是指像支撑ChatGPT应用的AI模型在响应用 户问题和请求时所进行的计算过程。英伟达在训练大型AI模型所需的芯片领域仍占据主导地位，而推理正成为竞争的新战场。 分析称，OpenAI及其他公司决定在推理芯片市场寻找替代方案，标志着对英伟达AI主导地位的一次重大考验。 周一，英伟达收跌近2.9%。 当前，OpenAI和英伟达这两家公司仍在进行投资谈判： 去年9月，英伟达表示，计划向OpenAI投入高达1000亿美元，作为一项交易的一部分。该交易将使英伟达获得这家初创公司的股份，同时为OpenAI提 供购买先进芯片所需的资金。 在此期间，OpenAI已与AMD等公司达成协议，采购可与英伟达竞 ...

Core Insights - 2026 is projected to be a pivotal year where AI inference workloads may surpass training workloads, with inference expected to account for the majority of AI capital expenditures by 2030, potentially reaching 75% of the estimated $1.2 trillion [1][4]. Group 1: AI Capital Expenditure and Market Performance - Despite concerns regarding funding, valuations, and interest rate fluctuations, the continuous growth in AI capital expenditures is driving strong performance in the semiconductor sector [1]. - The Philadelphia Semiconductor Index (SOX) has risen approximately 13% year-to-date, marking the second-best January performance in the past 20 years, significantly outperforming the S&P 500's 1% increase [1]. - The recent surge in semiconductor stocks is primarily led by storage chip manufacturers, semiconductor equipment suppliers, and analog chip producers, rather than major players like NVIDIA and Broadcom [1]. Group 2: Differentiated Opportunities for Chip Suppliers - The shift towards inference workloads will create differentiated opportunities for various types of chip suppliers, including GPUs, CPUs, and ASICs, while also impacting storage and semiconductor equipment suppliers [3]. - NVIDIA maintains a leading position with a comprehensive product lineup across both training and inference domains, supported by supply assurance advantages [5]. - AMD is viewed as a reliable second supplier of general-purpose chips, with recent stock price fluctuations attributed to market concerns over TSMC's 2nm process, which is still on track according to analysts [5]. Group 3: Optical Connectivity and Market Dynamics - Demand for optical connectivity is real, but recent price increases may be excessive; optical transceiver and component suppliers are among the strongest performers after storage chips [6]. - The necessity for optical connections is underscored by the expanding scale and bandwidth requirements of AI clusters, with NVIDIA's upcoming photonic switch expected to act as a potential catalyst [6]. - However, evidence of interest in co-packaged optics (CPO) from major cloud service providers is limited, primarily due to operational complexities and control over the bill of materials shifting to NVIDIA and Broadcom [6].

2025年，全球大模型token消耗量涨了100倍。每一笔消耗都意味着一次AI推理，而每一次推理的成本，正在成为AI公司能否盈 利的关键。 根据德勤报告，到2026年，推理算力在整体AI计算中的占比将超过训练，达到66%。大模型从"被训练出来"走向"被用起来"，推 理从技术配角变成了商业主力。 "训练市场是头部玩家的游戏，门槛越来越高，收敛得很快。"曦望董事长徐冰在采访中向21世纪经济报道记者表示，"但推理是 百花齐放的，需求看不到天花板。" 日前，曦望发布了公司新一代推理GPU——启望 S3。这家公司脱胎于商汤科技大芯片部门的公司，于2025年初独立运营，一年 内完成近30亿元战略融资，股东阵容兼具产业龙头与国资背景机构。 曦望选择了一条看似窄众的道路：All in推理，放弃训练。这在GPU公司竞相标榜"训推一体"或"算力领先"的语境里，像是一次 主动的战略收缩，而管理层认为这是聚焦。 国产AI芯片赛道正在进入一个更务实、更分化的新阶段。这背后，既有对市场趋势的预判，也有在现有技术、生态和供应链约 束下的务实考量。 "大模型的训练需要万卡甚至十万卡的大规模集群，成本极高，是少数巨头的游戏。而用为训练优化的昂 ...

另有AI芯片产业链人士也对《科创板日报》记者表示，未来，推理请求量与并发数将大幅增加，推理算力需求呈指数级攀升。 推理算力需求呈指数级攀升 "现阶段正处于以智能体为代表的人工智能新应用爆发初期，未来推理请求量与并发数大幅增加，推理算力需求呈指数级攀升。"一名AI产业链人士向《科 创板日报》记者表示，"随着推理范式变化，AI智能体将加速落地，其整体算力消耗可达同参数规模大语言模型的10倍以上，对智能算力的需求呈数量级 跃迁。 至2030年，预计AI推理在整个AI计算市场将占到80%的份额，而聚焦于极致推理的AI芯片，未来会有更强的爆发性。 随着大模型行业逐渐从大规模训练阶段走向推理落地阶段，业内分析普遍认为，2026年全球AI推理的需求将超过AI训练场景。 不久前，英伟达以200亿美元（约合1400亿元）收购一家AI推理芯片初创企业Groq的技术授权，并把Groq核心团队招入囊中，来补全推理算力拼图。 在近日的采访中，曦望董事长徐冰向《科创板日报》记者判断称，至2030年，预计AI推理在整个AI计算市场将占到80%的份额，而聚焦于极致推理的AI芯 片，未来会有更强的爆发性，并会对现有的算力系统造成冲击。 从市场 ...

智通财经APP获悉，东吴证券发布研报称，AIAgent落地速度正逐渐加速，其带来的沙箱创建、负载卸 载等需求有望大幅拉动CPU的需求。但同时，全球算力供应链产能紧张，多个环节涨价，使得CPU制造 成本增加。两方面影响下，该行认为CPU产业有望迎来大周期。 东吴证券主要观点如下： 风险提示：Agent落地进展不及预期；产能供应问题缓解。 DRAM生产转向HBM，消耗更多晶圆；与NAND需求攀升、交货期延长，库存告急，挤占CPU晶圆材 料供给。CPU部件PCB应用及加工材质的转变，使得钻针使用寿命缩短，消耗量暴增；CCL采用的树脂 体系、玻纤布与铜箔匹配复杂，新进入者良率提升缓慢与客户认证周期长，导致有效产能释放缓慢；二 者纷纷涨价，带动CPU价格上涨。 AI推理和Agent发展迅速，拉动高端多核CPU需求 CPU负载正从"人类节奏"转向"机器节奏"。Agentic AI远不是单次推理，而是动态推理+多步决策+外部 工具调用的循环，这比传统大模型调用更耗资源、负载更复杂、成本更高。这种资源调用增长，加之为 了安全防范而产生的高频沙箱隔离开销，使得CPU资源消耗呈现指数级放大。Deepseek提出Engram模 块 ...

Group 1 - The core viewpoint of the article highlights that SanDisk's financial performance exceeded expectations, with total revenue reaching $3 billion, a quarter-over-quarter increase of 31%, and a gross margin of 51.1%, up 21 percentage points from the previous quarter [1] - The data center business revenue surged by 64% quarter-over-quarter, accounting for 15% of total revenue, with management indicating that they will complete more certifications for large-scale cloud service providers in the upcoming quarters [1] - For FY3Q26, the company guides a midpoint revenue of $4.6 billion and earnings per share of $13, aligning with market expectations that have been revised upwards to $11-13 [1] Group 2 - The article discusses four underlying logic points regarding NAND demand, including the concept of "using storage to compute," particularly with KV Cache persistence, which significantly reduces computational power consumption during the prefill phase [3] - The shift in data generation from human production to self-generation by models, which is not limited by time, attention, or physical boundaries, is noted, with SanDisk stating that "data growth is accelerating as the temperature of data is rising" [4][3] - The increase in the value of data reuse is emphasized, with historical storage rates previously at only 1%, now significantly enhanced by LLM/RAG, leading to a substantial increase in storage rates [8] - The inflation of data under the same semantic density is highlighted, with the transformation of plaintext into embeddings and KV, resulting in capacity expansion by 5-1000 times, driven by AI workloads and increased NAND content requirements [8]

Core Viewpoint - The article highlights the advancements and market position of Alibaba's chip business, particularly the "Zhenwu" PPU chip, which has surpassed competitors in China's GPU market and is gaining traction in various applications, including AI and autonomous driving [2][4]. Group 1: Product Performance and Features - Alibaba's "Zhenwu" PPU chip has achieved a shipment volume of several hundred thousand units, surpassing competitors like Cambricon and establishing itself as a leader among domestic GPU manufacturers [2]. - The "Zhenwu" PPU chip features a self-developed parallel computing architecture and inter-chip communication technology, with a memory capacity of 96G HBM2e and an inter-chip bandwidth of 700 GB/s, making it suitable for AI training, inference, and autonomous driving applications [4]. - The chip has been deployed in large-scale for training and inference of the Qianwen large model, optimized with Alibaba Cloud's complete AI software stack, serving over 400 clients including State Grid of China, Chinese Academy of Sciences, XPeng Motors, and Sina Weibo [4]. Group 2: Market Reception and Competitive Edge - Industry insiders report that the overall performance of the "Zhenwu" PPU exceeds that of NVIDIA's A800 and is comparable to NVIDIA's H20, indicating a strong competitive position [4]. - The "Zhenwu" PPU is noted for its excellent stability and cost-effectiveness, receiving positive feedback in the industry, with a market showing signs of supply exceeding demand [4]. - The company, PingTouGe, was established in September 2018 as Alibaba's wholly-owned semiconductor chip business to advance its integrated cloud chip strategy [4].

Core Insights - Western Digital's gross margin surged to 46.1%, with incremental gross margin expectations reaching 75% due to price increases and cost reductions [2][4][35] - CEO Irving Tan announced that 2026 production capacity is sold out, with long-term agreements signed with three of the top five customers extending to 2027-2028 [4][10][12] - The demand for HDDs is expected to grow structurally driven by AI inference applications, which generate vast amounts of new data that require low-cost storage [2][18][20] Financial Performance - In Q2 of FY2026, Western Digital reported revenue of $3.02 billion and adjusted EPS of $2.13, both exceeding market expectations [3][4] - Net profit for the quarter reached $1.84 billion, or $4.73 per share, marking a 210% increase from $594 million ($1.27 per share) in the same quarter last year [4][35] - The company delivered 215 exabytes of data, a 22% year-over-year increase, including over 3.5 million units of the latest generation ePMR drives [33][34] Margin and Cost Dynamics - The gross margin improvement reflects a shift towards high-capacity drives and strict cost control across manufacturing and supply chains [35] - CFO Kris Sennesael confirmed that the incremental margin is around 75%, driven by a 2-3% increase in average selling price per terabyte and a 10% year-over-year decrease in manufacturing costs per terabyte [7][8][43] Long-term Agreements and Customer Relationships - Western Digital has secured all firm purchase orders for 2026 from its top seven customers and signed long-term agreements with three of the top five customers, indicating strong customer trust and recognition of value [10][12][46] - The long-term agreements include both pricing and quantity terms, reflecting a strategic approach to managing customer relationships in a tight supply environment [13][63] Technology and Market Trends - The company is accelerating the customer validation timeline for HAMR (Heat-Assisted Magnetic Recording) technology by six months due to supply pressures [17][56] - The transition from AI model training to inference applications is expected to create significant storage demand, benefiting HDDs as data centers return large amounts of inference data to HDDs [18][20][59] Future Outlook - For Q3 FY2026, Western Digital expects revenue of $3.2 billion, reflecting approximately 40% year-over-year growth, with gross margin projected between 47% and 48% [38] - The company continues to focus on supporting customer needs for exabyte-scale storage while completing the certification and release of next-generation ePMR and HAMR drives [30][38]

智通财经APP获悉，西部数据(WDC.US)召开2026财年第二季度财报电话会。CEO Irving Tan透露2026日 历年的产能基本已售罄，已与前七大客户签订了确定采购订单。公司还与其中两家签订了覆盖2027年的 长期协议（LTA），与另一家签订了覆盖2028年的协议。Irving Tan强调AI推理应用将驱动HDD需求结 构性增长，推理过程生成海量的新数据，这些数据需要被低成本地存储。 Q&A 问答 主持人： 下午好，感谢您的耐心等待。欢迎参加西部数据2026财年第二季度电话会议。目前，所有参会者都处于 仅听模式。稍后，我们将进行问答环节。（操作员说明）提醒一下，本次通话正在录音。现在，我将把 会议交给投资者关系副总裁Ambrish Srivastava先生。您可以开始了。 Ambrish Srivastava，投资者关系副总裁： 谢谢，大家下午好。今天与我一同参会的有西部数据首席执行官Irving Tan和首席财务官Kris Sennesael。 在开始之前，请注意今天的讨论将包含基于管理层当前假设和预期的前瞻性陈述，这些陈述受到各种风 险和不确定性的影响。这些前瞻性陈述包括对我们产品组合、业务计划 ...

1月29日，西部数据发布了2026财年第二季度财报，当季营收达到30.2亿美元，调整后每股收益2.13美元，双双超市场预期。 财报显示，西部数据第二财季净利润达到18.4亿美元，约合每股4.73美元，较去年同期的5.94亿美元（每股1.27美元）实现了210%增长。 尽管财务数据亮眼，但盘后股价下跌近3%。分析认为西部数据股价在1月份已大幅上涨超60%，利好出尽与获利回吐导致部分投资者选择在亮眼 财报公布后"卖出事实"以锁定利润。 在随后的财报电话会上，管理层表态关于"产能天花板"和"定价权"的极限。CEO Irving Tan直接指出：2026年产能已全部售罄，甚至有客户签下了 到2028年的供应长约。 利润狂飙，CFO确认75%增量毛利 财报显示公司Q2毛利率同比暴增770个基点至46.1%，且Q3指引进一步看高至48%。 富国银行资深分析师Aaron Rakers在现场直接"算账"： 根据47%-48%的指引，你们的增量毛利率（Incremental Margin）似乎维持在70%甚至75%的高位，这种暴利能持续吗？ CFO Kris Sennesael确认了分析师的测算： Aaron，你的数学很好。增 ...