我们为谷歌的成就感到欣喜，他们在AI领域取得长足进步，而我们 始终是谷歌可靠的供应伙伴 。 英伟达持续引领行业发展，我们 是唯 一能够兼 容所有AI模型、覆盖所有计算场景的硬件平台 。 西风 发自 凹非寺 量子位 | 公众号 QbitAI 一听到谷歌要抢走10%的年收入，英伟达罕见地慌了。 紧急发声，一边客套恭喜谷歌的AI进展，一边明晃晃强调自己的核心地位： 相较于专为特定AI框架或功能设计的专用芯片，英伟达解决方案具备 更卓越的性能、更广泛的适用性以及更灵活的通用性 。 到底是咋回事儿？ 众所周知，谷歌刚凭Nano Banana Pro、Gemini 3大获全胜，并且 Gemini 3全程使用谷歌自研TPU （而非英伟达GPU） 。 随后谷歌背后更大动作就曝光了，如意算盘打到了抢占 AI 芯片算力 市场 的高地上： 消息称，谷歌 开始向Meta与大型金融机构等，推介在其自有数据中心部署TPU的方案 。 而Meta正在与谷歌探讨使用TPU训练新AI模型， 计划在2027年斥资数十亿美元在其数据中心使用TPU ，并于明年从谷歌云租用谷歌芯片。 △ AI生成 两大巨头交锋，直接搅动整个AI产业牌桌。 知情人透露， ...

在AI大模型产业化加速的当下，算力需求正从训练阶段向推理阶段倾斜。 在近期举办的第二十二届中国国际半导体博览会（IC China 2025）上，英特尔公司副总裁、英特尔中国 研究院院长宋继强指出，从2025年开始智能体AI的相关算力逐渐攀升，未来会渐渐超过传统训练基座 大模型和训练微调大模型的算力，成为驱动AI算力增长的核心动力。 《中国经营报》记者观察发现，国内芯片产业的人士大多数认为，尽管英伟达、AMD在通用AI计算市 场仍占据主导地位，但在推理端，尤其是面向视频生成、边缘计算和行业应用的推理场景，正成为中国 AI芯片企业实现差异化突围的赛道。 国产化替代与全球化竞争并行 徐昀举例指出，当前国内AI芯片主流采用12纳米与7纳米先进工艺，而北美地区已逐步向2纳米工艺节 点推进；受工艺差距影响，国内单颗AI芯片的算力仅约为北美同类产品的30%，在内存容量、数据带宽 等关键性能指标上，也仅能达到北美芯片的40%至70%。 "2023年之前，训练与推理的算力需求占比约为6:4，但到了2025年，这一比例已实现反转。" 曦望 Sunrise商业产品负责人阎研表示，当前预训练大模型技术已接近成熟，高性能训练芯片的市场 ...

芯片制造商英伟达公布了2026财年第三季度570亿美元的营收，其中数据中心业务贡献最大，营收达510 亿美元，同比增长66%。 首席执行官黄仁勋表示，公司在人工智能工作负载方面持续增长，这些工作负载需要英伟达专业生产的 高性能图形处理器。 黄仁勋指出，由于预训练、后训练和推理能力的进步，AI推理正在指数级扩展。他说，随着AI系统现 在能够在生成答案之前"阅读、思考和推理"，推理变得越来越复杂。黄仁勋声称，计算需求的指数级增 长正在推动对英伟达平台的需求。 除了数据中心AI加速器GPU之外，公司的NVLink AI网络基础设施业务增长了162%，营收达82亿美 元。 黄仁勋表示："客户对NVLink Fusion的兴趣持续增长。我们在10月份宣布了与富士通的战略合作，将通 过NVL Fusion整合富士通的CPU和英伟达GPU，连接我们庞大的生态系统。我们还宣布了与英特尔的合 作，开发多代定制数据中心和PC产品，使用NVLink连接英伟达和英特尔的生态系统。" 然而，中国市场现在实际上已经关闭。黄仁勋说："由于地缘政治问题和中国市场竞争日益激烈，本季 度大额采购订单从未实现。虽然我们对目前阻止我们向中国运送更具 ...

Financial Data and Key Metrics Changes - The meeting discussed the election of directors and the ratification of Ernst & Young as the independent registered public accounting firm for fiscal year 2026, indicating a stable governance structure [12][19]. - The preliminary voting results showed that all proposals received affirmative votes from a majority of Oracle's shares present, reflecting shareholder confidence [17][18][20]. Business Line Data and Key Metrics Changes - The company highlighted its focus on AI, particularly in AI reasoning, which is expected to become increasingly important for Oracle's business [22][24]. - Oracle's database services are projected to grow significantly due to the integration of AI capabilities and partnerships with major cloud providers [35][36]. Market Data and Key Metrics Changes - Oracle's AI offerings are broad and encompass various areas, including model training, inferencing, and embedded AI features in applications, which positions the company favorably in the competitive landscape [31][39]. Company Strategy and Development Direction - The company is actively embedding AI features into its applications, making it easier for customers to adopt these technologies without additional costs [37][40]. - Oracle's strategy includes leveraging its extensive database capabilities and AI data platform to enhance customer interactions and data utilization [25][26]. Management's Comments on Operating Environment and Future Outlook - Management expressed confidence in the growth of the AI inferencing business and its potential impact on Oracle's future [22][24]. - The executives emphasized the importance of private enterprise data for AI applications, which Oracle is uniquely positioned to manage [29][30]. Other Important Information - The meeting included a reminder for shareholders to review the most recent Form 10-K and Form 10-Q for discussions on risks that may affect future results [21]. Q&A Session Summary Question: When will AI inferencing become more material to Oracle's business? - Management indicated that AI reasoning is expected to take off as models become more capable, and Oracle is well-positioned due to its data management capabilities [22][24]. Question: Why is Oracle winning more AI business than competitors? - The differentiation stems from Oracle's historical decisions in technology and architecture, which have created a scalable and cost-effective AI offering [28][29]. Question: What is driving the expected 8X growth in Oracle's database? - The growth is attributed to the expansion of Oracle Database services into other cloud environments and the increasing demand for AI-integrated database solutions [33][35][36]. Question: How will Oracle succeed in getting customers to adopt AI? - Oracle is embedding AI features directly into its applications, allowing for seamless adoption and immediate value for customers [37][40].

Core Insights - The demand for AI infrastructure is expanding significantly as AI applications evolve, with a shift from centralized cloud architectures to edge computing for real-time AI processing [1][2][5] - Akamai and NVIDIA have launched the Akamai Inference Cloud, a distributed generative edge platform designed for low-latency, real-time AI processing globally [1][5] - The AI inference workload is expected to far exceed training workloads, necessitating a reevaluation of computational infrastructure to support real-time AI processing demands [2][3] Industry Trends - The AI industry is transitioning from model development to practical application, with AI applications evolving from simple request-response models to complex multi-step reasoning and real-time decision-making [2][3] - Edge computing is becoming essential for AI inference, moving away from its previous role as a support for centralized cloud services to a primary function that enhances user experience and operational efficiency [2][3] Market Potential - The global edge AI market is projected to exceed $140 billion by 2032, a significant increase from $19.1 billion in 2023, indicating explosive growth [4] - The edge computing market could reach $3.61 trillion by 2032, with a compound annual growth rate (CAGR) of 30.4% [4] Competitive Landscape - Major tech companies, including Google, Microsoft, and Amazon, are actively investing in edge computing, leveraging their technological strengths and large user bases [5][6] - Akamai has established a global platform with over 4,200 edge nodes, enhancing its capability to support AI inference services and improve competitiveness in overseas markets [6]

Core Insights - The event "Storage Power China Tour" in Beijing focused on the challenges and innovative paths of storage power in the AI inference era, highlighting the importance of advanced storage as a core support for AI technology implementation [1] - The AI industry has transitioned from model creation to practical application, with inference costs becoming a bottleneck for large-scale deployment, driven by the exponential growth of token usage in various sectors [3] - Technical innovation is essential for overcoming industry pain points, with storage architecture evolving from passive storage to intelligent collaboration, exemplified by Huawei's Unified Cache Management (UCM) technology [4] Industry Challenges - The AI industry's shift to practical applications has led to three main challenges: the explosion of multimodal data creating storage capacity pressures, the high performance demands on storage systems, and the high costs of advanced storage media [3] - Traditional storage architectures struggle to meet the requirements for high throughput, low latency, and heterogeneous data integration, hindering AI application development [3] Technological Innovations - The UCM technology developed by Huawei represents a significant advancement, enabling a three-tier cache architecture that dramatically reduces token latency by up to 90% and increases system throughput by 22 times [4] - UCM's open-source initiative aims to lower barriers for small and medium enterprises to access advanced inference acceleration capabilities and promote unified technical standards [4] Ecosystem Development - A collaborative effort involving Huawei, China Mobile, and Inspur has led to the establishment of the "Advanced Storage AI Inference Working Group," focusing on technology research, standard formulation, and ecosystem building [5] - The Chinese storage industry has a solid foundation, with total storage capacity reaching 1680 EB by June 2025, and advanced storage accounting for 28% of this capacity, nearing the targets set in national development plans [5][6] Future Outlook - Advanced storage is evolving into a central component of the AI intelligent computing system, addressing performance, cost, and efficiency bottlenecks, thus making AI technology more accessible to small and medium enterprises [7] - The ongoing technological advancements and ecosystem improvements are expected to transform AI from a luxury for large enterprises into a necessity for smaller businesses, enhancing its practical value in real-world applications [7]

Core Insights - The conference focused on the role of advanced storage in empowering AI model development in the AI era [1][2] - Key experts from various organizations discussed the challenges and solutions related to AI inference and storage technology [2][3][4] Group 1: Advanced Storage and AI Inference - The chief expert from the China Academy of Information and Communications Technology emphasized that advanced storage is crucial for improving AI inference efficiency and controlling costs [2] - The national policies highlight the importance of advancing storage technology and enhancing the storage industry's capabilities [2] - A working group was established to promote collaboration and innovation in storage technology within the AI inference sector [2] Group 2: Technical Challenges and Solutions - Current challenges in AI inference include the need for upgraded KV Cache storage, multi-modal data collaboration, and bandwidth limitations [3] - China Mobile is implementing layered caching, high-speed data interconnects, and proprietary high-density servers to enhance storage efficiency and reduce costs [3] - Huawei's UCM inference memory data management technology addresses the challenges of data management, computational power supply, and cost reduction in AI applications [4] Group 3: Industry Collaboration and Future Directions - The conference facilitated discussions among industry experts from various companies, contributing to the consensus on the future direction of the storage industry [5] - The focus is on enhancing computational resource utilization and addressing issues related to high concurrency and low latency in AI inference [4][5] - The successful hosting of the conference is seen as a step towards fostering innovation and collaboration in the storage industry [5]

Core Points - Tesla shareholders approved Elon Musk's $1 trillion compensation plan during the annual meeting [2] - The shareholders also re-elected three board members and supported annual elections for all directors [2] - A non-binding shareholder proposal regarding investment in Musk's AI startup xAI received more votes in favor than against, but high abstention rates warrant further discussion [2] Group 1: Optimus Robot - Musk promoted the Optimus robot, predicting hundreds of billions of units will be deployed, claiming it will "eliminate poverty" [4] - The production cost of each Optimus robot is approximately $20,000 when adjusted for historical dollar value [4] Group 2: Full Self-Driving (FSD) Technology - Musk stated that Tesla's Full Self-Driving (FSD) technology has received "partial approval" in China, with full approval expected around February or March 2026 [5] Group 3: Cybercab Production - Tesla plans to start production of the Cybercab, a fully autonomous vehicle without pedals or steering wheels, in April 2026 [6] - The goal is to reduce the production time to under 10 seconds per vehicle, with a theoretical target of 5 seconds [6] Group 4: Chip Development - Musk emphasized the importance of low-cost, high-efficiency specialized chips for Tesla's robots, indicating discussions with Intel but no agreements yet [7] - Tesla's chips will be produced in Taiwan, South Korea, Arizona, and Texas [7] - Musk mentioned the potential construction of a "gigantic chip factory" to meet the company's chip production needs [8] Group 5: Automotive Focus - Despite diversifying beyond electric vehicles, Musk reiterated that cars remain a crucial part of Tesla's future, aiming for a significant increase in vehicle production [9] Group 6: Roadster Launch - Musk confirmed the upcoming launch of the new Roadster, initially announced in November 2017, with a demonstration planned for April 1 next year and mass production expected in 12 to 18 months [10] Group 7: Lunar and Martian Aspirations - Musk predicted that Tesla vehicles and Optimus robots would play a role in establishing bases on the Moon and Mars [12] Group 8: SpaceX IPO Consideration - Musk expressed that while operating a public company is challenging due to litigation risks, he is considering the possibility of SpaceX going public in the future [13] Group 9: AI Utilization in Vehicles - Musk envisions Tesla vehicles performing "AI reasoning" tasks while idle, potentially creating a distributed AI reasoning fleet that could generate income for owners [14] - He raised concerns about the future control of humanity if AI surpasses human intelligence [14] Market Reaction - Following the announcements, Tesla's stock rose by 0.88% in after-hours trading [14]

Investment Rating - The report indicates a positive outlook for Qualcomm, with expectations of outperforming the market in the upcoming periods [1]. Core Insights - Qualcomm's FY4Q25 results significantly exceeded market expectations, reporting revenue of $11.3 billion against a forecast of $10.76 billion, and a Non-GAAP EPS of $3 compared to the expected $2.87, showcasing robust profitability [1][7]. - The company has officially entered the AI datacenter market, focusing on inference workloads, with competitive advantages in power efficiency and compute density [2][8]. - Non-Apple related QCT revenue grew by 18% year-over-year, driven by strong demand for premium Android devices and increased content value [3][9]. - For FY1Q26, Qualcomm forecasts revenue between $11.8 billion and $12.6 billion, with expectations of continued growth in its QCT handset business and a focus on high-intensity R&D investments [4][10]. Summary by Sections Financial Performance - Qualcomm's QCT revenue reached $9.8 billion, with a quarter-over-quarter increase of 9% and a year-over-year increase of 13%. The EBT was $2.9 billion, reflecting a 17% year-over-year growth and a margin of 29% [1][7]. - The full FY25 Non-GAAP revenue was $44 billion, marking a 13% year-over-year increase, with EPS at $12.03, an 18% increase from the previous year [1][7]. AI Datacenter Strategy - Qualcomm's entry into the AI datacenter market includes the launch of AI 200 and AI 250 SoCs, targeting high efficiency and low-cost architectures, with the first customer, Humain, planning to deploy 200 MW of compute capacity starting in FY27 [2][8]. Non-Apple Revenue Growth - The Snapdragon 8 Elite Gen 5 platform has driven a strong recovery in the premium Android market, with significant contributions from brands like Xiaomi and Honor. Management remains optimistic about sustained growth in premium Android, IoT, and automotive segments [3][9]. Future Outlook - Qualcomm anticipates Q1 FY26 revenue of $11.8–12.6 billion, with QCT revenue projected at $10.3–10.9 billion and EBT margins of 30–32%. The company emphasizes ongoing R&D investments in AI datacenters, edge AI, and other growth engines [4][10].

Core Insights - The "Storage Power China Tour" event held in Beijing on November 4 attracted nearly 20 industry representatives, focusing on how advanced storage can reduce costs and improve efficiency for AI inference [1] - Key challenges in AI inference include the upgrade of KVCache storage needs, multi-modal data collaboration, insufficient bandwidth for computing-storage collaboration, load variability, and cost control [1] - Huawei's open-source UCM (Unified Cache Manager) technology is viewed as a critical solution to address these industry pain points, focusing on multi-level caching and inference memory management [1] Industry Developments - The UCM technology has recently been open-sourced in the Magic Engine community, featuring four key capabilities that can reduce first-round token latency by up to 90%, increase system throughput by up to 22 times, and achieve a tenfold context window expansion [2] - The foundational framework and toolchain of UCM are available in the ModelEngine community, allowing developers to access source code and technical documentation to enhance the technology architecture and industry ecosystem [2] - The open-sourcing of UCM is seen as a significant step beyond mere technical sharing, enabling developers and enterprises to access leading AI inference acceleration capabilities at lower costs and with greater convenience, promoting the widespread adoption of AI inference technology [2]