作者 | 龙 玥 编辑 | 硬 AI 一场由算力驱动的AI能力大跃升可能正在酝酿。 硬·AI 报告认为，投资者需要为2026年可能出现的AI能力阶梯式提升做好准备。 报告描述了即将到来的算力规模：一个由Blackwell GPU组成的1000兆瓦数据中心，其算力将超过5000 exaFLOPs（每秒五百京次浮点运算）。相比之下，美国政府一台名为"Frontier"的超级计算机算力仅略高 于1 exaFLOP。这种量级的算力增长，是市场预期AI能力将出现非线性提升的核心依据。 报告称，尽管许多LLM开发者普遍认同算力投入将带来能力提升，但也有怀疑论者认为，前沿模型的智 能、创造力和解决问题的能力可能存在上限。 据硬AI，摩根士丹利在一份最新报告中表示，市场可能严重低估了即将在2026年出现的一项人工智能领域 的重大利好——由算力指数级增长驱动的模型能力"非线性"飞跃。 根据这份由Stephen C Byrd等分析师撰写的报告， 多家美国大型语言模型（LLM）开发商计划到2025年 底，将其用于训练前沿模型的算力提升约10倍。这一前所未有的算力投入，预计将在2026年上半年产出 结果，构成一个"未被充分重视的催化 ...

来源：市场投研资讯 （来源：上海证券研究） 市场行情回顾 过去一周（09.15-09.19），SW电子指数上涨2.96%，板块整体跑赢沪深300指数3.40个百分点，从六大 子板块来看，消费电子、电子化学品Ⅱ、光学光电子、半导体、元件、其他电子Ⅱ涨跌幅分别为 4.85%、3.61%、2.89%、2.79%、1.37%、0.74%。 中美贸易摩擦加剧、终端需求不及预期、国产替代不及预期。 报告名称：《台积电领衔晶圆代工2.0市场，英伟达50亿美元注资英特尔——电子行业周报 （2025.09.15-2025.09.19）》 分析师：陈凯 SAC编号：S0870525070001 研报发布日期：2025年9月22日 发布机构： 上海证券有限责任公司 核心观点 25Q2台积电营收突破302亿美元，晶圆代工2.0市占率达到38%。9月16日，根据芯智讯援引市场研究机 构CounterpointResearch最新公布的数据显示，受益于人工智能（AI）、高性能计算（HPC）芯片旺盛需 求，2025年第二季全球半导体代工市场达到了417亿美元。其中，台积电的市占率高达70.2%，当季营 收突破302亿美元，较上季增长18. ...

Core Insights - The SW Electronics Index increased by 2.96% from September 15 to September 19, outperforming the CSI 300 Index by 3.40 percentage points [2] - In the six sub-sectors, the performance was as follows: Consumer Electronics (4.85%), Electronic Chemicals II (3.61%), Optical Electronics (2.89%), Semiconductors (2.79%), Components (1.37%), and Other Electronics II (0.74%) [2] Semiconductor Market Overview - TSMC's revenue surpassed $30.2 billion in Q2 2025, with a market share of 38% in the Foundry 2.0 segment [3] - The global semiconductor foundry market is projected to reach $41.7 billion by Q2 2025, with TSMC holding a dominant market share of 70.2% [3] - TSMC's revenue grew by 18.5% compared to the previous quarter, with nearly 75% of its revenue coming from advanced process technologies below 7nm [3] - Major clients include NVIDIA, AMD, and Apple, indicating strong demand for advanced chips driven by AI and high-performance computing [3] - Samsung's efforts in 2nm GAA technology are hindered by a lack of large-scale production orders, limiting its ability to challenge TSMC's market position [3] Strategic Partnerships - NVIDIA announced a $5 billion investment in Intel to enhance collaboration in the data center and personal computing sectors [4] - Intel will customize x86 CPUs for NVIDIA, which will integrate these into its AI infrastructure platform [4] - This partnership aims to merge NVIDIA's AI capabilities with Intel's extensive x86 ecosystem, potentially positioning NVIDIA as a significant shareholder in Intel [5] Investment Recommendations - The electronic semiconductor sector is expected to experience a comprehensive recovery by 2025, with an improved competitive landscape and profitability [6] - Recommended stocks include those in semiconductor design with low PE/PEG ratios, such as Zhongke Lanyun and Juchip Technology, as well as key materials and carbon-silicon industry leaders [6]

Group 1 - The core focus of the news is on the significant investments by major US tech companies in the UK to establish AI infrastructure, marking a shift from mere political gestures to tangible technological deployment [1][3]. - Microsoft announced a $30 billion investment (approximately £22 billion) for AI data centers, cloud computing facilities, and local R&D teams [2]. - Nvidia plans to deploy 120,000 Blackwell GPUs in the UK and invest £500 million in local AI infrastructure company Nscale [2][19]. Group 2 - The total investment from major companies, including Google and Salesforce, exceeds £31 billion (approximately $42 billion), indicating a comprehensive tech investment agreement spanning AI, energy, policy, and chips [3]. - The investments represent a national-level industrial layout rather than simple corporate expansion, with the US AI giants transitioning the concept of "sovereign AI" into reality [3][4]. - There are concerns about whether the UK is building its own AI capabilities or merely becoming a node in the global layout of US companies [4]. Group 3 - Microsoft CEO Satya Nadella emphasized that the $30 billion investment is not for market speculation but to build foundational computing infrastructure [5][12]. - The investment is divided into three parts: hardware (land, data centers), software (local sales and R&D), and human resources (building and training AI teams) [8][9]. - Nadella expressed that while AI has potential, realizing its economic value requires time and organizational changes [10][11]. Group 4 - Nvidia CEO Jensen Huang highlighted the importance of data sovereignty, suggesting that the UK should utilize its own data to train large models [17][18]. - Nvidia's deployment of GPUs is not just about selling hardware but about helping the UK establish a complete data center ecosystem [19][20]. - Huang pointed out that the UK has the potential to develop its own AI capabilities, provided there is investment in foundational infrastructure [22][23]. Group 5 - OpenAI's Stargate UK project aims to establish local large model infrastructure in the UK, marking a shift from global API services to localized deployments [26][30]. - The project will support the development of sovereign AI, ensuring that high-quality models can be trained and run locally [27][28]. - This new approach signifies a transformation in AI roles, integrating deeply with local policies and regulations [30][31]. Group 6 - The UK has gained significant investments and infrastructure development, but concerns remain about who truly controls the core capabilities [35][36]. - While the UK benefits from job creation and infrastructure, the ultimate control over the technology and training remains with US companies [37][38]. - The collaboration raises questions about whether it represents genuine partnership or dependency on US tech giants [39][40]. Group 7 - The investment wave signals a shift in AI competition from model performance to deployment capabilities [42]. - The UK has positioned itself as a key node in the global AI landscape, with the northeastern region emerging as a new AI industrial hub [42]. - The collaboration model highlights the need for countries to assess their roles as either co-builders or mere hosts in the global tech strategy [43].

Core Viewpoint - Nvidia is positioned to become the world's first $10 trillion company, with the introduction of its new Rubin CPX AI chip potentially accelerating this growth trajectory [1][8]. Group 1: New Technology Introduction - Nvidia unveiled the Rubin CPX, a new class of GPU designed for massive-context processing, capable of handling over 1 million tokens [3][4]. - This technology is expected to enhance AI coding assistants, transforming them into sophisticated systems that can manage large-scale software projects and improve the quality of long-form video creation [5][9]. Group 2: Economic Impact - Nvidia claims that Rubin CPX could deliver a return on investment of 30x to 50x at full scale, significantly improving AI inference economics [6]. - The introduction of Rubin CPX is anticipated to create substantial new use cases, attracting interest from various AI companies [8][9]. Group 3: Market Growth Projections - To reach a market cap of $10 trillion by 2030, Nvidia would need a compound annual growth rate (CAGR) of approximately 18.3%, with current earnings showing a year-over-year increase of 59% in Q2 2025 [12]. - Even with a projected earnings growth of around 41% in 2026, Nvidia is expected to maintain strong growth, especially if Rubin CPX and its successors perform as anticipated [13].

Investment Rating - The report maintains an "Outperform" rating for the communication industry, indicating an expected performance exceeding the benchmark index by more than 10% [1][43]. Core Insights - Nvidia's recent financial results continue to drive global computing investment sentiment, with a strong sequential growth in network revenue, which positively impacts the optical module market [3][15]. - Alibaba Cloud's Q2 performance exceeded expectations, with a 26% year-on-year revenue increase to 33.4 billion yuan, marking the highest growth rate in three years, and a significant rise in capital expenditures [5][17]. - The emergence of supernodes in domestic chip development is expected to accelerate market growth for copper connections and optical modules, enhancing the value of server manufacturing [7][18]. Summary by Sections Industry Dynamics - Nvidia's Q2 revenue reached $46.7 billion, a 56% year-on-year increase, with data center revenue also growing by 56% [3][15]. - Google showcased its TPUv7 at Hotchips2025, indicating advancements in optical switching technology that could enhance performance significantly [4][16]. - Alibaba Cloud's capital expenditures surged to 38.6 billion yuan in Q2, reflecting a 220% increase year-on-year, driven by robust AI-related revenue growth [5][17]. Market Performance - The overall market saw significant gains during the week of August 25-29, 2025, with the Shenwan Communication Index rising by 12.38% [8][19]. - The top-performing sectors included optical cables and modules, with weekly increases of 52.59% and 29.99%, respectively [8][19]. Company Recommendations - Key companies to watch in the domestic computing server sector include ZTE, Unisoc, and Huakong Technology [7][19]. - For overseas supernode opportunities, companies like Zhongji Xuchuang and New Yisheng are highlighted [7][19].

Core Insights - The article highlights the rapid rise of Cambrian Technology, surpassing Kweichow Moutai to become the highest-priced stock in A-shares, driven by the booming AI market [1] - NVIDIA's stock price fell despite impressive Q2 2026 financial results, with revenue reaching $46.7 billion, a 6% increase from Q1 and a 56% year-over-year growth [2][4] - NVIDIA's CEO Jensen Huang emphasizes the company's transformation into an AI infrastructure provider, with expectations of AI infrastructure investments reaching $3 to $4 trillion by the end of the decade [18][19] Financial Performance - NVIDIA's data center revenue was $41.1 billion, a 5% increase from Q1 and a 56% year-over-year growth [8] - The company has consistently exceeded revenue expectations, leading to heightened market expectations for future performance [4][5] - NVIDIA's revenue from the Chinese market decreased to $2.769 billion, down nearly $900 million from the previous year, with its contribution to total data center revenue dropping to a "low single-digit percentage" [24][25] Product Development - NVIDIA has developed the Blackwell NVLink 72 system, which significantly enhances performance and energy efficiency [10][11] - The new Blackwell architecture's B100/B200 series offers a 2.5x performance improvement over the H100 [11] - NVIDIA is transitioning to producing compliant chips for the Chinese market, including a reduced-performance version of the Blackwell architecture [26][27] Market Trends - The demand for AI computing power is expected to grow exponentially, driven by the proliferation of inference and intelligent AI applications [21] - NVIDIA's CUDA platform and AI model frameworks have become essential tools for AI developers, creating a strong ecosystem that is difficult for customers to replace [22][23] - The Chinese market presents a significant opportunity for NVIDIA, estimated at around $50 billion this year, with a projected annual growth rate of 50% [29] Competitive Landscape - Domestic competitors are emerging, with companies like DeepSeek developing models tailored to local chip architectures [32][33] - The introduction of new parameter formats, such as UE8M0 FP8 by DeepSeek and NVFP4 by NVIDIA, indicates a competitive push in the AI training space [36][38] - As local chip manufacturers collaborate to create compatible software stacks, confidence in domestic solutions is expected to rise [43]

Group 1: Nvidia (NVDA) - Nvidia reported Q2 revenue of $46.7 billion, slightly below Goldman Sachs' expectation of $47 billion but above Wall Street's consensus of $46.5 billion [1] - Data center revenue reached $41.1 billion, a year-over-year increase of 56%, while gaming revenue was $4.3 billion, exceeding analyst expectations of $3.9 billion [1] - Gross margin was reported at 72.4%, in line with expectations, and operating margin was 64.5%, surpassing forecasts [1] - Nvidia's earnings per share (EPS) was $1.04, consistent with Goldman Sachs' prediction of $1.05 and above Wall Street's expectation of $1.02 [1] - For Q3, Nvidia expects revenue midpoint of $54 billion, aligning with Wall Street's expectations but below Goldman Sachs' forecast of $57 billion [1] - Following the earnings report, Nvidia's stock price may experience a slight decline, according to Goldman Sachs [1] - Nvidia's management confirmed that no H20 products were shipped to China during the quarter [1] Group 2: Broadcom (AVGO) - Broadcom announced multiple updates related to VMware Cloud Foundation, integrating private AI services into VMware Cloud Foundation 9.0 [2] - The company plans to incorporate Nvidia's Blackwell GPU into VMware Cloud Foundation to support advanced AI model deployment in private cloud environments [2] - Goldman Sachs reiterated a "buy" rating for Broadcom with a target price of $340, indicating the company's commitment to enhancing AI capabilities and strengthening security measures in VMware Cloud Foundation [2] - Following Nvidia's earnings report, Nvidia's stock price dropped by 5%, although it has risen approximately 35% year-to-date, significantly outperforming the Nasdaq index's 12% increase [2]

Core Viewpoint - FlashAttention-4, introduced by Tri Dao at the Hot Chips 2025 conference, demonstrates significant performance improvements over previous versions and competitors, particularly in the context of NVIDIA's GPU architecture [1][2][10]. Summary by Sections FlashAttention-4 Introduction - FlashAttention-4 is reported to be up to 22% faster than NVIDIA's cuDNN library implementation on the Blackwell architecture [2]. - The new version incorporates two key algorithmic improvements: a new online softmax algorithm that skips 90% of output rescaling and a software simulation for better throughput [4][5]. Performance Enhancements - The kernel developed by Tri Dao's team outperforms NVIDIA's latest cuBLAS 13.0 library in specific computation scenarios, particularly when the reduction dimension K is small [7]. - FlashAttention-4 utilizes CUTLASS CuTe Python DSL, which is significantly more challenging to port to ROCm HIP compared to CUDA C++ [6]. Competitive Landscape - The development of FlashAttention is seen as a core advantage for NVIDIA, as Tri Dao and his team primarily use NVIDIA GPUs and have open-sourced much of their work for the developer community [10]. - There are implications for AMD, suggesting that financial incentives may be necessary to encourage Tri Dao's team to develop for ROCm [10]. Historical Context and Evolution - FlashAttention was first introduced in 2022, addressing the quadratic time and memory overhead of traditional attention mechanisms by reducing memory complexity from O(N²) to O(N) [12]. - Subsequent versions, including FlashAttention-2 and FlashAttention-3, have continued to enhance performance, with FlashAttention-2 achieving speed improvements of 2-4 times over its predecessor [21]. Technical Innovations - FlashAttention-3 achieved a speed increase of 1.5-2.0 times over FlashAttention-2, reaching up to 740 TFLOPS on H100 GPUs [23]. - FlashAttention-4 introduces native support for Blackwell GPUs, addressing previous compilation and performance issues [24]. Community Engagement - The GitHub repository for FlashAttention has garnered over 19,100 stars, indicating strong community interest and engagement [25].

Group 1: Product Launch and Features - Nvidia has launched its Jetson AGX Thor developer kit and production modules, a next-generation robotics platform designed to power millions of robots with advanced AI capabilities [1][4] - The Blackwell GPU-powered platform delivers up to 2,070 FP4 teraflops of AI compute with 128GB of memory in a 130-watt power envelope, offering 7.5 times the AI compute and 3.5 times better energy efficiency than its predecessor, Jetson Orin [2] - Jetson Thor is designed to run multiple generative AI models at the edge, enabling robots and humanoids to interact intelligently and operate autonomously in complex real-world environments [3] Group 2: Market Adoption and Pricing - Industry leaders such as Amazon Robotics, Boston Dynamics, Agility Robotics, Meta Platforms, and Caterpillar have adopted Jetson Thor, while companies like OpenAI and John Deere are evaluating the system [4] - The Jetson Thor is available starting at $3,499 and integrates with Nvidia's robotics software stack to accelerate the development of next-generation humanoid and industrial robots [4] Group 3: Financial Outlook - JPMorgan analyst Harlan Sur has a bullish outlook on Nvidia, projecting July-quarter revenue of $46–$47 billion, driven largely by the ramp-up of GB200 rack shipments [5] - Sur expects October-quarter revenue guidance of $53–$54 billion+, as Nvidia scales GB200 volumes to 8,000–9,000 racks, with full-year Blackwell shipments reaching 28,000–30,000 racks [5] - Gross margins are projected to rise to 73% in the fourth quarter, moving toward the mid-70% range by year-end, citing supply chain efficiencies [7]