2026 年 2 月 27 日 陈伟光 S0340520060001 电话：0769-22119430 邮箱： chenweiguang@dgzq.com.cn SAC 执业证书编号： S0340521020001 电话：0769-22110619 邮箱： luoweibin@dgzq.com.cn S0340524070002 电话：0769-22119302 邮箱： chenzhanqian@dgzq.com.cn 超配（维持） 散热需求向互连系统延伸，连接器散热成为重要补充 深 度 研 AI 集群互连散热专题报告 证 券 研 究 报 告 资料来源：iFind，东莞证券研究所 相关报告 投资要点： 本报告的风险等级为中高风险。 本报告的信息均来自已公开信息，关于信息的准确性与完整性，建议投资者谨慎判断，据此入市，风险自担。 请务必阅读末页声明。 通信行业 SAC 执业证书编号： 罗炜斌 AI集群功耗上扬，集群散热需求增长。AI算力需求呈指数级爆发直接推 动了AI集群功耗上扬，从单芯片到机柜级别的功耗密度的激增已经超越 了传统数据中心的设计极限。以英伟达产品为例，2026年是其AI硬件产 品将从H100/H ...

Core Viewpoint - The article argues that Elon Musk, with his vast wealth and influence, poses a significant challenge to China, more so than any U.S. president, due to his control over critical technologies and industries that could reshape global order [1][3]. Group 1: Musk's Wealth and Influence - Musk's recent merger of xAI and SpaceX resulted in a valuation of $1.25 trillion, making him the first individual to surpass $800 billion in wealth [3]. - Musk's portfolio includes significant stakes in Tesla and other ventures, positioning him as a key player in the future of technology and capital [3][5]. Group 2: Strategic Industries and Technologies - Musk's companies are not limited to automotive and aerospace; they encompass AI, space communication, and low-orbit internet, which are pivotal for global order [5][20]. - Tesla's Shanghai factory is projected to deliver 916,000 vehicles in 2024, accounting for half of global deliveries, while also collecting over 3 billion kilometers of autonomous driving data in China [7][9]. Group 3: National Security Implications - The Starlink project, initially aimed at providing internet access to remote areas, has deployed thousands of satellites that could potentially be used for military purposes, raising concerns about data sovereignty and security [9][11]. - Musk's xAI aims to integrate AI into various applications, creating a closed-loop ecosystem that could dominate standards and control key technologies, posing a challenge for Chinese companies [11][18]. Group 4: China's Response and Opportunities - The article suggests that Musk's presence in China has stimulated local innovation, exemplified by the rapid development of the electric vehicle supply chain [13][15]. - Despite challenges, Chinese companies are adapting and developing their own low-orbit satellite systems and AI capabilities, indicating a competitive response to Musk's influence [16][22]. Group 5: Future Competition Dynamics - The competition between Musk's enterprises and Chinese firms will not only be about market share but also about defining technological standards and controlling communication channels [20][22]. - The article emphasizes the need for China to recognize the new logic of cross-industry competition and to make breakthroughs in multiple fields to effectively respond to Musk's influence [22].

Core Viewpoint - The global technology industry is experiencing a stark contrast, with major players like Microsoft, Google, and Amazon announcing significant AI capital expenditure plans, yet facing market declines due to ROI concerns, resulting in a collective market value loss of nearly $900 billion [1] Group 1: Lenovo's Performance Amidst Market Challenges - Lenovo Group's third-quarter performance for the fiscal year 2025/26, reported on February 12, 2026, showcased a record revenue of $22.2 billion (157.5 billion RMB), an 18% year-on-year increase, with adjusted net profit growth reaching 36%, double the revenue growth rate [1][2] - Lenovo's resilience during supply chain disruptions and the downturn of global tech giants is attributed to its strong delivery capabilities and operational discipline, allowing it to maintain steady performance [2][3] - The company has consistently demonstrated strong delivery resilience during past supply chain crises, regaining its position as the world's leading PC manufacturer even when the market faced stagnation [2][3] Group 2: Supply Chain Management and Operational Strategy - Lenovo's unique operational strategy, termed "global resources, local delivery," enables it to integrate resources across various regions, ensuring consistent procurement and delivery despite supply chain disruptions [4] - The company's "ODM+" model allows it to mitigate localized disruptions and maintain operational continuity, earning recognition as a global supply chain leader [4] - Lenovo's proactive asset management, including signing annual framework agreements with key suppliers, has secured supply and cost stability, providing a competitive edge during market volatility [7][8] Group 3: Market Position and Competitive Advantage - Lenovo's global PC market share increased from 23.7% in the fiscal year 2024/25 to 25.3%, enhancing its bargaining power with suppliers during periods of component shortages [6] - The company's scale advantage allows it to achieve optimal procurement prices, positioning it favorably against competitors amid rising costs [6][8] - Lenovo's customer base, predominantly composed of enterprise and government clients, enables it to absorb cost pressures more effectively than competitors reliant on price-sensitive consumer markets [8][9] Group 4: Long-term Operational Discipline - Lenovo's operational discipline and risk management culture have allowed it to avoid common pitfalls during supply chain disruptions, maintaining stable inventory levels and cash flow [9][10] - The company has successfully locked in low-cost components through strategic inventory management, preventing significant financial losses during price fluctuations [10] - Lenovo's ability to navigate supply chain challenges is attributed to its evolution beyond a traditional PC company, leveraging its position within the global supply chain and AI ecosystem [11][12] Group 5: Future Outlook and AI Integration - Lenovo is well-positioned to capitalize on the AI industry's growth, integrating hardware and software to meet emerging demands for computational power and infrastructure [12][13] - The company's strategy of combining AI PCs, smartphones, and servers with a robust service framework is expected to enhance its order stability and cash flow, solidifying its leadership in the global AI ecosystem [12][13]

Core Viewpoint - The article discusses the rapid development and funding of a 3D AI chip company, 算苗科技 (Suanmiao Technology), which has completed two rounds of financing totaling nearly 1 billion RMB, aimed at developing domestically produced 3D computing chips for AI applications [3][10]. Group 1: Company Overview - 算苗科技 focuses on the research and development of 3D computing chips, with its core product being a customized chip for AI model inference [4]. - The company aims to address the "memory wall" issue that limits AI model computation, as current AI chips face significant inefficiencies due to memory bandwidth constraints [4][5]. - 算苗科技's A4 chip has demonstrated a throughput of 1.26 to 2.19 times that of NVIDIA's H200 in inference tasks on major open-source models [5]. Group 2: Funding and Market Position - The recent funding rounds were led by prominent investors, including Source Code Capital and Shixi Capital, indicating strong market interest and support for the company's vision [3][10]. - The company is positioned to leverage its expertise in 3D IC technology to create a competitive edge in the AI chip market, which is expected to grow significantly [10][19]. Group 3: Technological Innovation - 算苗科技 utilizes a 3D stacked architecture that allows for significantly higher memory bandwidth (up to 32 TB/s), which is crucial for AI model inference [4][13]. - The company’s approach contrasts with traditional GPU architectures, focusing on specialized ASIC designs that optimize performance for specific tasks rather than general-purpose computing [14][15]. Group 4: Strategic Focus - The company has chosen to concentrate on AI model inference rather than training, as it anticipates that 90% of future AI computing demand will be for inference tasks [15][18]. - 算苗科技 believes that the future of AI computing lies in architectural innovation, particularly through 3D stacking and ASIC optimization, which aligns with the growing demand for efficient computing solutions [28][29].

Core Viewpoint - The article discusses the emerging trend of space computing power in the global AI competition, highlighting advancements from both American and Chinese companies in deploying AI models in space [1][4][13]. Group 1: Space Computing Power Developments - Starcloud, backed by Nvidia, has successfully run a large model in space, marking a significant milestone in space computing power [1][4]. - Guoxing Aerospace has announced the launch of the world's first silicon-based intelligent agent service network in space, planning to deploy 2,800 satellites to support billions of silicon-based intelligent agents [2][4]. - The total computing power from the planned satellites will reach 100,000 P-level for inference and 1,000,000 P-level for training, with full deployment expected by 2035 [4][6]. Group 2: Technological Differences - Starcloud's approach involves deploying large models on the ground before sending them to space, while Guoxing Aerospace can deploy general large models directly in orbit and update them as needed [9][10]. - This capability allows for real-time updates and operational flexibility, akin to over-the-air updates in smartphones [9][10]. Group 3: Advantages of Space Computing Power - Space computing power can significantly reduce costs and save land resources, as it operates without the constraints of terrestrial data centers [13]. - It offers energy efficiency by utilizing solar power directly in space, avoiding the high energy consumption associated with ground-based data centers [13]. - The real-time service capabilities of space computing power can enhance applications in various sectors, such as providing fishermen with timely information about fish movements [14][16]. Group 4: Challenges and Technical Considerations - The development of space computing power faces challenges such as hardware selection, the need for on-orbit hardware replacement mechanisms, and the unique environmental conditions of space [19][21]. - Issues like heat management and protection against high-energy particles must be addressed to ensure the reliability and accuracy of space-based computing systems [21][22]. Group 5: Future Outlook - The integration of space computing power with open-source large models presents a unique opportunity for China to establish a leading position in this emerging field [23][24]. - The ongoing advancements in both space computing and AI models are expected to drive significant changes in various industries, promoting broader access to AI technologies [17][24].

Core Insights - The integration of AI inference chips is becoming crucial for automotive intelligence as autonomous driving approaches [2][10] - The demand for inference chips is expected to significantly increase by 2026 due to the rapid growth of automotive intelligence needs [3] Inference Demand Surge - AI model training has been a key growth driver for the AI chip market, with high-end chips like NVIDIA's H100 and H200 being highly sought after, often resulting in multi-million dollar orders [4] - Inference chips have now surpassed training chips in demand, becoming the new mainstay for data center computing power and smart driving applications, as companies focus on translating large models into practical applications [4][5] Automotive Intelligence Key to Success - Autonomous vehicles are evolving into highly integrated "smart mobile terminals" that require real-time decision-making capabilities, supported by the powerful computing power of inference chips [6] - A Level 4 autonomous vehicle can generate data volumes of several gigabytes per second, necessitating rapid processing and analysis for effective driving decisions [6][7] Performance and Efficiency of Inference Chips - Inference chips are designed for edge computing, allowing for immediate data processing without relying on cloud transmission, which is critical for timely decision-making in autonomous driving [7] - New generation inference chips utilize advanced architectures and manufacturing processes, such as 7nm technology, to provide high performance while significantly reducing energy consumption [8] Customization for Autonomous Driving - Inference chips must be tailored for core tasks in autonomous driving, such as visual recognition and decision control, through customized neural network accelerators to enhance processing efficiency and accuracy [9] Industry Transformation with Inference Chips - Inference chips represent a pivotal point in AI industry development, acting as a bridge from research to market application and playing an essential role in automotive intelligence [10] - Achieving automotive-grade certification is a significant hurdle for inference chips, requiring rigorous environmental testing to ensure reliability and stability throughout the vehicle's lifecycle [10][11] Challenges and Future Outlook - Algorithm adaptation is a key challenge for inference chips in automotive applications, necessitating close collaboration between chip manufacturers and automotive companies to optimize performance [11] - The rise of inference chips marks a new phase in the AI and autonomous driving industry, addressing core issues such as cost, latency, and privacy, and enabling deeper integration of AI technologies into operational contexts [11][12] - As AI technology and automotive hardware converge, the future application prospects for inference chips will expand, with increasing competition among automotive companies to develop more competitive autonomous driving solutions [12]

Core Insights - In early 2025, significant developments in the AI chip sector were reported, including Elon Musk's confirmation of Tesla's (TSLA.US) revival of the Dojo 3 supercomputer project, aiming to become the largest AI chip manufacturer globally, and Cerebras Systems' multi-year procurement agreement with OpenAI worth over $10 billion, promising 750 megawatts of computing power by 2028 [1][2]. Group 1: AI Chip Evolution - The evolution of AI chips is characterized by two distinct designs: Cerebras' wafer-scale integration and Tesla's Dojo, which represents a hybrid approach between single-chip and GPU clusters [3]. - The divergence stems from different solutions to the "memory wall" and "interconnect bottleneck" challenges, with traditional GPU architectures facing limitations in memory bandwidth compared to computational power [3][4]. Group 2: Cerebras' Innovations - Cerebras' WSE-3 chip features 40 trillion transistors, 900,000 AI cores, and 44GB of on-chip SRAM, achieving a bandwidth of 214 Pb/s, significantly outperforming NVIDIA's H100 [4]. - The design addresses yield issues associated with large wafers by minimizing the size of each AI core and employing redundancy to maintain performance despite defects [4]. Group 3: Tesla's Strategic Shift - Tesla's Dojo project faced setbacks but was revived with a new focus on "space AI computing," moving away from its original goal of competing with NVIDIA's GPU clusters [7][8]. - The AI5 chip, designed with a 3nm process, is expected to be produced by the end of 2026, aiming for performance comparable to NVIDIA's Hopper architecture [8]. Group 4: Market Dynamics and Competition - The AI chip market is becoming increasingly crowded, with competitors like AMD and NVIDIA rapidly advancing their offerings, which poses challenges for alternative architectures like wafer-scale systems [16][19]. - Cerebras aims to differentiate itself by focusing on low-latency inference systems, capitalizing on the growing demand for real-time AI applications [16][14]. Group 5: Strategic Partnerships - Cerebras' partnership with OpenAI, involving a $10 billion commitment for computing power, highlights the increasing importance of low-latency inference capabilities in the AI landscape [11][12]. - The collaboration reflects a broader trend of established tech companies integrating promising AI chip startups into their ecosystems, which may reshape the competitive landscape [20][21].

Core Insights - The article discusses significant developments in the AI chip sector, highlighting Tesla's revival of the Dojo 3 supercomputer project and Cerebras Systems' multi-billion dollar agreement with OpenAI for AI computing power [1][10]. Group 1: AI Chip Developments - Tesla's Dojo 3 project aims to position the company as a leading AI chip manufacturer, with a focus on "space artificial intelligence computing" rather than traditional training models [6][8]. - Cerebras Systems has secured a contract with OpenAI worth over $10 billion, promising to deliver 750 megawatts of computing power by 2028, emphasizing the growing demand for low-latency inference capabilities [10][11]. Group 2: Chip Architecture and Performance - The distinction between two types of large chips is made: Cerebras' wafer-scale integration and Tesla's wafer-scale system, each addressing the "memory wall" and "interconnect bottleneck" challenges differently [2][4]. - Cerebras' WSE-3 chip boasts 40 trillion transistors and 900,000 AI cores, achieving a memory bandwidth of 21 PB/s, significantly outperforming NVIDIA's H100 [3][11]. Group 3: Strategic Shifts - Tesla's shift in strategy reflects a recalibration of resources, moving away from competing directly with NVIDIA's GPU clusters to focusing on specialized applications in space computing [7][8]. - Cerebras' approach to positioning itself as a provider of dedicated inference machines allows it to capitalize on the emerging demand for low-latency processing, differentiating itself from traditional training platforms [15][19]. Group 4: Market Dynamics and Competition - The AI chip market is becoming increasingly crowded, with competitors like AMD and NVIDIA rapidly advancing their offerings, which poses challenges for alternative architectures like those from Cerebras and Tesla [15][19]. - The collaboration between OpenAI and Cerebras is seen as a strategic move to secure a foothold in the burgeoning inference market, which is expected to dominate AI computing needs in the future [10][19]. Group 5: Future Outlook - The advancements in packaging technology, such as TSMC's CoWoS, are expected to blur the lines between large and small chip architectures, potentially reshaping the competitive landscape [16][19]. - The article concludes that both Tesla and Cerebras are not merely trying to replicate NVIDIA's success but are instead seeking to find value in niches overlooked by general solutions, indicating a long-term battle for survival and innovation in the AI chip market [20].

Core Viewpoint - TSMC's recent financial report has provided strong validation for the AI industry, showcasing significant growth and robust capital expenditure plans that signal future demand for chip manufacturing [6][9]. Financial Performance - TSMC's Q4 2025 financial results exceeded expectations, with revenue growth for eight consecutive quarters and a gross margin surpassing 60%, comparable to software giants [6][12]. - The company has projected capital expenditures of $52 billion to $56 billion for 2026, a substantial increase from $40.9 billion in 2025, indicating confidence in future orders from clients like NVIDIA and AMD [9][28]. Market Dynamics - TSMC's capital expenditure is primarily allocated for building production lines and purchasing equipment, which typically takes 2-3 years to yield results, thus reflecting anticipated growth in orders from chip design companies [9][28]. - The demand for AI computing chips has surged, leading to a shortage of 3nm capacity, with TSMC reportedly halting new 3nm orders due to full capacity bookings for the next two years [18][25]. Competitive Landscape - TSMC's dominance in the 3nm process technology is underscored by its ability to maintain high gross margins, driven by strong demand from AI chip manufacturers [14][18]. - Competitors like Samsung and Intel have struggled to keep pace, with TSMC's 3nm technology remaining unmatched in terms of performance and yield [18][20]. Advanced Packaging - TSMC's advanced packaging technology, particularly CoWoS, has become critical for AI chips, with significant demand leading to a substantial increase in its capital expenditure allocation for this segment [19][20]. - The company has captured a significant share of the advanced packaging market, with NVIDIA alone accounting for approximately 57.4% of TSMC's current capacity [22][20]. Client Relationships - TSMC's collaboration with NVIDIA has evolved from design processes to system-level integration, indicating a deepening partnership that may see NVIDIA surpassing Apple as TSMC's largest client by 2026 [25][28]. - The shift in client dynamics highlights TSMC's reliance on major customers for securing orders and mitigating risks associated with advanced process technologies [25][28].

Key Points Summary of NVIDIA Conference Call Company Overview - **Company**: NVIDIA - **Key Product**: VR 200 architecture with significant innovations in key components and system design Core Insights and Arguments - **VR 200 Architecture**: The VR 200 architecture features a new Ruby design, upgrading from L6 to L10 level delivery, enhancing integration and functionality of the motherboard [2][4] - **Power Consumption**: The new GPUs may reach power consumption levels of 1,000 to 1,500 watts, necessitating widespread adoption of liquid cooling technology [2][5] - **Supply Chain Impact**: Increased demand for raw materials like copper and silver due to higher PCB layer counts (increased from 18 to 24 or 28 layers) [4][5] - **Standardization Effects**: NVIDIA's standardized system design reduces flexibility for brands like Dell and ASUS, potentially leading to increased product homogeneity and pushing some brands towards competitors like AMD or Google [6][7] - **Market Position**: Supermicro maintains a leading position in the domestic GPU server market due to its stable partnership with NVIDIA, while Dell remains a significant player globally [8] - **Manufacturing Partnerships**: Foxconn continues to be a key manufacturing partner for NVIDIA, producing H100, H200 GPUs, and B100, B200 modules, with expected visibility improvements in the coming months [9][10] - **Supply Chain Management**: NVIDIA collaborates closely with Foxconn, deploying automated equipment to ensure efficient production and assembly [12] Additional Important Insights - **Power Supply Improvements**: The VR 200 cabinet utilizes silicon carbide (SiC) technology to enhance power conversion efficiency to 80%-90%, reducing power loss and improving system stability [3][13] - **Future Orders**: NVIDIA anticipates a significant increase in H200 GPU orders in 2026, with an expected domestic market scale of 2 million units, equivalent to 250,000 servers [3][18] - **Delivery Schedule**: The delivery process involves multiple steps, with an estimated one-month timeline from order to final delivery, aiming for approximately 500,000 units per quarter starting from April 2026 [19] - **MPS Market Share**: MPS (New Yuan System) has seen an increase in its share within NVIDIA's supply chain, indicating growing market interest [20] - **Liquid Cooling Systems**: The Robin series integrates more GPUs per CPU compared to the GB300, requiring enhanced liquid cooling solutions to manage increased thermal output [21]