Core Insights - The competition in the AI large model sector is shifting towards AI inference, marking the beginning of the AI inference era, with edge computing emerging as a new battleground in this field [1][2]. AI Inference Era - Major tech companies have been active in the AI inference space since last year, with OpenAI launching the O1 inference model, Anthropic introducing the "Computer Use" agent feature, and DeepSeek's R1 inference model gaining global attention [2]. - NVIDIA showcased its first inference model and software at the GTC conference, indicating a clear shift in focus towards AI inference capabilities [2][4]. Demand for AI Inference - According to a Barclays report, the demand for AI inference computing is expected to rise rapidly, potentially accounting for over 70% of the total computing demand for general artificial intelligence, surpassing training computing needs by 4.5 times [4]. - NVIDIA's founder Jensen Huang predicts that the computational power required for inference could exceed last year's estimates by 100 times [4]. Challenges and Solutions in AI Model Deployment - Prior to DeepSeek's introduction, deploying and training AI large models faced challenges such as high capital requirements and the need for extensive computational resources, making it difficult for small and medium enterprises to develop their own ecosystems [4]. - DeepSeek's approach utilizes large-scale cross-node expert parallelism and reinforcement learning to reduce reliance on manual input and data deficiencies, while its open-source model significantly lowers deployment costs to the range of hundreds of calories per thousand calories [4]. Advantages of Edge Computing - AI inference requires low latency and proximity to end-users, making edge or edge cloud environments advantageous for running workloads [5]. - Edge computing enhances data interaction and AI inference efficiency while ensuring information security, as it is geographically closer to users [5][6]. Market Competition and Player Strategies - The AI inference market is rapidly evolving, with key competitors including AI hardware manufacturers, model developers, and AI service providers focusing on edge computing [7]. - Companies like Apple and Qualcomm are developing edge AI chips for applications in AI smartphones and robotics, while Intel and Alibaba Cloud are offering edge AI inference solutions to enhance speed and efficiency [7][8]. Case Study: Wangsu Technology - Wangsu Technology, a leading player in edge computing, has been exploring this field since 2011 and has established a comprehensive layout from resources to applications [8]. - With nearly 3,000 global nodes and abundant GPU resources, Wangsu can significantly improve model interaction efficiency by 2 to 3 times [8]. - The company's edge AI platform has been applied across various industries, including healthcare and media, demonstrating the potential for AI inference to drive innovation and efficiency [8].

点击注册小程序 特别申明： 本订阅号中所涉及的证券研究信息由光大证券研究所编写，仅面向光大证券专业投资者客户，用作新媒体形势下研究 信息和研究观点的沟通交流。非光大证券专业投资者客户，请勿订阅、接收或使用本订阅号中的任何信息。本订阅号 难以设置访问权限，若给您造成不便，敬请谅解。光大证券研究所不会因关注、收到或阅读本订阅号推送内容而视相 关人员为光大证券的客户。 报告摘要 北京时间3月19日凌晨，英伟达举办2025年GTC大会，黄仁勋在圣何塞 SAP 中心发表的现场主题演讲，关 注代理式AI、机器人、加速计算等领域的未来发展。此外，该大会还包括1000多场具有启发性意义的会 议，以及400多项展示、技术实战培训和大量独特的交流活动。 提出Agentic AI，新的推理范式将继续推动全球数据中心建设 黄仁勋按照"Generative AI（生成式AI）、Agentic AI（智能体）、Physical AI（具身AI）"三个阶段的进 化路线，将Agentic AI描述为AI技术发展的中间态。Scaling Law的发展需要投入更多的数据、更大规模的 算力资源训练出更好的模型，训练规模越大，模型越智能，预计全球数 ...

Core Viewpoint - SoftBank has agreed to acquire Ampere Computing for $6.5 billion, indicating a strong belief in the potential of Ampere's chips to play a significant role in artificial intelligence and data centers [1][2]. Group 1: Acquisition Details - The acquisition reflects SoftBank's commitment to advancing AI technology, with CEO Masayoshi Son emphasizing the need for breakthrough computing capabilities [1]. - Ampere, founded eight years ago, specializes in data center chips based on Arm Holdings technology, which is widely used in smartphones [1]. - SoftBank plans to operate Ampere as a wholly-owned subsidiary [1]. Group 2: Market Context - The acquisition comes amid a surge in demand for chips that support AI applications like OpenAI's ChatGPT [2]. - SoftBank has announced several transactions aimed at increasing its influence in the AI sector, including a $500 billion investment plan to establish data centers in the U.S. [2]. - Oracle, a major investor and customer of Ampere, is involved in the "Star Gate" initiative alongside SoftBank and OpenAI [2]. Group 3: Competitive Landscape - Intel, AMD, and Arm design microprocessors that play a crucial role in AI, often working alongside GPUs from Nvidia [3]. - Nvidia is promoting Arm processors as alternatives to Intel and AMD chips for AI tasks, which could reshape the market [3]. - IDC predicts that the market for AI microprocessors will grow from $12.5 billion in 2025 to $33 billion by 2030 [3]. Group 4: Ampere's Position - Ampere's microprocessors target the general data center market, with a new chip named Aurora designed for AI inference applications [4]. - Major tech companies like Amazon, Google, and Microsoft are focusing on developing their own Arm-based microprocessors, although Oracle continues to support Ampere [4][5]. - Oracle holds a 29% stake in Ampere, with an investment value of $1.5 billion after accounting for losses [4].

Core Viewpoint - High-tech companies consistently develop roadmaps to mitigate risks associated with technology planning and adoption, especially in the semiconductor industry, where performance and capacity limitations can hinder business operations [1][2]. Group 1: Nvidia's Roadmap - Nvidia has established an extensive roadmap that includes GPU, CPU, and networking technologies, aimed at addressing the growing demands of AI training and inference [3][5]. - The roadmap indicates that the "Blackwell" B300 GPU will enhance memory capacity by 50% and increase FP4 performance to 150 petaflops, compared to previous models [7][11]. - The upcoming "Vera" CV100 Arm processor is expected to feature 88 custom Arm cores, doubling the NVLink C2C connection speed to 1.8 TB/s, enhancing overall system performance [8][12]. Group 2: Future Developments - The "Rubin" R100 GPU will offer 288 GB of HBM4 memory and a bandwidth increase of 62.5% to 13 TB/s, significantly improving performance for AI workloads [9][10]. - By 2027, the "Rubin Ultra" GPU is projected to achieve 100 petaflops of FP4 performance, with a memory capacity of 1 TB, indicating substantial advancements in processing power [14][15]. - The VR300 NVL576 system, set for release in 2027, is anticipated to deliver 21 times the performance of current systems, with a total bandwidth of 4.6 PB/s [17][18]. Group 3: Networking and Connectivity - The ConnectX-8 SmartNIC will operate at 800 Gb/s, doubling the speed of its predecessor, enhancing network capabilities for data-intensive applications [8]. - The NVSwitch 7 ports are expected to double bandwidth to 7.2 TB/s, facilitating faster data transfer between GPUs and CPUs [18]. Group 4: Market Implications - Nvidia's roadmap serves as a strategic tool to reassure customers and investors of its commitment to innovation and performance, especially as competitors develop their own AI accelerators [2][4]. - The increasing complexity of semiconductor manufacturing and the need for advanced networking solutions highlight the competitive landscape in the AI and high-performance computing sectors [1][4].

Core Viewpoint - Nvidia's innovations in AI inference technologies, including the introduction of inference Token expansion, inference stack, Dynamo technology, and Co-Packaged Optics (CPO), are expected to significantly reduce the total cost of ownership for AI systems, thereby solidifying Nvidia's leading position in the global AI ecosystem [2][4][68]. Group 1: Inference Token Expansion - The rapid advancement of AI models has accelerated, with improvements in the last six months surpassing those of the previous six months. This trend is driven by three expansion laws: pre-training, post-training, and inference-time expansion [8]. - Nvidia aims to achieve a 35-fold improvement in inference cost efficiency, supporting model training and deployment [10]. - As AI costs decrease, the demand for AI capabilities is expected to increase, demonstrating the classic example of Jevons Paradox [10][11]. Group 2: Innovations in Hardware and Software - Nvidia's new mathematical rules introduced by CEO Jensen Huang include metrics for FLOPs sparsity, bidirectional bandwidth measurement, and a new method for counting GPU chips based on the number of chips in a package [15][16]. - The Blackwell Ultra B300 and Rubin series showcase significant performance improvements, with the B300 achieving over 50% enhancement in FP4 FLOPs density and maintaining an 8 TB/s bandwidth [20][26]. - The introduction of the inference stack and Dynamo technology is expected to greatly enhance inference throughput and efficiency, with improvements in smart routing, GPU planning, and communication algorithms [53][56]. Group 3: Co-Packaged Optics (CPO) Technology - CPO technology is anticipated to significantly lower power consumption and improve network scalability by allowing for a flatter network structure, which can lead to up to 12% power savings in large deployments [75][76]. - Nvidia's CPO solutions are expected to enhance the number of GPUs that can be interconnected, paving the way for networks exceeding 576 GPUs [77]. Group 4: Cost Reduction and Market Position - Nvidia's advancements have led to a performance increase of 68 times and a cost reduction of 87% compared to previous generations, with the Rubin series projected to achieve a 900-fold performance increase and a 99.97% cost reduction [69]. - The overall trend indicates that as Nvidia continues to innovate, it will maintain a competitive edge over rivals, reinforcing its position as a leader in the AI hardware market [80].

Core Viewpoint - Nvidia's dominance in the AI chip market is being challenged by emerging competitors like DeepSeek, as the focus shifts from training to inference in AI computing demands [1][2]. Group 1: Market Dynamics - The AI chip market is experiencing a shift from training to inference, with new models like DeepSeek's R1 consuming more computational resources during inference requests [2]. - Major tech companies and startups are developing custom processors to disrupt Nvidia's market position, indicating a growing competitive landscape [2][5]. - Morgan Stanley analysts predict that over 75% of power and computing demand in U.S. data centers will be directed towards inference in the coming years, suggesting a significant market transition [3]. Group 2: Financial Projections - Barclays analysts estimate that capital expenditure on "frontier AI" for inference will surpass that for training, increasing from $122.6 billion in 2025 to $208.2 billion in 2026 [4]. - By 2028, Nvidia's competitors are expected to capture nearly $200 billion in chip spending for inference, as Nvidia may only meet 50% of the inference computing demand in the long term [5]. Group 3: Nvidia's Strategy - Nvidia's CEO asserts that the company's chips are equally powerful for both inference and training, targeting new market opportunities with their latest Blackwell chip designed for inference tasks [6][7]. - The cost of using specific AI levels has decreased significantly, with estimates suggesting a tenfold reduction in costs every 12 months, leading to increased usage [7]. - Nvidia claims its inference performance has improved by 200 times over the past two years, with millions of users accessing AI products through its GPUs [8]. Group 4: Competitive Landscape - Unlike Nvidia's general-purpose GPUs, inference accelerators perform best when optimized for specific AI models, which may pose risks for startups betting on the wrong AI architectures [9]. - The industry is expected to see the emergence of complex silicon hybrids, as companies seek flexibility to adapt to changing model architectures [10].

Core Viewpoint - Nvidia's CEO Jensen Huang expressed excitement about the potential demand for AI inference, which is expected to far exceed current large language models (LLMs), potentially requiring millions of times more computing power [1][5]. Group 1: AI Inference and Demand - The demand for inference will significantly increase, especially for long-thought inference AI models, which may require several orders of magnitude more computing power than pre-training [5]. - Nvidia's Blackwell architecture is designed for inference AI, improving inference performance by 25 times compared to Hopper while reducing costs by 20 times [6][34]. - The DeepSeek-R1 inference model has generated global enthusiasm and is an outstanding innovation, being open-sourced as a world-class inference AI model [1]. Group 2: Financial Performance and Projections - Nvidia reported record revenue of $39.3 billion for the fourth quarter, a 12% quarter-over-quarter increase and a 78% year-over-year increase, exceeding expectations [32]. - The data center revenue for fiscal year 2025 is projected to be $115.2 billion, doubling from the previous fiscal year [32]. - Nvidia's CFO Colette Kress expects profit margins to improve once Blackwell production increases, with margins projected to be in the mid-70% range by the end of 2025 [2][11]. Group 3: Product Development and Supply Chain - The supply chain issues related to the Blackwell series chips have been fully resolved, allowing for the next training and subsequent product development to proceed without hindrance [1]. - Blackwell Ultra is planned for release in the second half of 2025, featuring improvements in networking, memory, and processors [16][60]. - Nvidia's production involves 350 factories and 1.5 million components, achieving $11 billion in revenue last quarter [8][53]. Group 4: Market Dynamics and Growth Areas - The global demand for AI technology remains strong, with the Chinese market's revenue remaining stable [20][68]. - Emerging fields such as enterprise AI, agent AI, and physical AI are expected to drive long-term demand growth [14][24]. - Nvidia's full-stack AI solutions will support enterprises throughout the entire AI workflow, from pre-training to inference [25]. Group 5: Infrastructure and Future Outlook - The current AI infrastructure is still utilizing various Nvidia products, with a gradual update expected as AI technology evolves [26][27]. - Nvidia's CUDA platform ensures compatibility across different generations of GPUs, facilitating a flexible update process [28]. - The company anticipates significant growth in data center and gaming businesses in the first quarter, driven by strong demand for Blackwell [44].

Core Viewpoint - The company reported strong financial results for Q4 2025, with revenue reaching $39.3 billion, a 12% quarter-over-quarter increase and a 78% year-over-year increase, leading to an annual revenue of $130.5 billion, up 114% [1] Group 1: Financial Performance - Q4 data center revenue was $35.6 billion, marking a record high with a 16% quarter-over-quarter increase and a 93% year-over-year increase, driven by the release of the Blackwell architecture and expansion of Hopper 200 [2] - Q4 gaming revenue was $2.5 billion, down 22% quarter-over-quarter and 11% year-over-year, but annual revenue reached $11.4 billion, up 9% [2] - Professional visualization revenue for Q4 was $511 million, with a 5% quarter-over-quarter increase and a 10% year-over-year increase, totaling $1.9 billion for the year, up 21% [2] - Automotive revenue reached a record $570 million in Q4, with a 27% quarter-over-quarter increase and a 103% year-over-year increase, totaling $1.7 billion for the year, up 55% [2] Group 2: Product and Technology Developments - The Blackwell architecture contributed $11 billion in revenue for the quarter, emphasizing its impact on performance and cost efficiency in AI inference workloads [3] - The company launched a cluster of 100,000 GPU instances for inference and model customization, catering to the growing demand for AI applications across various industries [3] - The AI inference platform supports large-scale datasets, particularly in finance, healthcare, and retail, addressing the need for efficient processing [3] Group 3: Future Outlook - The company expects Q1 2026 total revenue to reach $43 billion, with a 2% quarter-over-quarter increase, and gross margin projected between 70.6% and 71% [3] - Operating expenses are anticipated to rise to a median of $3 billion for the year, with a Q1 expected tax rate of 17% [4] - Shareholder returns will include stock buybacks and cash dividends totaling $8.1 billion for the fiscal year [4]

Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The TMT sector outperformed the CSI 300 index, with significant gains driven by the release of DeepSeek's models, particularly in the computer and media sub-sectors [3][4] - DeepSeek's V3 and R1 models are reshaping the competitive landscape of AI large models, potentially leading to breakthroughs in vertical application scenarios such as AI healthcare, education, and finance [3][4] - The rapid user growth of DeepSeek, reaching 22.15 million daily active users within 20 days of launch, indicates a surge in demand for inference capabilities, which is expected to drive growth in computing power requirements [3][4] Market Weekly Review - The TMT sector saw significant weekly gains, with the computer industry leading at 22.29%, followed by media at 17.43% and electronics at 6.43% [3][4] - Notable stock performances included Qingyun Technology with a 208.19% increase and Light Media with a 264.43% increase [3][4] Industry News and Key Company Announcements Important Industry Events - The report highlights several key events in the AI sector, including the launch of new models by major companies and significant user growth for DeepSeek [15][17] - The AI technology exhibition in Dubai and the AI action summit in Paris are noted as important gatherings for industry leaders [16][17] Industry News - DeepSeek's rapid user growth and the introduction of new models are expected to enhance the competitive dynamics in the AI market [17][21] - The report mentions various companies, including JD Cloud and Huawei, integrating DeepSeek's models into their services, indicating a trend towards broader adoption of AI technologies [17][21] Company Announcements - Several companies, including Kingsoft and Cloud Tianyi, are reported to be integrating DeepSeek's models into their products, showcasing the growing influence of DeepSeek in the industry [34][36] - The report also notes that DeepSeek's API services are being adopted by various cloud service providers, further expanding its market reach [34][36]