英伟达 CES 发布了什么？星环科技为何受益 20260106 摘要 英伟达通过 PU 和 SSD 优化，旨在提升 GPU 计算效率，特别是在线学 习等新模型架构的数据处理能力，通过低成本 SSD 替换 DRAM，实现 更高效的数据存储。 英伟达新架构通过提升内存使用效率，打破内存墙，增加热数据需求， 显著提升向量数据库的数据流量，尤其是在 H200 芯片大量采用的情况 下，业务增量可能达数百亿级别。 星环科技作为国内领先的独立第三方向量数据库厂商，受益于英伟达新 技术和按流量计费模式，有望充分利用存算一体化带来的流量增长，实 现业务空间成百倍放大。 向量数据库与传统数据库的主要区别在于按流量计费而非按节点收费， 更适合实时训练和在线学习等应用场景，其商业模式因灵活性和经济性 更具吸引力。 英伟达收购 Groq 并采用 Atrium 方式优化 HBM 交互层，使得未来模型 架构中的固定权重更新更加高效，并促进 SSD 与 HBM 之间的数据传输 速度，大幅提高系统性能。 向量数据库与传统数据库最大的不同在于其按流量计费，而非按节点收费。传 统数据库主要管理冷数据，确保数据不丢失、不变形，并且需要持续付费。而 ...

Core Insights - The global tech industry is experiencing a shift from cloud-based AI to edge AI, driven by the limitations of cloud dependency and the need for real-time processing in critical applications [1][4][18] - The current trend emphasizes the development of smaller, more efficient AI models that can operate independently on edge devices, rather than relying on large cloud models [16][18] Group 1: Challenges of Cloud Dependency - Cloud-based AI systems face significant latency issues, which can be detrimental in time-sensitive applications like autonomous driving [2][4] - Privacy concerns arise from the need to transmit sensitive data to cloud servers, making edge computing a more attractive option for users [2][4] Group 2: The Shift to Edge AI - The industry is moving towards a "cloud-edge-end" architecture, where complex tasks are handled by cloud models while real-time tasks are managed by edge devices [7][18] - Edge AI must overcome the "impossible triangle" of high intelligence, low latency, and low power consumption, necessitating innovative solutions [7][8] Group 3: Techniques for Edge AI Implementation - Knowledge distillation is a key technique that allows smaller models to retain the intelligence of larger models by learning essential features and reasoning paths [8][10] - Extreme quantization reduces model size and increases speed by compressing model weights, allowing for efficient processing on edge devices [10][11] - Structural pruning eliminates redundant connections in neural networks, further optimizing performance for edge applications [10][11] Group 4: Hardware Innovations - The "memory wall" issue in traditional architectures leads to inefficiencies, prompting the development of specialized architectures that integrate storage and computation [11][13] - Companies are exploring dedicated chip designs that optimize performance for specific AI tasks, enhancing efficiency in edge computing [13][14] Group 5: Industry Evolution - The focus is shifting from general-purpose AI models to specialized models that excel in specific applications, improving reliability and performance [15][16] - The Chinese AI industry is collectively recognizing the importance of practical applications over sheer model size, leading to a more grounded approach to AI development [16][18]

Report Industry Investment Rating No relevant content provided. Core Views of the Report - The report focuses on the empowerment of intelligent computing centers, elaborating on the latest development trends around demand scenarios, key capabilities, and implementation ecosystems to further release the empowerment effect of intelligent computing centers and promote the deep integration of AI and the real economy [5]. - Facing the "14th Five-Year Plan", the artificial intelligence computing infrastructure has three important development trends: clear demand scenarios for optimal resource allocation, focused key capabilities for improved service levels, and aggregated implementation ecosystems for accelerated value release [24]. - In the future, the demand scenarios of artificial intelligence computing infrastructure will become more diverse and complex, key capabilities will be more intensive and soft, and the implementation ecosystem will be more aggregated and collaborative [75]. Summary by Directory 1. Evolution Trend of Artificial Intelligence Computing Infrastructure - **Technological Innovation: Upgrading of Tri - in - One Intelligent Computing Facilities**: China's artificial intelligence computing infrastructure is evolving towards large - scale clustering, green and low - carbon development, and high - speed interconnection. For example, Huawei's Ascend 384 super - node and ZTE's Nebula intelligent computing super - node achieve high - speed interconnection of computing cards; the liquid - cooling technology in the China Mobile data center reduces energy consumption [12][13][14]. - **Layout Optimization: Coordinated Development of National Intelligent Computing Facilities**: Policy guidance promotes the high - quality development of intelligent computing centers. The scale of intelligent computing centers continues to grow, and regional intelligent computing is deployed in a more coordinated and intensive manner. For instance, as of June 2025, the total rack scale of computing centers in use in China reaches 1.085 million standard racks, and the intelligent computing scale is 788 EFlops [16][17]. - **Industrial Upgrade: Collaborative Development of the Entire Intelligent Computing Industry Chain**: The intelligent computing industry is growing rapidly, with upstream hardware achieving domestic breakthroughs, mid - stream facilities being built on a large scale, and downstream applications accelerating penetration into various industries. Three major operators and AI giants are actively deploying intelligent computing [18][19][20]. 2. Important Trends in the Empowerment of Artificial Intelligence Computing Infrastructure - **Clearer Demand Scenarios for Optimal Allocation of Intelligent Computing Resources**: The positioning of demand scenarios is becoming clearer, promoting the precise empowerment of intelligent computing centers. The construction of artificial intelligence computing infrastructure is shifting from "building well" to "using well", and the rights and responsibilities of all parties are becoming more explicit [25]. - **Focused Key Capabilities for Improved Intelligent Computing Service Levels**: The supply of key capabilities is being strengthened. In terms of basic support, innovation services, and operation guarantee, the service capabilities of intelligent computing centers are continuously improving, promoting the value - closed - loop and long - term development of intelligent computing centers [26][27]. - **Aggregated Implementation Ecosystems for Accelerated Release of Intelligent Computing Value**: The ecological system is being integrated, and the collaborative mechanism is being improved. The construction of artificial intelligence computing infrastructure is evolving towards an integrated solution of "computing power + algorithm + data + scenario + service", and a sustainable and high - value partner network is being initially established [28]. 3. Demand Scenarios of Artificial Intelligence Computing Infrastructure - **Large - Model Pre - training Scenario**: Training large - scale pre - trained models (with over a thousand billion parameters) requires high - end ten - thousand - card cluster centers with E - level computing capabilities. Domestic operators and AI manufacturers are actively building such clusters [30][31][33]. - **Large - Model Fine - tuning Scenario**: Small - scale intelligent computing centers (with a computing capacity of 100 PFlops) can effectively support the fine - tuning of industry models. Most domestic intelligent computing centers are focusing on this scenario [34][36]. - **Large - Model Inference Scenario**: Cloud - side inference dominates the current inference demand scenarios. Different inference application scenarios have different requirements for inference models and intelligent computing centers, and specialized intelligent computing centers for inference are emerging [37][39][40]. 4. Key Capabilities of Artificial Intelligence Computing Infrastructure - **Basic Support Capabilities**: Training scenarios focus on cluster computing power effectiveness, stability, single - cluster computing power scale, and compatibility with mainstream computing frameworks. Inference scenarios focus on throughput, latency, and the heterogeneity of intelligent computing cards [44][45][46]. - **Innovative Service Capabilities**: Training scenarios emphasize high - efficiency cloud services, efficient model migration, and diverse data governance. Inference scenarios focus on the pooling and scheduling capabilities of intelligent computing resources and efficient model migration and deployment [50][51][52]. - **Operation Guarantee Capabilities**: Both training and inference scenarios focus on the flexibility of computing power scheduling, the cost - effectiveness of computing power leasing, and security and compliance. Training scenarios also pay attention to the richness of cooperative partners [55][56][57]. 5. Implementation Ecosystem of Artificial Intelligence Computing Infrastructure - **Collaboration between Intelligent Computing and Data Elements**: Collaborating closely with high - value data is the core for intelligent computing centers to improve basic support capabilities. For example, the Wenzhou Artificial Intelligence Computing Center and the Guian New Area are promoting the transformation of high - quality data resources into intelligent computing ecological capabilities [60][61]. - **Collaboration between Intelligent Computing and Algorithm Models**: Collaborating with high - level algorithm models is the key for intelligent computing centers to improve innovative service capabilities. For example, the Chongqing Artificial Intelligence Innovation Center and the Wuling Mountain (Lichuan) Artificial Intelligence Computing Center are promoting the development and application of industry - specific models [63][64][65]. - **Collaboration between Intelligent Computing and Cross - domain Intelligent Computing**: Promoting cross - domain intelligent computing interconnection and collaboration is an important exploration for the improvement of intelligent computing center operation capabilities. Operators' intelligent computing centers have achieved practical breakthroughs in long - distance interconnection [66][67]. - **Collaboration between Intelligent Computing and Industry Scenarios**: Collaborating closely with industry scenarios is the core driving force for the continuous evolution and upgrading of the intelligent computing center ecosystem. The Chang'an Automobile Intelligent Computing Center and the Yunnan Communications Investment Intelligent Computing Center are typical examples of in - depth collaboration [68][70]. - **Collaboration between Intelligent Computing and Regional Industries**: Collaborating with regional industries is an important guarantee for intelligent computing centers to achieve multi - dimensional and full - scenario empowerment. Intelligent computing centers in Ningbo, Wuhan, and Dalian are promoting regional industrial development [71][73]. 6. Development Outlook - **More Diverse and Complex Demand Scenarios**: The demand scenarios of artificial intelligence computing infrastructure will become more diverse, complex, and deeply integrated. There will be higher requirements for computing power, storage, industry integration, and cloud - edge - end collaboration. Different stakeholders should play different roles [76][77]. - **More Intensive and Soft Key Capabilities**: The artificial intelligence computing infrastructure is shifting from extensive hardware stacking to refined service improvement, including large - scale clustering, resource pooling, open - source development, and service - orientation. Industry organizations and operators should take corresponding measures [78][79][80]. - **More Aggregated and Collaborative Implementation Ecosystems**: The implementation of artificial intelligence computing infrastructure empowerment depends on a more aggregated and collaborative ecosystem, including multi - party participation, joint innovation, and industrial cultivation. Government departments and operators should play their roles [81][82][83].

Market Dynamics - The selection process for the new Federal Reserve Chair is in its final weeks, with Kevin Hassett seen as the frontrunner. The Treasury Secretary indicated that the nomination could be announced before the end of the year, with traders increasing bets on interest rate cuts [1] - U.S. stock indices rose for the third consecutive day, with Google challenging Nvidia in the AI model space. Google's stock saw a significant increase, while Nvidia's stock dropped [1] - The onshore and offshore RMB appreciated against the USD, reaching over a year-high, with the Nasdaq Golden Dragon China Index also seeing gains [2] - Southbound funds recorded a net purchase of approximately 11.166 billion HKD, with Alibaba and Kuaishou receiving significant inflows [3] Industry Developments - The Singapore National AI Strategy is shifting its focus to Alibaba's Qwen open-source architecture, marking a significant expansion of Chinese open-source AI models globally [2] - The global wafer foundry market is expected to reach $199.4 billion by 2025, driven by strong AI demand, with a projected CAGR of 14.3% from 2025 to 2030 [6] - The demand for SSDs is increasing, with North American cloud service providers planning to expand their use of SSDs for both warm and cold data applications [12] - The memory market is experiencing severe shortages, with expectations of continued price increases for DRAM and storage flash memory [12] Company Updates - Alibaba's cloud intelligence group reported a revenue of 39.82 billion CNY for Q2 of FY2026, a 34% year-on-year increase, with AI-related product revenue growing for nine consecutive quarters [4][5] - NIO reported a record delivery of 87,071 vehicles in Q3, a 40.8% year-on-year increase, with revenue reaching 21.79 billion CNY [14] - Meta Platforms is reportedly considering a significant investment in Google's TPU for its data center construction [8] - The company CFM noted a significant price increase in SSDs, driven by AI demand, with expectations of continued shortages in the memory market [12]

Summary of Longxin Storage Conference Call Industry Overview - The global DRAM market is experiencing an upward demand trend, driven by traditional demand recovery and emerging applications such as artificial intelligence, with an expected compound annual growth rate (CAGR) of nearly 5% [3][4] - The Chinese market accounts for over 30% of the global DRAM market, with a projected growth rate of around 8%, primarily supported by the consumer electronics and automotive industries [3][4] Company Insights: Longxin Storage - Longxin Storage, established in 2016, is currently undergoing IPO counseling and is expected to accelerate its listing process [2] - The company holds less than 10% of the global DRAM market share but has significant growth potential, particularly in the domestic market, where its share could increase to over 30% [5][6] Production Capacity - As of the end of 2024, global DRAM monthly production capacity is approximately 1.8 million wafers, expected to rise to 1.9-2 million wafers by the end of 2025 [6] - Longxin Storage's monthly production capacity is projected to grow from 200,000 wafers at the end of 2024 to 300,000 wafers by the end of 2025, representing about 15.6% of global capacity and a year-on-year increase of 50% [6] Product Development - Longxin Storage is transitioning from DDR4 to DDR5, launching a new 16GB DDR5 product using a 16nm process [7] - The expected market share for DDR5 shipments is projected to increase from nearly 1% in Q1 2025 to around 7% by Q4 2025, while LPDDR product share is expected to rise from 0.5% to 9% [7] Upstream Equipment and Material Opportunities - The expansion of Longxin Storage's capacity and product iteration will drive demand for upstream equipment and materials [8] - Key companies to watch in the semiconductor equipment sector include North Huachuang, Zhongwei Company, Tuojing Technology, and Huahai Qingke [8][9] Future Investment Opportunities - The HBM (High Bandwidth Memory) sector is highlighted as a significant area for investment, with expectations for domestic HBM supply chain breakthroughs by 2026 [10] - Specific investment opportunities in the HBM supply chain include wafer manufacturing companies like North Huachuang and Zhongwei, and testing and packaging companies like Jingzhida and Xinyuanwei [11][12] Conclusion - Longxin Storage is positioned for substantial growth within the DRAM market, with a focus on expanding production capacity and transitioning to advanced memory technologies. The overall DRAM market is set for growth, particularly in China, with various upstream and HBM-related investment opportunities emerging in the semiconductor sector.

Core Viewpoint - The company has made a strategic investment in Shenzhen Xinzhan Technology Development Co., Ltd. (referred to as "Xinzhan") by increasing its capital participation, which is expected to enhance its competitiveness in the semiconductor storage sector [1][2] Group 1: Investment Details - The company has invested a total of RMB 20 million, acquiring a 5.55% stake in Xinzhan, which has a newly registered capital of RMB 333,333.33 [1] - The investment is part of a consortium that includes Shixi Capital, Huadeng, and Xinchuan Technology Center [1] Group 2: Industry Context - Xinzhan specializes in the research and development of high-performance enterprise-level SSD products, which are in high demand due to the growth in artificial intelligence and cloud infrastructure [1] - The enterprise storage sector is expected to experience strong growth, with applications in data centers for internet, cloud services, finance, and telecommunications, as well as in consumer electronics like smartphones and PCs [1] Group 3: Strategic Rationale - The investment aims to leverage Xinzhan's advantages in semiconductor high-performance enterprise storage control chips, modules, and data services, while combining it with the company's expertise in packaging and testing [2] - This collaboration is intended to drive technological innovation and business expansion in the semiconductor storage field, thereby enhancing the company's core competitiveness [2]

Summary of Key Points from the Conference Call Industry and Company Overview - The conference call primarily discusses advancements in the autonomous driving chip technology by Huawei, focusing on the new generation of chips and their implications for the automotive industry. Core Insights and Arguments 1. **Next-Generation Chip Performance**: Huawei's new generation chips will offer 500-800 TOPS computing power, utilizing a single-chip solution to replace the dual-chip approach, which addresses transmission limitations and reduces costs, with expected pricing slightly above $10,000, lower than dual-chip solutions [1][4] 2. **Chip Architecture**: The vehicle-side chip architecture is based on the Da Vinci architecture, optimized for integer operations rather than floating-point operations, leading to significant cost differences [1][5] 3. **Algorithm Transition**: Huawei's autonomous driving algorithms are transitioning from a two-stage structure to an end-cloud collaborative Vivo framework, enhancing generalization capabilities in complex scenarios [1][13] 4. **Data Quality Importance**: High-quality data labeling and engineering are crucial for improving training outcomes, with simulation-generated high-quality scenarios being a key method [16] 5. **Chip Development Plans**: The next MDG1,000 chip will significantly enhance computing power and bandwidth, moving from 100 GB/s to 200-280 GB/s, with a focus on integrated storage and computing [2] 6. **Single vs. Dual Chip Advantages**: The new single-chip solution offers advantages over dual-chip configurations, including cost efficiency and improved performance in various driving conditions [3][4] 7. **L3 and L4 Autonomous Driving Plans**: L3 level autonomous driving is expected to launch by the end of this year or early next year, while L4 level technology is in testing, with plans for gradual rollout in high-value models [11][32] 8. **Sensor Fusion Strategy**: Huawei emphasizes a multi-sensor fusion approach, integrating lidar, cameras, and radar to enhance perception and safety in complex driving environments [22][23] Additional Important Content 1. **Market Positioning**: Huawei's focus is on specific automotive applications, contrasting with competitors like NVIDIA, which cater to a broader range of customer needs [9] 2. **Regulatory Challenges**: Current regulations do not fully support L3 capabilities, impacting the public declaration of such features despite the technology being ready [28][31] 3. **Future Technology Integration**: The fifth-generation lidar is set to be introduced this year, with plans for integration into mass-produced models, although actual deployment may vary based on hardware configurations [29][30] 4. **Performance Metrics**: The current multi-modal large language model parameters are around 1 billion, significantly lower than competitors like Tesla, which has models with parameters in the tens of billions [14][19] This summary encapsulates the key points discussed in the conference call, highlighting Huawei's advancements in autonomous driving technology and the implications for the automotive industry.