Core Viewpoint - The article discusses the emergence and significance of the "SuperNode" concept in the AI computing market, highlighting the competitive landscape among domestic manufacturers aiming to match or surpass Nvidia's offerings [1][11]. Group 1: SuperNode Concept - The term "SuperNode" refers to high-performance computing systems that integrate multiple AI training chips within a single cabinet, enabling efficient parallel computing [5][7]. - Domestic manufacturers have rapidly adopted the SuperNode concept, with various companies showcasing their solutions at industry events, indicating a collective push towards advanced AI computing capabilities [2][4]. Group 2: Performance Metrics - Companies are emphasizing the performance metrics of their SuperNode products, with Huawei's 384 SuperNode reportedly offering 1.67 times the computing power of similar Nvidia devices [3][12]. - The scale of integration, indicated by numbers like "384" or "640," reflects the number of AI training chips within a single system, serving as a key performance indicator for manufacturers [7][8]. Group 3: Challenges and Solutions - The industry faces a "communication wall" where a significant portion of computing time is spent waiting for data transfer, necessitating the development of SuperNodes to enhance communication efficiency [6][9]. - The transition from traditional computing methods to SuperNode architectures is driven by the need for higher performance in training large AI models, with manufacturers exploring both Scale-Up and Scale-Out strategies [7][8]. Group 4: Competitive Landscape - Domestic firms are positioning their SuperNode products against Nvidia's offerings, with Huawei's Atlas950 expected to outperform Nvidia's NVL144 in several key metrics [11][12]. - The competition is not only about performance but also about innovative engineering solutions to manage power consumption and heat dissipation in densely packed systems [13][15]. Group 5: Market Demand - The primary demand for AI computing resources is expected to come from large internet companies and state-led cloud services, which are likely to drive the market in the next few years [20][21]. - There are concerns about the sustainability of this demand, as companies may face challenges in justifying high capital expenditures for advanced computing resources [21][22]. Group 6: Future Outlook - The article suggests that while hardware challenges exist, the real test for domestic manufacturers will be in developing robust software ecosystems to support their SuperNode offerings [19][22]. - There is optimism about the potential for AI applications in sectors like robotics and advanced manufacturing, which could drive sustained demand for high-performance computing solutions [22].

Core Insights - The AI computing power market is increasingly focused on "SuperNode" technology, with multiple companies showcasing their solutions at various conferences throughout 2023 [2][3] - The emergence of SuperNodes is driven by the need to overcome bottlenecks in training large AI models, particularly the "communication wall" that arises during parallel computing [4][9] - Domestic companies are adopting SuperNode technology as a practical solution to enhance overall computing power, compensating for limitations in single-chip performance [10][12] Group 1: SuperNode Technology - SuperNode refers to a high-density computing solution that integrates multiple AI chips within a single cabinet, allowing them to function as a unified system [6][7] - The design of SuperNodes involves two main approaches: Scale-Up, which increases resources within a single cabinet, and Scale-Out, which connects multiple cabinets [5][8] - The numbers associated with SuperNodes (e.g., "384", "640") indicate the number of AI training chips integrated within a single system, serving as a key metric for performance and density [7][8] Group 2: Industry Competition - Companies like Huawei and Inspur are positioning their SuperNode products as superior to NVIDIA's offerings, with Huawei claiming its Atlas 950 will outperform NVIDIA's NVL144 in multiple performance metrics [10][11] - The competitive landscape is marked by aggressive parameter comparisons, with domestic firms striving to achieve higher integration density within their SuperNode solutions [12][14] - The engineering challenges of integrating numerous high-power chips into a single cabinet necessitate advanced cooling and power supply technologies [12][14] Group 3: Market Demand and Challenges - The primary demand for AI computing power is expected to come from large internet companies and state-led cloud services, which have the infrastructure to support high-end computing needs [19][20] - Despite the strong demand, there are concerns about the sustainability of investments in AI computing infrastructure, particularly regarding the potential for overbuilding [20][22] - The software ecosystem remains a significant challenge for domestic manufacturers, as effective software solutions are crucial for the successful deployment of high-density computing systems [18][22]

Core Viewpoint - China is accelerating its investment in the artificial intelligence (AI) sector, supported by national policies and R&D investments from major tech companies and local suppliers, which is expected to drive the development of domestic computing power and AI models [1][2]. Group 1: Investment and Development - UBS highlights that despite uncertainties in imported AI chips, domestic computing power is continuously developing due to government support and investments from major tech firms [1]. - Alibaba and Baidu are favored by UBS for their ongoing progress in self-developed chips, which will strengthen their positions in the AI value chain [1]. Group 2: Technological Advancements - Recent technological advantages include improvements in domestic GPU performance through internal R&D and local suppliers, as well as system-level enhancements via supernode scaling [2]. - The design of supernodes, such as Alibaba's Panjiu 128 and Huawei's Ascend 384, significantly increases GPU quantities per cabinet, compensating for performance gaps in individual domestic GPUs [2]. Group 3: AI Model Development - AI model developers are optimizing algorithms for domestic GPUs, with DeepSeek's latest v3.2 model utilizing the TileLang programming language to better fit the local algorithm ecosystem [2]. - Most internet companies are accelerating the development of ASICs to optimize workloads and improve cost-effectiveness [2]. Group 4: Hardware and Software Ecosystem - A recent survey of AI chip experts revealed that domestic GPUs are now comparable to NVIDIA's Ampere in performance, although they still lag behind the Blackwell series [3]. - Some domestic chip manufacturers have established their own software stacks or added CUDA compatibility, enhancing engineer migration efficiency, though fragmentation limits scalability [3]. Group 5: Supply Chain and Market Position - China's capabilities in advanced process technology and high-bandwidth memory production are still in early stages, impacting supply chain strength [3]. - Besides Alibaba and Baidu, UBS also sees potential in iFlytek for its advancements in integrating domestic hardware with large model development, and prefers Horizon Robotics, Northern Huachuang, and Zhongwei Company [3].

Core Viewpoint - China is accelerating investments in the artificial intelligence (AI) sector, with UBS highlighting that despite uncertainties in imported AI chips, domestic computing power is continuously developing due to national policy support and investments from major tech companies and local suppliers [1][2]. Group 1: Investment and Development - UBS is optimistic about Alibaba (09988) and Baidu (09888), believing that progress in self-developed chips will strengthen their positions in the AI value chain and that they will continue to invest in AI [1]. - Recent technological advantages include ongoing investments from Chinese internet companies in internal R&D and local GPU suppliers, which are rapidly improving despite existing gaps in chip performance [2]. Group 2: Hardware and Software Ecosystem - The performance of domestic cutting-edge GPUs has reached parity with NVIDIA's Ampere, with next-generation products targeting Hopper, although they still lag behind the Blackwell series [3]. - Some domestic chip manufacturers have established their own software stacks or added CUDA compatibility through translation tools, enhancing engineers' migration efficiency, though the fragmented ecosystem limits scalability [3]. Group 3: Supply Chain and Market Position - In addition to Alibaba and Baidu, UBS is optimistic about iFlytek (002230.SZ) due to its leading progress in integrating domestic hardware with large model development [3]. - The company also favors Horizon Robotics (09660), Northern Huachuang (002371.SZ), and Zhongwei Company (688012.SH) for their unique market positions [3].