智通财经网·2025-12-19 00:55Core Insights - The report from CITIC Securities indicates that the supernode solution is expected to scale rapidly, serving as a fundamental computing unit for future AI infrastructure, with advantages such as efficient communication bandwidth and native memory semantics [1][2] Group 1: Supernode Development - The MoE (Mixture of Experts) architecture imposes new hardware requirements, leading to the emergence of scale-up supernodes [2] - Supernodes face complex systemic challenges compared to traditional eight-card servers, including heat dissipation, stability issues from mixed optical and copper interconnects, and long-term reliability concerns [2][3] - The current phase of supernode solutions is characterized by a variety of competing technologies, with domestic solutions like Huawei's CloudMatrix384 and Alibaba's Panjiu emerging [3] Group 2: Technical Challenges and Solutions - As computing density increases, liquid cooling solutions with a PUE closer to 1, such as phase change immersion cooling, may see greater development opportunities if stability issues can be resolved [4] - The complexity of supernode servers has significantly increased, requiring deep consideration of chip integration, heat dissipation, and interconnects, transforming server manufacturers into core system integrators [5] Group 3: Investment Strategy - The supernode technology is in its early stages, with the MoE architecture likely to become mainstream, presenting new adaptive requirements for hardware development [7] - The report suggests that companies with customization capabilities and supply chain management skills in the server manufacturing sector are likely to see significant growth opportunities [7]