KAIYUAN SECURITIES·2026-01-28 11:15Investment Rating - The industry investment rating is maintained as "Positive" [2] Core Insights - The report emphasizes the transition to the "SuperNode" era in computing infrastructure, driven by the increasing computational demands of AI models, which are evolving from pre-training to multi-faceted scaling [14][17] - The SuperNode architecture relies on enhanced interconnectivity, integrating multiple computing chips into a logical "super GPU/ASIC" to overcome the limitations of traditional single-server setups [39] - The report identifies three main investment themes: "Network End + AIDC + Computing End" and four key sectors: "Optical + Liquid Cooling + Power Supply + Chips" [7] Summary by Sections 1. Transition to SuperNode Era - The evolution of AI models necessitates a shift in computational infrastructure, with demands for higher performance and efficiency [14] - The report highlights the rapid growth in model parameters and training data, indicating a need for advanced computational clusters [14] 2. SuperNode Architecture - SuperNode clusters are defined as large-scale GPU/ASIC systems that expand from thousands to millions of computational units, utilizing both Scale Up and Scale Out strategies [17] - The architecture includes components such as computing nodes, switch trays, and power supply units, which are essential for the operation of SuperNode systems [80] 3. Upgrades in Key Components - The report notes that the increasing penetration of SuperNode servers will drive demand for high-power supplies, liquid cooling systems, and optical communication technologies [5][80] - Specific companies are highlighted as beneficiaries in the sectors of optical modules, liquid cooling, server power supplies, and exchange chips [7] 4. Domestic SuperNode Development - The report discusses the progress of domestic SuperNode solutions, particularly Huawei's Atlas 900 A3 and Atlas 960, which are designed to enhance the performance of domestic AI clusters [6][112] - The performance of domestic chips is shown to improve significantly when deployed in SuperNode configurations, effectively compensating for individual chip limitations [112]