Investment Rating - The report maintains a "Positive" outlook on the AI computing industry, indicating an expectation for the sector to outperform the overall market [1]. Core Insights - The introduction of the Rubin CPX is specifically optimized for the Prefill phase of AI tasks, potentially reducing BOM costs to as low as one-fourth of previous models. It is designed for large-scale context processing tasks such as AI video generation and software development [3][4]. - The Rubin CPX is projected to deliver 20 PFLOPS of NVFP4 dense computing power with 128GB GDDR7 memory, and it is expected to be integrated into the VR200 Oberon NVL144 or operate as a standalone cabinet for AI clusters, optimizing TCO for end customers [3][4]. - The report highlights four key factors contributing to the improved token economics: enhanced computational efficiency, cost reduction through GDDR replacing HBM, the use of mature FCBGA packaging to avoid costs associated with CoWoS, and savings from not requiring NVLink for scaling [3][4][10]. Summary by Sections Section 1: CPX Optimization - Rubin CPX is tailored for the Prefill phase, significantly lowering the cost of initial token output and KV Cache generation [3][4]. - The PD separation mechanism allows for distinct hardware nodes for Prefill and Decode phases, optimizing resource allocation and reducing costs [8][9]. Section 2: Technological Enhancements - The report discusses the increase in computing density and integration complexity, which drives demand for interconnects, liquid cooling, and assembly [13][23]. - Key changes include the introduction of a new PCB orthogonal mid-board, the elimination of Overpass cables for improved maintenance, and the shift to liquid cooling due to increased power density [14][19][23]. Section 3: Investment Recommendations - The report suggests focusing on companies involved in PCB/CCL increments such as Shenghong Technology and Fuzhou Technology, connector increments like Luxshare Precision, and those enhancing liquid cooling solutions like Invec and BYD Electronics [28][29].
AI算力行业跟踪点评:英伟达RubinCPX:TCO与算力密度再进一步,揭示PCB/液冷、组装增量