Investment Rating - The report indicates a positive outlook for the AI computing industry, suggesting an "Overweight" rating for the sector [3]. Core Insights - NVIDIA has detailed its AI computing technology roadmap at the Washington GTC, enhancing confidence in its execution. The roadmap includes the preview of the next-generation Vera Rubin architecture and emphasizes local manufacturing progress in the U.S. [3] - The Blackwell platform has shipped approximately 6 million GPUs to date, with expectations of 20 million GPUs shipped from Blackwell and Rubin combined by CY25-26, generating over $500 billion in revenue, which is five times the lifetime revenue of the Hopper chip [3][5]. - The report highlights the differentiation of computing products into Context and Generation GPUs, with Context GPUs (CPX) optimized for compute-intensive Prefill stages, aiming for higher system economics [3][4]. Summary by Sections AI Computing Technology Roadmap - NVIDIA's roadmap shows a clear path for the development of AI computing products, with significant advancements in GPU architecture and interconnect technology [3][6]. - The NVLink technology is set to evolve to 400G SerDes channel rates, surpassing standard Ethernet capabilities, indicating a strong demand for high-speed interconnects as AI cluster scales grow [3][13]. Hardware Design Innovations - The Vera Rubin platform aims for complete cable-free connections within trays, enhancing deployment ease and reliability while optimizing space for other components [3][19]. - The report suggests a focus on PCB and connector ASP inflation logic, indicating potential growth in related component manufacturers [3][21]. Investment Opportunities - The report recommends monitoring companies involved in PCB production, connectors, and liquid cooling solutions, including Shenghong Technology, Lixun Precision, and BYD Electronics, as they stand to benefit from the advancements in AI computing infrastructure [3][21].
Rubin更新:Context处理器纳入路线图、SerDes448G升级、定调无缆化:AI算力行业跟踪点评之2