摩根士丹利：人工智能供应链-GB300 和 H20 需要了解的

Investment Rating - Industry view is rated as In-Line [7] Core Insights - Key industry trends include GB300 system assembly yield, China's H20 shipment license, and upcoming AI capital expenditures from cloud service providers (CSPs) [7] - TSMC maintains a mid-40% five-year CAGR target for AI semiconductor revenue despite uncertainties surrounding H20 [2][6] - Huawei's Ascend 920 chip is set for mass production in 2H25, indicating ongoing evolution in China's AI GPU landscape [3][12] Summary by Sections AI Supply Chain Dynamics - The forecast for H20 baseboard assembly has been cut in half, raising concerns about potential production cuts at TSMC and Amkor [2] - If the H20 license is delayed, TSMC may redirect unused wafer capacity to H200 production if demand persists [2] Hardware Developments - The GB300 server racks are reverting to the Bianca board design, which may improve assembly yields [4][5] - Analysts suggest that the reversion to the Bianca board indicates improved yields for the GB200/Bianca board [5] Stock Market Implications - Updates on AI capital expenditures from global CSPs will be crucial for investor confidence [6] - Key Asian supply chain stocks like TSMC, Hon Hai, and Wistron appear undervalued but require new catalysts for performance [6] GPU Market in China - Huawei's new AI GPU, Ascend 920, will utilize a 6nm process and is expected to replace H20 chips, which are currently restricted [12][15] - The CloudMatrix 384 system from Huawei aims to compete with Nvidia's offerings, showcasing significant computing power [13] Future Projections - The demand for CoWoS (Chip on Wafer on Substrate) is projected to grow significantly, with NVIDIA leading the demand [18][22] - AI computing wafer consumption is expected to reach up to $16.1 billion in revenue by 2025 [37]