硬AI·2025-10-21 10:26Core Insights - The core argument of the article is that the bottleneck in AI development has shifted from chip manufacturing and packaging to downstream infrastructure, including data center power supply, liquid cooling, high bandwidth memory (HBM), server racks, and optical modules [2][4][9]. Upstream Capacity Expansion - Chip manufacturing and packaging have significantly expanded, alleviating previous concerns about supply shortages [5][6]. - TSMC has reported strong AI demand and is working to close the supply-demand gap, with a lead time of only six months for expanding CoWoS capacity [6][9]. - The report predicts that global CoWoS demand will reach 1.154 million wafers by 2026, a 70% year-on-year increase, indicating a robust supply response [6][12]. Downstream Infrastructure Challenges - As chip supply is no longer the main issue, the focus has shifted to the availability of data center space, power, and supporting infrastructure, which have longer construction cycles than chip manufacturing [9][12]. - The deployment of large-scale GPU clusters presents significant challenges in power consumption and heat dissipation, leading to a preference for liquid cooling solutions and high-voltage direct current (HVDC) power supply systems [9][12]. - The demand for HBM is expected to explode, with global consumption projected to reach 26 billion GB by 2026, with NVIDIA alone accounting for 54% of this demand [9][12]. Investment Opportunities - The shift in focus towards downstream infrastructure opens new investment opportunities beyond traditional chip manufacturers, emphasizing the importance of companies that excel in power, cooling, storage, memory, and networking [12][13]. - Global cloud service capital expenditure is expected to grow by 31% to $582 billion by 2026, significantly higher than the market's general expectation of 16% [12]. - AI server capital expenditure could see approximately 70% year-on-year growth if AI servers' share of capital expenditure increases [12][13].