大厂正在抛弃GPU

Core Insights - The global AI infrastructure market is facing a severe supply shortage, particularly for GPUs, with an expected order volume of 2 million units this year against only 700,000 available units [1] - The demand for self-developed ASICs by cloud service companies is projected to grow at a rate of 44.6%, surpassing the 16.1% growth rate for GPUs, indicating a structural shift towards ASIC adoption due to GPU supply constraints [1] - The supply chain risks for GPUs are expected to peak this year, with production processes and high bandwidth memory (HBM) being interlinked, meaning any bottleneck could disrupt overall supply [1] Group 1 - TSMC is expanding its advanced packaging production lines, crucial for AI accelerators, but the gap between rapidly growing order volumes and actual shipments will persist due to the time required for capacity expansion [2] - ASIC chips, initially led by Google's TPU, are gaining attention as they are designed for specific AI workloads, offering advantages in energy efficiency, performance, and total cost of ownership (TCO) in the long run [2] - The AI accelerator market for ASIC users is expected to maintain a compound annual growth rate (CAGR) of approximately 28% until 2030, with the generative AI ASIC market projected to grow from about $24.9 billion in 2024 to approximately $186.7 billion by 2032, reflecting an annual growth rate of around 28.6% [2] Group 2 - This year is viewed as a critical turning point for the ASIC market, with industry executives noting that the current GPU supply shortage is a short-term issue but will have long-term implications on decision-making [3] - Major tech companies are increasingly viewing GPUs as strategic assets rather than stable commodities, leading to a shift towards reducing GPU dependency and increasing the share of ASICs in new data center investment plans [3]