HBM，陡生变数

Core Viewpoint - The demand for Google's Tensor Processing Units (TPUs) is expected to be strong, leading to significant growth in the High Bandwidth Memory (HBM) supply chain, with Google projected to become a key demand source in the HBM market, second only to NVIDIA [2] Group 1: Market Dynamics - Bank of America has raised its forecast for Google's TPU shipments from 4 million to 4.6 million units for this year, which is double the previous prediction of 2.3 million units for 2025 [2] - It is anticipated that Google's share in the HBM market may exceed 30% this year, indicating structural growth in the Application-Specific Integrated Circuit (ASIC) market [2] - The shift towards ASICs like TPUs, optimized for specific tasks such as training and inference, is becoming crucial in the evolving AI landscape [2] Group 2: HBM Supply Chain Competition - Samsung Electronics and SK Hynix are competing to maintain their dominance in the HBM supply chain, with both companies supplying HBM3E to Google and working on HBM4 and HBM4E [3] - Samsung has officially announced shipments of HBM4, while SK Hynix has begun mass production of HBM4 and is optimizing with major clients [3] - There are predictions that Google may skip HBM4 and directly adopt HBM4E, which could accelerate the development of HBM4E by Samsung and SK Hynix [3] Group 3: Diversification of Demand - The demand for accelerators is diversifying, with AMD signing a contract to provide Meta with AI accelerators, which is expected to further broaden the sources of HBM demand [4] - Since last year, Samsung has been supplying 12-layer HBM3E products to AMD's flagship AI accelerator MI350 series, indicating a growing diversification in HBM demand sources [4] - The diversification of HBM demand sources is seen as positive for the memory industry, enhancing the bargaining power of Korean manufacturers [5]