Morgan Stanley predicts a significant shift in global demand for high-bandwidth memory (HBM, high-speed stacked memory used in AI chips) in 2027. Google's TPU is projected to account for 35% of global demand, slightly lower than Nvidia's 36%, bringing their combined share to 71%. AMD is expected to account for 16%, Amazon AWS for 5%, Creative Electronics customers for approximately 3%, and Microsoft and Meta for approximately 2% each. In terms of chip type, GPUs are projected to account for 52% of HBM demand, while custom AI chips (ASICs) will account for 47%, essentially splitting the market evenly. The HBM configuration per chip is also continuously increasing: the AMD MI500 is expected to have 768GB, the Google TPU v9 576GB, and the Nvidia Rubin Ultra 384GB. This signifies a shift in HBM demand from being dominated by Nvidia GPUs to being driven by custom chips from both Nvidia and major cloud vendors, with Google now becoming a comparable HBM consumer to Nvidia.
