Signal AI data centre demand outruns compute supply and exhausts memory chips

Summary

KISDI reports that the supply of computing power currently falls short of demand from AI data centres. Bottlenecks have appeared in system semiconductors such as GPUs that are essential for AI and in server memory such as high-bandwidth memory. Where one HBM stack used to be attached to one GPU, configurations now connect twelve or more, so memory chips are being consumed at an unprecedented pace. The strain has spread to general-purpose DRAM and NAND, where supply itself has become a bottleneck. The institute uses this shortage as current evidence for its view that the present memory upturn is driven by demand for AI infrastructure.