Accelerator and memory scarcity

Leading accelerator and HBM production remains concentrated and cannot instantly follow demand.

Modern AI performance is set by the whole system: accelerators, high-bandwidth memory, networking and software. If one layer cannot scale, expensive compute units sit idle, so nominal FLOPS are not the same as usable capacity.

Sources

NVIDIA · Rubin platformOpen primary source