GPU

The GPU is the main chip that performs the heavy parallel math of training and running AI models — effectively the unit in which AI compute is measured. Demand for it vastly exceeds supply, which is what makes it the most visible scarce object in the industry. But its scarcity is instructive rather than final: the GPU's importance to the system is not the same as its being the true bottleneck. Trace the shortage down and the binding constraints have already moved beneath it, into the memory stacked beside it and the packaging that assembles it.

If GPU allocation falls short, training and inference capacity stalls across the whole field, and access concentrates among those who can secure supply — the cloud builders and neoclouds like CoreWeave, Nebius and Microsoft. NVIDIA designs the dominant part, TSMC fabricates it, and test houses like Advantest gate its output; challengers such as Biren and MetaX are trying to break in. The visible scarce thing, in other words, hides the real one below it.