Training and serving large models requires semiconductors made of ultra-pure silicon and compound materials, power conversion built increasingly on gallium nitride, distribution networks made of copper, optical interconnect doped with germanium, and generation capacity that increasingly includes nuclear.
Reading the AI build-out as a materials programme rather than a software one changes which constraints appear binding. Chips are one constraint. Transformers, copper and firm power are others, and they are constrained by mines and permits rather than by fabs.