[pixl97]

>haven't been running their own LLMs for the last five years

Because the hardware has not existed.

This said by accident I've seen hardware that was brought to a testing company by federal marshals that was massively parallel custom hardware that was likely for signal processing a lot of channels at once. So there is plenty of custom hardware out there, but these items have not been produced at the scale needed (from what anyone can tell) and, again from what we can tell, they don't have the general processing capability that GPU/TPU driven LLMs have.