Another consequence of inference unit economics improving vs training: it’s profitable to run inference clouds on scrappy datacenters like stranded gas pads, old crypto mines, or telco closets. Any powered shell with decent chips can be turned into a token-refining revenue gen.