“This isn't about marketing claims, this is just basic hardware benchmarking. The inference hardware is getting more efficient per watt.”

Lobsters · Sep 22 · Essay "Tokens Too Cheap to Meter" sparks debate over AI cost collapse