OpenAI burns <5B$/year in inference. They're raising that money to hyperscale larger models.