inference.

Serve models on nodes close to your users, autoscaling with traffic. You pay per token, not per idle server.

Bring your own model or serve from the catalog. Either way, a proof lands for every batch, and you're billed for exactly the compute you touched.