you get better models with more compute.

its not just PoW at inference. It's PoW of inference + training.