Do you have any evidence that inference revenue is growing faster than training costs? RLVR is significantly less compute-efficient than token-prediction pretraining - especially as labs are trying to train models to achieve agentic tasks which take tens of minutes per rollout.