a lot of that compute is used for inference, which is demand-based