It's a combination of tokens consumed, dataset + model storage cost, and inference + training compute cost.