The majority of energy consumed is for training the AI models, not providing output from those models.

This means the water consumption is not tied to usage and prompts. Also it means resource consumption to train models is temporary, relative to the model.