We now have audited data on water consumption for AI. Over the 18 month lifespan of Mistral Large 2, a 128B model, all water usage (including chats, training; hardware & data centers) took as much water as 678 US households use yearly. Each query is only 1/100 of a teaspoon.

A page of AI output from Mistral uses 0.000000857% of the water of a hamburger.

I messed up on my teaspoon conversion, it is 9 teaspoons of water for a query (OpenAI reported 1/15 of a teaspoon, not sure if they are more efficient or measuring differently). 9 teaspoons is equivalent to 0.001875% of a hamburger.