
A 2024 estimate put a 100-word GPT-4 email at 519 millilitres of water, while training GPT-3 may have directly consumed enough for 350,000 people to drink for a day
The 500-millilitre-per-query figure for ChatGPT comes from a 2024 UC Riverside and Washington Post collaboration, and training a single large model can evaporate 5.4 million gallons — enough drinking water for 300,000 people for a day. The fight over that number is now a projected $58 billion US infrastructure problem.















