PSA: Found out GPT-style models burn through roughly 500 ml of water for every 10 to 50 chat replies
I was reading a paper from UC Riverside last week about water use in data centers and it stopped me cold. The number that got me: cooling a mid-size AI data center can pull down up to 5 million gallons a day in peak summer, which is like a small town's whole supply. I ran that math against my own usage and my dumb little habit of asking a chatbot to reword emails now feels different. Has anyone here actually seen a solid breakdown of how much water per query versus per training run? I keep hearing different numbers and I want to know which ones to trust.