Serious question, the stat that AI training runs use as much power as 30,000 homes shocked me
I was reading a piece from the Lawrence Berkeley National Lab last month and it said a single big AI training run can pull as much electricity as 30,000 homes use in a day. That number stuck with me because I spend my weeks sizing breaker panels and running load calcs for buildings, and I kept trying to picture the service entrance for something like that. The part that got me was how they said it's not just the training, it's the cooling and the constant inference calls after the model goes live. I'm not anti AI at all, I use it for quoting jobs, but I had no idea the power draw was in that range. Has anyone seen good numbers on how much of that load is actually being offset by newer chips or better cooling setups?