A few people asked me to make an image of how much irrigated farmland would use the same water required for ALL ChatGPT usage, including every part of the process. I did a botec and my best guess right now is that inference uses about as much water as training, and power generation uses ~5x as much water as the data centers themselves, so it looks something like this. The water cost of manufacturing chips is marginal compared to how much water they use over their lifetimes.