Is AI water crisis actually a hard requirement, or just how we build?

I was recently reading about how a chatgpt query supposedly uses just a fifteenth of a teaspoon of water. It's true, but that only counts on-site cooling, not the water behind the electricity. If you zoom out, texas data centers were heading toward 399 billion gallons a year by 2030. that's lake mead drawn down 16 feet and it's starting to shape local decisions. Across the USA, more than 50 cities have now paused or turned down new data center builds, a lot of it over water, especially in places already dealing with drought Here's what i keep turning over: cooling scales with concentration, so thousands of GPUs in one building dump the whole thermal load on a single watershed, but a cast AI report this year put average GPU utilization at around 5 percent and roughly 10 million servers sit idle and batch inference, which is the fastest-growing workload, is atomizable, it doesn't need everything in one place the way training does so maybe the water problem isn't a hard requirement of how AI works and maybe it's just how we happen to build. if the work were spread across idle hardware that's already cooled and plugged in worldwide, the thermal load would spread with it, and no single town's aquifer would get drained for one company's building i've been messing around with exactly this idea, distributed compute with an orchestration layer to coordinate nodes across locations, and the early results genuinely surprised are good how do you all feel about P2P compute networks where we can consume idle supply and turn it into compute power?