Data center water “use” numbers are largely misunderstood, because about 80% of the water attributed to them is withdrawn by power plants and then returned, not consumed.
On-site data center cooling is the real consumptive water load, and in Maricopa County that’s tiny compared to golf courses: 0.12% of daily water vs 3.8%, yet data centers generate ~50x more tax revenue per unit of water.
Per-prompt AI water impact is closer to 1 milliliter than a full bottle, so individual chatbot use barely nudges a person’s water footprint unless they live in a prompt window all day.
The real AI infrastructure risk is electricity: utilities like APS already say they can’t meet all the requested data center power, and grid build‑outs are driving higher residential bills in key regions.
There’s a tradeoff: places like Loudoun County pay more for power infrastructure but also become far wealthier from data center tax revenue, which rarely shows up in the same conversation as costs.
Masley argues policy should force hyperscalers to internalize grid and infrastructure costs instead of pushing them onto ratepayers, while avoiding the opposite extreme of “build everything” national security rhetoric.
For anyone tracking AI infra, the piece is useful context on how to correctly frame water vs power debates and who should really pay for the next wave of GPU-heavy data centers.
Source: Data centers aren’t the water villains you think they are, environmentalist says
