Every ChatGPT Query Has a Water Bill
Amazon finally published its data center water usage numbers — 2.5 billion gallons in 2025. That’s not a typo. Billion. With a B.
The timing isn’t accidental. Seattle just passed a one-year moratorium on new data center construction, pushed in part by Amazon’s own employees who watched their city’s water infrastructure strain under the weight of the cloud. Amazon’s report is the company holding up its hands and saying “look, we’re actually more efficient than the other guys.”
And sure, the comparison chart is there: Amazon at 0.12 liters per kWh, Microsoft higher, Google higher, Meta higher. Google’s number is specifically for Gemini data centers, which is a convenient carve-out — compare our whole fleet against their most intensive facilities, and we look great. That’s the kind of graphs you get when your PR team writes the methodology.
The part that bugs me isn’t the water usage itself. It’s the framing. Amazon says “90 percent of the time our data centers use air cooling.” The other 10 percent — the hottest hours of the hottest days — they switch to evaporative cooling, which is the water-intensive bit. That sounds reasonable until you remember what a “hottest day” looks like in Northern Virginia, which is the world’s largest data center market and also a region facing serious water stress. Or in Arizona, where the ground is literally sinking from groundwater depletion.
The cloud isn’t a cloud. It’s a building. With pipes.
Counterargument: Amazon points to a peer-reviewed paper showing their data centers are seven times more water-efficient than the industry average. They’re raising server temperature tolerances to reduce cooling needs. They’re investing in water replenishment projects. These are real efforts, and efficiency improvements at their scale matter.
Fair. But here’s the thing — the industry average is terrible. Being seven times better than terrible is still not great. And as AI workloads grow, the trend line is going up, not down. Every model training run, every inference call, every ChatGPT query leaves a water trace. We’re building a world where thinking is metered in gallons.
The bigger problem is that water isn’t priced into the economics of AI. Data center operators pay for the water they use, sure, but they don’t pay for the externalities — the aquifer depletion, the municipal water supply competition, the environmental cost of consuming 2.5 billion gallons that could have gone to farms or homes. These costs are socialized while the profits are private.
Seattle’s moratorium is the canary. Not because Seattle is running out of water tomorrow, but because the mechanism for rationing it is about to shift from “we build what we want” to “the city says no.” That’s a regulatory cost that none of the AI capex projections have priced in.
I don’t think data centers should stop using water. I think the water should cost what it actually costs. If every gallon had its full social price tag, the calculus on air cooling vs. evaporative cooling would shift. The hottest-day problem would start looking like a design problem worth solving rather than a PR problem worth spinning.
You can’t abstract away physics. The cloud lives in a building, the building needs power and cooling, and the cooling needs water. The only honest conversation is about who pays the full bill.
And right now? The bill is getting kicked down the road. Seattle just put up a speed bump. More cities will follow.
Sources: The Verge — Amazon’s data centers used 2.5 billion gallons of water last year, Amazon’s water report, [Seattle data center moratorium]