Water: the hidden cost of AI
Computers turn electricity into heat, and one of the cheapest ways to get rid of heat is to evaporate water. That puts data centers in competition with farms and towns, sometimes in the driest places in America.
Plain-English explainer by Joe Miller, a nuclear engineer. Not investment advice. Figures are the latest public numbers we could find and change fast; sources are at the bottom of each page.
How data centers use water
- Direct use, for cooling: many data centers use evaporative cooling towers, like a giant swamp cooler. Water that evaporates is gone from the local supply.
- Indirect use, at the power plant: coal, gas and nuclear plants also use water for cooling, so the electricity a data center buys carries a hidden water cost. Studies find this indirect water is often larger than the direct use.
- Making the chips: chip factories use large amounts of ultra-pure water, but that happens far away, mostly in Taiwan and Arizona.
How much water
A single large campus with evaporative cooling can use 1 to 5 million gallons a day in summer. One proposed campus near Scranton, Pennsylvania, could have drawn up to 3.3 million gallons a day from a local lake.
How much water per AI question?
Much less than the scary headlines say. Google reported that a typical Gemini text question uses about 0.26 milliliters of water, about five drops, and 0.24 watt-hours of electricity. OpenAI's CEO gave a similar figure for ChatGPT. Critics point out that these numbers leave out the water at the power plant, which can be several times more. Either way, the issue isn't your one question; it's billions of questions a day, plus training, concentrated in a few towns.
Where it hurts
About 20% of U.S. data centers were already in water-stressed areas five years ago, and the boom is pushing into dry, cheap-land places like Arizona, West Texas and Nevada. In wet regions like Virginia or Ohio, water is less of a worry than electricity.
The local fights
| Place | What happened |
|---|---|
| Tucson, Arizona | City council killed "Project Blue" in August 2025 after residents objected to its water and power use. |
| Lackawanna County, Pennsylvania | Residents fought a campus that could draw 3.3 million gallons a day from Lake Scranton. |
| California | Bills to make data centers report their water use; the governor vetoed one in 2025, and two more came up in 2026. |
| Pima County, Arizona | 120-day pause on new data centers (September 2026). |
| Nationwide | In a Gallup poll, about 18% of people opposed to local data centers named water as their reason. |
Using less water
- Closed-loop and "zero-water" cooling: liquid circulates in sealed pipes and dumps heat through fans, like a car radiator. Microsoft has said its new designs use almost no water for cooling, at the cost of more electricity.
- Direct-to-chip liquid cooling: needed anyway for the hottest AI racks and more efficient than air.
- Recycled water: treated wastewater instead of drinking water.
- Cooler places: building where outside air does most of the cooling.
- Cleaner power: wind and solar use almost no water, which cuts the hidden indirect use.
Questions to ask if one is coming to your town
- How many gallons a day, at peak summer, and from which source?
- Will it use drinking water, groundwater, or recycled water?
- What happens in a drought? Who gets cut first, the data center or the town?
- Will the company report its water use every year, publicly?
Sources
Checked November 16, 2026.
- Utility Dive: U.S. data centers could 4x water use by 2028 (July 21, 2026)
- Lawrence Berkeley National Laboratory: 2024 U.S. Data Center Energy Usage Report
- Google: Measuring the environmental impact of delivering AI at Google scale (Aug. 2025)
- The Register: Google games numbers to make AI look less thirsty (Aug. 22, 2025)
- CalMatters: California data center water transparency bills (Aug. 2026)
- LLM Pulse: U.S. data center moratoriums and blocked projects (2026)