Andy Masley breaks down the numbers behind AI data‑center water use, showing that the sector consumes a tiny fraction of U.S. freshwater and that most of the impact comes from electricity generation, not the cooling systems themselves. He also points out where media coverage gets the story wrong and highlights how data centers can actually bring net benefits to water‑stressed communities.
The claim
Recent headlines have warned that AI‑driven data centers are “guzzling” water and threatening local supplies. The narrative usually follows a pattern: a big number (millions of gallons) is quoted, the water is described as potable, and the story links the usage to droughts or rising household bills.
What the data actually say
- National scale – In 2023 all U.S. data centers withdrew roughly 200–250 million gallons of freshwater per day. The United States uses about 132 billion gallons daily, so data centers accounted for 0.2 % of total consumptive use.
- AI’s share – AI workloads consume about 20 % of data‑center electricity. Assuming water use scales with power, AI’s direct water draw is roughly 0.04 % of the national total, or about 10.6 million gallons per day.
- Off‑site water – The majority of water linked to data‑center operations is used by nearby power plants to generate electricity (≈80 %). This water is typically non‑potable and returned to the source unchanged.
- Potable water – Only a small slice (≈3 %) of the water used on‑site is treated to drinking‑water standards. Converting raw water to potable water costs roughly $2–$5 per 1,000 gallons, meaning the extra cost of AI‑related potable water is on the order of $100,000 per day – a fraction of a single tech company’s revenue.
How the numbers get distorted
| Misleading tactic | What actually happens |
|---|---|
| Citing permits as usage – Permits list a maximum draw under worst‑case conditions. Real‑world consumption is far lower. | |
| Comparing to households – Saying AI uses “as much water as X million homes” ignores that household water is only ~8 % of an individual’s total footprint (most of it is embedded in food, manufacturing, etc.). | |
| Omitting indirect water – Some stories only show on‑site use, making the figure look tiny, then claim a “hidden” cost without clarifying that the hidden water is non‑potable electricity water. | |
| Scaling up without context – Multiplying a per‑prompt estimate (≈2 mL) by total prompts gives a large absolute number, but that still represents <0.001 % of a typical American’s daily water footprint. |
Local impacts are modest
The few places where data centers have been linked to higher water bills (e.g., Newton County, GA) show that the facility used ~2 % of the county’s water, while a pharmaceutical plant used ~4 %. In most high‑stress regions—Maricopa County, AZ; Loudoun County, VA—the data‑center share is 0.1–0.2 % of total withdrawals.
Economic upside
Data centers generate far more tax revenue per gallon of water than traditional users. In Maricopa County, water‑intensive golf courses consume 30 × more water than data centers but bring in only a fraction of the tax base. When a data center partners with a municipality, the added revenue often funds water‑system upgrades that benefit all residents (e.g., Google’s upgrades in The Dalles, OR; Microsoft’s reuse utility in Quincy, WA; AWS’s agricultural water‑reuse pipelines in Umatilla, OR).
Why the focus on water can be misleading
- Water vs. energy trade‑off – Switching from evaporative cooling to air cooling reduces water use but raises electricity demand, which has a larger climate impact.
- Value of the service – Even if a technology were socially undesirable, its water footprint must be judged against the total water use of comparable activities. A single AI prompt uses less water than the production of a pair of jeans.
- Infrastructure economics – In water‑abundant regions, adding a large, reliable water customer (a data center) improves economies of scale, lowering per‑unit costs for everyone. In scarcity‑driven areas, the marginal increase is still tiny compared to baseline demand.
Bottom line
AI does consume water, but the scale is minuscule compared with the overall U.S. water system and even compared with other industrial users. The real environmental pressure comes from electricity generation, not the cooling loops inside the data center. When data centers are sited responsibly and contribute tax revenue, they can help fund water‑system improvements rather than degrade water access.
“If you’re worried about AI’s water use, look at the numbers – the problem is orders of magnitude smaller than most people think.” – Andy Masley
Further reading
- Lawrence Berkeley National Lab, Data Center Energy and Water Use 2024 (PDF)
- Brian Potter, Water and Data Centers: An Update (blog)
- Hannah Ritchie, Global Water Use by Sector (Our World in Data)
- Matt Yglesias, Data Centers and Water (The Atlantic)
Comments
Please log in or register to join the discussion