Water, data centers, and AI

In short

• One chatbot prompt uses about 1.3 mL of water at this site's central estimate, about a quarter of a U.S. teaspoon. The range used here is 0.26 to 7.5 mL.
• On site, U.S. data centers consumed about 66 billion liters (17 billion gallons) of water in 2023, mostly evaporated in cooling. That's about 0.058% of the water consumed by the three largest uses in the lower 48 (2010 to 2020 average). The power plants that supplied their electricity consumed about 800 billion liters more.
• Water that evaporates from cooling returns as rain or snow, usually within days, but not necessarily to the basin it left.

The arithmetic is in the methodology.

Put a number in context

Enter a volume of water and the period it covers. The tool restates it in other units, converts it to a yearly amount, and compares that with reference values from this page: national ones first, then Texas ones, since both examples are in San Antonio. Each comparison uses the reference's own measure, so read the Reference column before comparing across rows.

What the figures measure

Data centers use water in two places. On site, cooling towers and evaporative coolers evaporate water to carry heat away from the servers. At the power plants that make their electricity, water evaporates in cooling systems and from hydropower reservoirs. Dry cooling uses little water on site but more electricity. Closed-loop systems vary: some still use water to chill the coolant.

The figures on this page measure different things:

Withdrawal

Water taken from a river, lake, aquifer, or utility. Some of it returns, for example as cooling-tower discharge or wastewater.

Consumption

Withdrawn water that doesn't return to its source, mostly because it evaporates. Data-center and power-plant figures on this page are consumption unless marked otherwise.

Evaporation

Water lost from a lake or reservoir surface. Gross evaporation counts all of it. Net evaporation subtracts the rain that falls on the lake.

Lifecycle footprint

All the water behind a product, from growing it to making it. For crops, most of it is rain stored in the soil. The irrigation share is the part comparable with cooling water.

Figures of different kinds aren't interchangeable, so the tables keep the kind beside each figure.

AI tasks and water

Water per task is the task's data-center electricity times a water factor per watt-hour: 1.15 mL for on-site cooling only (low), 4.29 mL with U.S. average power-plant water added (central), and 7.48 mL from Li et al.'s Arizona figures (high). One figure, Google's 0.26 mL for a median text prompt, was disclosed by a company. The rest are derived. Energy per task is on the AI energy page.

Show ranges

Published per-prompt figures

Most of the spread comes from three choices: whether power-plant water is counted, the energy assumed per prompt, and the location.

Household reference points

• One person's home use in the U.S., 82 gallons (310 liters) a day, equals the central water estimate for about 240,000 chatbot prompts.
• Twenty chatbot prompts a day at the central estimate come to about 26 mL a day, or 9.4 liters a year.
• A cup of coffee's irrigation water, about 1.3 liters, equals the central estimate for about 1,000 chatbot prompts.

Scale reference

Per day divides a yearly figure by 365.25. Gallons are U.S. gallons, 3.785 liters each. The published figure is under each name. For a sense of size, an Olympic pool at the minimum competition dimensions, 50 by 25 by 2 meters, holds 2,500 cubic meters: 660,000 gallons (2.5 million liters).

United States

1. The 2015 USGS compilation is the latest that covers every use. The 2010 to 2020 figures are newer USGS models of the three largest uses in the lower 48. USGS publishes 2020 estimates for the other uses as separate data sets, and no combined 2020 total was found.
2. No federal national total for reservoir evaporation was found. The 721-reservoir figure is a peer-reviewed estimate. It leaves out smaller reservoirs and natural lakes, and it ends in 2015.
3. The power-plant figure for U.S. data centers includes hydropower reservoir evaporation. Grid water factors that leave hydropower out are much lower.

Regional examples

Figures for one state, city, or lake.

1. Texas reservoir evaporation is a 2014 simulation with 1940s to 1990s hydrology. Newer reservoir-specific data shows evaporation rates rising about 1.1 inches a decade, but no statewide total from it has been published.
2. The two San Antonio figures cover different sites and periods: the reported figure covers two data centers over two calendar years, and Microsoft's covers its own datacenters from July 2024 to June 2025.
3. Lake Mead's volume is a two-year measurement from 2010 to 2012. Lake Powell's is a rounded estimate. Both lakes' areas change with their levels.

Product water footprints

These are lifecycle footprints: all the water used to grow and make one item, as global averages for 1996 to 2005. Most of it is rain stored in the soil where the crop grows. The irrigation share, water pumped or diverted and then consumed, is the part comparable with cooling water.

The water cycle

Evaporation moves water rather than destroying it. The open question is where the water falls again.

Taken together: water evaporated from a cooling tower or a reservoir returns as rain or snow within days, but mostly not to the basin it left. That is why hydrologists count evaporation as consumption, and why the same volume matters more in a dry basin than in a wet one.

Local context

Consumption matters most where water is scarce. In a basin with water to spare, evaporation from cooling has little local effect. In a water-stressed basin, the same volume is a larger share of what's available, and it can leave the basin as vapor. Figures for a state or a country say little about any one town.