Every conversation about AI data centers eventually reaches water. The problem is that several different water stories are often compressed into one alarming number.
There is water inside cooling loops. There can be water evaporated from cooling towers. There is water withdrawn from a utility or well, and a different amount actually consumed instead of returned. There is also water used elsewhere to generate electricity for the facility. Those are related questions, but they are not interchangeable.
If we want to understand the tradeoffs, we have to follow the heat.
One modeled comparison shows how much the cooling system matters. EPRI's May 2025 brief, Water Usage in Data Centers, models 150 MW facilities three ways:
| Cooling approach | Site WUE | Modeled water use per day | Per year |
|---|---|---|---|
| Average U.S. efficiency, 2023 aggregate | 0.36 L/kWh | ~340,000 gallons | ~120 million gallons |
| Water-cooled chillers | 2.8 L/kWh | ~2.7 million gallons | ~970 million gallons |
| IT liquid cooling with dry coolers | 0.2 L/kWh | ~190,000 gallons | ~69 million gallons |
These are modeled scenarios for similar-sized 150 MW facilities, counted as direct on-site consumption — water used for cooling and humidification at the site, not water used to generate its electricity. They are not measurements of any real facility, and EPRI notes the estimates vary with climate and site-specific factors. The 0.36 figure is Berkeley Lab's modeled aggregate U.S. average — just over 0.36 litres per kilowatt-hour for 2023 — not a precise efficiency for any single facility, and these examples are illustrations, not a universal range. EPRI describes each example as a 150 MW data center and does not state whether that is IT load or total facility load, or what utilization it assumes — the figures above are EPRI's stated modeled results.
The heat has to complete a journey
AI processors turn electricity into heat while they work. The first job of a cooling system is to pick up that heat before the hardware gets too hot. In a traditional air-cooled room, fans move cool air through the servers and carry warm air away. In a direct-to-chip system, a liquid loop brings cold plates close to the processors and collects the heat more directly. That shift is driven by density — see Inside a 120 kW AI rack: power and cooling, on what one 120 kW rack demands of the building around it.
That is only the first leg. The warm liquid usually transfers its heat through an exchanger into another loop serving the building. Pumps then move the heat toward chillers, dry coolers, cooling towers, or a combination of equipment outdoors.
To see that journey equipment by equipment, see Following the heat through a liquid-cooled data center.
This is why seeing pipes at the rack does not tell you how much water a site consumes. A sealed loop can circulate the same coolant repeatedly. It is like the coolant in a car radiator: the liquid carries heat, but the system does not intentionally throw away the entire tank on every trip.
The major water question is often at the last step—how the facility rejects heat to the outdoor environment.
Evaporation is efficient, but the water leaves
A cooling tower uses evaporation to remove heat. A small portion of the circulating water becomes vapor and carries energy into the atmosphere. Additional water is discharged as “blowdown” so minerals do not build up indefinitely, and fresh make-up water replaces what was lost.
Evaporation can be very effective, especially when the outdoor air is suitable. It can reduce the electricity required to reject heat compared with relying entirely on mechanical refrigeration. But the evaporated water is consumed from the local water system’s point of view; it does not immediately return to the same watershed as liquid.
Dry coolers work more like a large outdoor radiator. Fans push air across heat-exchanger surfaces without deliberately evaporating a continuous supply of water. They can sharply reduce on-site water use, but may require more fan power, more equipment area, or higher operating temperatures. Performance also changes with climate.
Hybrid systems can switch between modes. Some use water only during the hottest periods. Others combine chillers, towers, and dry coolers to balance energy, water, cost, and reliability.
There is no single cooling diagram that represents every data center. Even two facilities owned by the same company may make different choices because Phoenix, Seattle, and Northern Virginia do not offer the same air temperature, humidity, water availability, or electric grid.
Withdrawal and consumption answer different questions
Water withdrawal is the amount taken from a source. Water consumption is the portion not returned promptly to that source, often because it evaporated or became part of another product or process.
A facility can withdraw a large volume and return much of it, or withdraw less and consume a high share. Both measurements matter. Withdrawal can stress pipes, treatment plants, wells, and stream flows. Consumption changes how much water remains locally available.
The source matters too. Potable drinking water, reclaimed wastewater, and non-potable industrial supplies do not carry the same community tradeoff. The U.S. Department of Energy recommends examining alternative make-up sources for cooling towers, including condensate and suitably treated recycled water, where those options are technically and locally appropriate. Often, yes. EPRI states that most data centers today use potable — drinking — water for cooling. One company's reporting illustrates how uneven the alternative is: in the United States, only one of Google's 18 data center locations used reclaimed wastewater in 2023. That is one company in one year, not the whole industry. Reclaimed or recycled water is used at some sites, but its use varies by location. So the question is not only how much water a site uses, but which water — potable, reclaimed, or private groundwater — and what else depends on that source.
A gallon also means something different in a water-rich region than in a drought-prone basin. National totals can describe scale, but water conflict is usually local. The relevant questions are which source supplies the site, when demand peaks, what other users depend on that source, and how the system behaves during drought.
Saving water can move the burden to electricity
Cooling is full of tradeoffs. A design that uses less on-site water can require more electricity. That additional electricity may have its own water footprint at power plants, depending on the regional generation mix. A design that lowers energy use through evaporation can consume more water at the facility.
Neither outcome automatically makes one system good and the other bad. It means a serious comparison needs clear boundaries. Are we counting only water at the data center fence line? Are we including water used to generate electricity? Are we comparing annual totals or the hottest afternoon of the year?
Water Usage Effectiveness, or WUE, is the standard way to compare sites. It is the litres of water a site consumes in a year divided by the kilowatt-hours of energy its IT equipment uses in that year. A lower number means less site water per unit of computing energy. Berkeley Lab modeled the aggregate U.S. average site WUE at just over 0.36 litres per kilowatt-hour in 2023. Individual cooling designs span a much wider range: EPRI reports site WUE values from 0.1 to 9.0 litres per kilowatt-hour across sources, and models a water-cooled chiller plant at 2.8 against IT liquid cooling with dry coolers at 0.2 for similar-sized 150 MW facilities. WUE has two limits worth remembering. First, site WUE counts only water at the facility — it says nothing about the water used to generate the facility's electricity, which is a separate calculation (source WUE) and depends on where the power comes from. Second, it says nothing about which water is used: a low WUE achieved with drinking water is a different community tradeoff from the same number achieved with reclaimed water. Compare WUE only between sites measured over the same period, with the same boundary — and, as Berkeley Lab notes, a low WUE is not automatically good, because the cooling system that uses the least water may use the most electricity.
For a related efficiency measure, see PUE: what an efficiency score reveals—and misses. PUE is an energy ratio and WUE is a water ratio — their boundaries and denominators differ, and neither ratio by itself tells you a facility's total resource use.
The honest answer to “How much water does an AI data center use?” is usually another question: which data center, using which cooling system, in which climate, measured at which boundary?
The questions communities should ask
A developer saying “closed loop” is not the end of the conversation. The rack loop may be closed while an outdoor cooling tower still evaporates water. Likewise, “air cooled” can describe the server side while a chilled-water plant serves the room.
Useful questions are more specific:
What is the maximum and expected annual water demand? How much is withdrawal versus consumption? Is the source potable, reclaimed, or private? Which cooling mode operates on normal days, and which mode operates during peak heat? What happens during drought restrictions? Are estimates based on the first building or the full planned campus?
Those answers can be reported without exposing proprietary computing details. They let a community evaluate real infrastructure instead of arguing over labels.
Liquid cooling is becoming more important because AI hardware is becoming denser. That can make heat collection inside the building more efficient. It does not, by itself, determine the site’s water use. The heat still has to reach the outdoors, and the design of that final handoff is where much of the water story lives.
AI is not weightless. But its water footprint is not one universal number either.
Sources and further reading
- U.S. Department of Energy: Cooling water efficiency opportunities for federal data centers
- U.S. Department of Energy: Cooling tower management
- Berkeley Lab: U.S. data center energy and water modeling
- Berkeley Lab: United States Data Center Energy Usage Report — 2025 Update
- Electric Power Research Institute, Water Usage in Data Centers, Technology Innovation Spotlight, May 2025 (EPRI product 3002033251)
- Shehabi, A. et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, December 2024 (LBNL-2001637)

