Why Does AI Use Water?

AI uses water in two main ways. Data centers that run AI models use water to cool their computer chips, often by evaporating it in cooling towers. Power plants that generate the electricity those data centers need also use water for cooling. The software itself uses no water. The water use comes from the physical machines that run it and the power that keeps them on.

How AI uses water: the two main paths

Large AI models are trained, and usually run, on specialized chips in data centers. Those chips draw large amounts of electricity and turn much of it into heat. Removing that heat, and producing that electricity, is where the water comes in.

Direct water: cooling the data center

Direct water is the water a data center uses on site, mostly to keep its equipment from overheating.

A common design works like this:

1. Chips and servers heat the air or liquid around them.
2. A water loop carries that heat away from the equipment.
3. The warm water flows to a cooling tower.
4. Some of the water evaporates, and the evaporation carries the heat into the outside air.
5. The cooled water returns to pick up more heat, and fresh water is added to replace what evaporated.

The U.S. Department of Energy describes this process in its guidance for federal data centers. Evaporation sheds heat very effectively, which is why it is widely used and why a data center can consume large volumes of water.

Engineers measure on-site water efficiency with a metric called Water Usage Effectiveness, or WUE: liters of water used per kilowatt-hour of electricity consumed by the computing equipment. A lower WUE means less water for the same amount of computing.

Indirect water: generating the electricity

Indirect water is the water used somewhere else to produce the electricity a data center consumes.

Many power plants, including coal, natural gas and nuclear plants, boil water into steam to spin turbines. According to the U.S. Energy Information Administration, that steam must then be cooled back into water, using water drawn from nearby rivers, lakes or oceans. The more electricity a data center uses, the more of this power plant water sits behind it.

Indirect water can be larger than direct water. In one widely cited academic estimate, researchers at UC Riverside and UT Arlington calculated that for a GPT-3 request served from an average U.S. location, about 2.2 milliliters of water was used on site and about 14.7 milliliters was used off site to generate the electricity. The estimate assumed an energy use per request and describes an older model, but it shows why counting only on-site cooling water can miss most of the total.

Withdrawn water vs consumed water

Reports on AI and water consumption often mix up two measurements:

• Water withdrawal is water taken from a river, lake or underground source. Some of it is later returned.
• Water consumption is water that is not returned to the local water source, mainly because it evaporated.

The U.S. Geological Survey defines consumptive use as water that is evaporated or otherwise "removed from the immediate water environment." Evaporated water is not destroyed. It rises into the atmosphere and eventually falls again as rain or snow, but often far from where it left. For a community that depends on a local river or aquifer, that water is gone.

When you see a large number about AI and water usage, check which of these two it measures. Withdrawal figures are usually much bigger.

Training vs inference: when AI uses water

AI uses water at two stages, because it uses electricity at two stages.

Training is the process of building a model by running it over huge amounts of data, often for weeks or months on thousands of chips. The same UC Riverside and UT Arlington researchers estimated that training GPT-3 in Microsoft's U.S. data centers could have consumed about 5.4 million liters of water, including about 700,000 liters consumed on site. That is a modeled estimate for a model released in 2020, not a figure disclosed by the companies, and training footprints for newer models are rarely published. Training mostly happens before release, although development involves many experimental training runs and models are updated over time.

Inference is what happens every time someone uses a finished model: asking a chatbot a question, generating an image, or summarizing a document. Each request uses a small amount of electricity, but popular AI products handle enormous numbers of requests. Research from Google and Meta published before the current wave of chatbots found that running models for users often uses more energy than training them. Meta researchers reported in 2021 that Facebook's AI infrastructure power capacity was split roughly 10% experimentation, 20% training and 70% inference.

Why estimates of AI water use vary so widely

Two data centers running the same AI model can have very different water footprints, for several reasons.

Cooling technology. Evaporative cooling towers consume the most water on site. Air-cooled systems and dry coolers, which release heat directly into outside air, use far less. Newer designs circulate liquid in a sealed loop to cool the chips directly. Microsoft said in December 2024 that its newest data center design "consumes zero water for cooling," with pilot sites planned for 2026. The design applies to new sites, not existing buildings, and still uses water for things like restrooms.

The energy and water trade-off. Using less water often means using more electricity. Microsoft said its closed-loop design will raise the energy used for cooling. Google said in 2022 that its water-cooled data centers use about 10% less energy than many air-cooled ones. Because electricity carries its own indirect water, the lowest-water choice on site is not always the lowest-water choice overall.

Climate and season. Hot, dry conditions increase evaporation. Cooler climates let data centers use outside air for cooling more of the year. In the GPT-3 estimate above, the water per request varied about sevenfold depending on location, reflecting both local climate and the local power grid.

The local power grid. Power plants that turn water into steam need cooling water, so the mix of plants on the local grid affects how much indirect water sits behind each unit of electricity.

Water stress. A liter used in a wet region has a different local impact than a liter used where water is scarce. A 2021 peer-reviewed study, using data from before the recent AI build-out, estimated that one-fifth of U.S. data center servers' direct water footprint came from moderately to highly water-stressed watersheds.

The model and the task. Larger models and heavier tasks, such as long reasoning, image generation or video, generally need more computing than a short text reply. Mistral AI reported in 2025 that a model 10 times bigger has roughly 10 times the impact for the same amount of generated text.

How much water does AI use per prompt?

There is no single number for water per prompt. Published figures differ by more than a hundredfold, largely because they measure different things and assume different models, hardware, locations and cooling systems.

A figure that counts only on-site cooling will look much smaller than one that includes power plant water or, in a lifecycle estimate, the water used to manufacture chips and servers. A claim about water per prompt means little until you know what it includes.

Does AI use a lot of water?

A single AI request uses a small amount of water. The totals matter because of scale and because the water use is concentrated in specific places.

Google reported consuming 10.9 billion gallons (41 billion liters) of water across its data centers and offices in 2025, up 34% from 2024. Microsoft reported consuming 8,170 megaliters company-wide in its 2025 fiscal year, up from 6,693 megaliters the year before. Both figures cover all company operations, not only AI.

Across the United States, the Environmental Law Institute, citing Lawrence Berkeley National Laboratory's 2024 report, says U.S. data centers directly consumed about 66 billion liters of water in 2023, up from 21.2 billion liters in 2014. It adds that producing their electricity consumed about 800 billion liters indirectly in 2023. Those figures cover all data centers, not just AI.

Whether that amounts to "a lot" depends on location. The same volume can be manageable where water is plentiful and a real strain in a drought-prone region.

Some companies fund water replenishment projects to offset their use. Google, for example, says its projects replenished roughly 78% of its 2025 freshwater consumption. Replenishment projects can be in different places from the data centers that consume the water, so they are not the same as using less water locally.

Where data center water comes from

Many data centers use clean freshwater, often drinking-quality water from a municipal supply. The researchers behind the GPT-3 estimate explain that clean water helps avoid clogged pipes and bacterial growth in cooling systems.

Water quality matters because evaporation leaves minerals behind. The Department of Energy explains that as cooling tower water evaporates, dissolved solids such as calcium, magnesium and silica build up. Too much buildup causes scale and corrosion, so operators regularly drain off some concentrated water and add fresh water. Cooling water is also commonly treated with scale, corrosion and biological growth inhibitors.

Some data centers use other sources:

• Recycled water. Google says it cools its data center in Douglas County, Georgia, with recycled water, a practice it began there in 2012.
• Seawater. Google's data center in Hamina, Finland, uses a cooling system that draws seawater from the Gulf of Finland.
• Reclaimed or non-potable water. Microsoft says it has expanded its use of reclaimed and recycled water in Texas, Washington, California and Singapore.

Common misconceptions about AI and water

"AI drinks water." AI models are software. The water is used by the cooling systems of the buildings that run the software and by the power plants that supply their electricity.

"Every prompt uses a bottle of water." This comes from the academic estimate that GPT-3 needed about a 500 mL bottle of water for roughly 10 to 50 medium-length responses, depending on when and where it ran. It covered an older model, included power plant water, and was never meant as a fixed rule for today's products.

"The numbers are settled." They are not. Lawrence Berkeley National Laboratory has said a lack of available data significantly limits its analysis of data center energy use. The UN Environment Programme has said there is a dearth of reliable information on AI's environmental impact.

"Efficiency will solve it." Efficiency is improving fast. The International Energy Agency reported in April 2026 that energy use per AI task has been dropping by at least an order of magnitude a year. Total demand is still rising, though: the IEA found data center electricity use grew 17% in 2025. More efficient requests do not guarantee lower totals when usage grows faster.

Water is one part of AI's environmental footprint. Electricity, carbon emissions and hardware are covered in How Does AI Affect the Environment?. For how training and inference fit into the way AI systems work, see How Does AI Work?.