Does AI really use water, and how?

A chip that computes heats up: almost all the electricity a server receives comes out as heat, which has to be removed from the room. To reject it into the outdoor air, operators have 2 main families of systems. Some reject it by warming the air, with fans and heat exchangers, like a car radiator. Others reject it by evaporating water, in cooling towers or by misting the incoming air.

Evaporation is effective because it absorbs a lot of heat: about 2,300 kJ per litre of water evaporated. If all the heat from one kWh of electricity were removed this way, about 1.6 litre of waterwould have to be evaporated, according to our calculation. This is an order of magnitude. Actual values can exceed it: the heat of the cooling equipment adds to that of the servers, and misting also cools the outdoor air. The LBNL simulations thus give 2 to more than 3 litres per kWh for water-cooled chillers without an economiser. The Lawrence Berkeley National Laboratory (LBNL), the reference body in the United States, sums up the trade-off: air-cooled chillers use no water but more energy; evaporative systems are generally more energy efficient.

In a cooling tower, the water that evaporates leaves its minerals in the water that remains. To prevent scale, the operator regularly discharges part of it, the blowdown, and replaces it with fresh water. The blowdown returns to the sewers or the environment; the vapour does not. According to Shaolei Ren's team at the University of California, Riverside, about 80% of the water withdrawn evaporates when the water is of good quality. At Google, data centres consumed 10,523 million gallons out of 13,562 withdrawn in 2025, or 78%; Microsoft, across all its sites, 8,170 megalitres out of 13,266 in its fiscal year 2025, 62%.

The difference between the 2 concepts explains part of the large gaps between figures. The same team projects for AI in 2027, using US electricity factors, a withdrawal of 4.2 to 6.6 billion m³, but a consumption of 0.38 to 0.60 billion m³, 11 times less. The withdrawal is inflated by the water that power plants pump and then return to the river. A water figure can only be compared with a figure of the same type, withdrawn or consumed.

Many data centres connect to the drinking water network of their municipality. In 2025, of the 13,562 million gallons withdrawn by Google's data centres, 81 % were drinking water, 12% non-potable water and 7% recycled wastewater. Microsoft buys almost all its water from suppliers, 13,192 of its 13,266 megalitres, without publishing the share of drinking water in that total. In France, Arcep notes that the volume withdrawn by data centres is almost entirely drinking water.

Operators who want to spare drinking water have 2 options: another source, recycled wastewater, industrial water or seawater, and better reuse of water in the tower. Seawater cools without withdrawing fresh water, but it is only possible on the coast; wastewater requires a nearby treatment plant and an agreement with it.

The water efficiency of a site is measured in litres per kWh consumed by the servers (WUE, for "water usage effectiveness"). Operators do not define it in the same way: Google and the LBNL report water consumed, Microsoft and AWS water withdrawn, part of which returns to the environment. The published values stay below the order of magnitude of 1.6 litre, because large sites use evaporation for only part of the year and cool the rest of the time with outdoor air.

The annual average hides 2 effects. In summer, when evaporation runs the most, is also the season when rivers and aquifers are at their lowest. And the US average is rising: the LBNL expects it to reach between 0.45 and 0.48 litre per kWh after 2023, with the growing share of very large data centres and liquid cooling of AI servers.

Thermal power plants, whether nuclear, coal or gas, are also cooled with water. They withdraw a lot of it and evaporate part of it. The LBNL estimates that in 2023, producing the electricity of US data centres consumed about 800 billion litres of water, 4.52 litres per kWh, against 66 billion on the sites themselves: about 12 times more. This calculation counts the water evaporated from hydroelectric dam reservoirs, a contested convention. A study from 2021, on US data from 2018, finds a lower ratio: site water there is about a quarter of the total. In France, Arcep also estimates that the water behind the electricity of data centres far exceeds their direct withdrawal, 575,000 m³ in 2024. Elsewhere, the ratio depends on the electricity mix and on how hydroelectricity is counted.

The water of a data centre therefore also depends on its electricity. A site powered by wind and photovoltaic has almost no indirect water; a site powered by thermal plants with towers has several litres per kWh. In France, cooling nuclear power plants consumes about 0.5 billion m³ per year, 12% of the water consumed in the country on average from 2010 to 2019, against 58% for agriculture, according to the SDES estimate. The same reasoning applies to carbon: our map of electricity by country shows how much the country of the servers changes the footprint.

Since August 2024, Microsoft has been designing its new data centres with liquid cooling as close as possible to the chips, in a closed circuit: the water, filled once at construction, circulates between the servers and chillers without evaporating. Microsoft estimates that each data centre will thus avoid more than 125 million litres of water per year, a figure calculated with its average withdrawal WUE, 0.30 litre per kWh. Microsoft also writes that replacing evaporation with mechanical cooling will increase the electricity consumed by its sites for the same computing power.

This additional electricity itself has indirect water, and a carbon cost, which depend on the grid. The choice between water and electricity is therefore made site by site: where water is scarce and electricity is low-carbon, cooling without water has the advantage; elsewhere, the answer depends on the 2 footprints. Our analysis of river basins shows that several contested projects chose air cooling.

Semiconductor factories use ultrapure water to rinse silicon wafers, and water to cool themselves. In Singapore, the average water recycling rate reaches 45% in wafer factories and 23% in semiconductor factories, cited by Shaolei Ren's team.

Few analyses attribute this water to a use of AI. Mistral AI's life cycle assessment is one: its 45 mL per response includes the water from manufacturing the servers, from electricity and from cooling, without the breakdown being published. The gap of about 170 with Google's 0.26 mL comes from these scopes, detailed in our analysis of the gaps ; Mistral AI does not publish the share of each route.