How Much Water Does a Data Center Use, and Why?

The short answer

A data center uses water in two places: on its own site, mostly to reject heat, and at the power stations that make its electricity. The second is usually the larger of the two. The International Energy Agency’s most-quoted benchmark puts it in one line: an average 100 MW hyperscale data center in the United States consumes around 2 million litres a day in total, equivalent to about 6,500 households, and over 60% of that is indirect water used for its electricity.

On site, the figure that matters is water usage effectiveness, or WUE: litres of water per kilowatt-hour of energy delivered to the servers. Lawrence Berkeley National Laboratory (LBNL) estimates the US average at about 0.36 L/kWh in 2023. At that rate, every megawatt of IT load running flat out for a day evaporates about 8,640 litres on site, and a 100 MW campus about 864,000 litres. Those are arithmetic, not measurements. Real sites run below full load, and the spread between cooling designs is wider than the spread between sizes.

Across the industry, the IEA estimates that data centers worldwide consume around 560 billion litres of water a year, counting cooling, electricity supply and chip manufacturing, and that this could rise to around 1,200 billion litres a year by 2030 in its base case. The rest of this guide explains where those numbers come from, why some sites consume almost nothing on site while others consume several times the average, and where water has already become a condition for getting a permit.

Why do data centers need water?

Every watt a server draws ends up as heat. A rack of conventional servers draws several kilowatts; a rack of AI accelerators can draw many times that. That heat has to leave the building continuously, because chips throttle and then fail as they get hotter. Cooling is therefore not an accessory in a data center. It is the second-largest system after the electrical one, and it decides most of the water story.

There are only a few ways to move heat out of a building at that scale. You can blow outside air across the equipment when the weather allows. You can run refrigeration machines (chillers) that dump heat into the outdoor air through fans and coils. Or you can let water evaporate. Evaporation is the reason water is so useful: when water turns into vapour it carries a large amount of heat away with it, so a cooling tower can reject a building’s heat with much less fan and compressor power than a dry system. Data centers that consume a lot of water are usually doing so to save electricity.

Water enters a data center’s cooling system in several forms:

• Cooling towers. Warm water from the building’s chillers or heat exchangers is sprayed over a fill material while fans pull air through it. Part of the water evaporates, which cools the rest. The evaporated part is gone, and some extra water is regularly drained off (“blowdown”) because evaporation concentrates minerals that would otherwise scale the equipment. Both losses have to be replaced with fresh make-up water.
• Adiabatic and evaporative coolers. On hot days, water is sprayed onto pads or into the air stream that cools the building, lowering the air temperature by evaporation before it reaches coils or servers. These systems use little or no water in cool weather and a lot during heatwaves.
• Humidification. Very dry air raises the risk of static discharge, so some halls add moisture in winter. It is a small use compared with cooling, but LBNL notes it is often the main reason for what water consumption remains in otherwise dry designs.
• Domestic use. Toilets, kitchens and landscaping. Microsoft’s own description of its zero-water cooling design is careful to say that water for these purposes still continues.

So the plain answer to why data centers use so much water is this: they produce heat on the scale of a small industrial plant, and evaporation is the cheapest way to get rid of it. The plain answer to why some use very little is that they chose to pay for the alternative in electricity, or they are in a climate cold enough that outside air does most of the work.

Consumption, withdrawal, site and source

Water numbers about data centers are easy to misread because four different quantities are quoted under the same word.

Withdrawal is all the water taken from a river, aquifer or utility. Consumption is the part of that withdrawal that is not returned to its source, mostly because it evaporated. The IEA uses exactly these definitions in its report, and its charts show global withdrawals for data centers running at several times consumption. A once-through cooling system that pumps water from a river and returns it slightly warmer withdraws a lot and consumes little. A cooling tower withdraws less but consumes most of what it takes.

Site water is what the data center itself uses. Source water, often called indirect water, is the water consumed to generate the electricity the data center buys. LBNL calls these “WUE (site)” and “WUE (source)”. Its 2024 report stresses that including both reflects the true water cost of a data center, but that the source figure is complex to calculate and highly dependent on where the electricity comes from. A data center supplied mainly by wind and solar has a much smaller indirect footprint than one supplied by thermal power stations, because thermal plants evaporate water in their own cooling towers. The IEA makes the same point from the other direction: solar PV and wind use one-hundredth of the water that fossil sources use, or less.

A third layer sits further upstream. Making chips needs large volumes of ultra-pure water, and the IEA counts that too. In its accounting of global data center water consumption in 2023, about two-thirds was associated with primary energy supply and electricity generation, about a quarter with direct cooling at the data centers, and the remainder with semiconductor and chip manufacturing.

When a company, a campaigner or a county official quotes a single number, the first question is which of these it is. A claim of zero water usually means zero water for cooling, consumed on site. A claim that a data center “uses” millions of litres may be withdrawal, consumption, or a total that includes electricity. None of these is dishonest in itself. They are different measurements, and they should not be compared with each other.

How water use is measured: WUE

The industry metric is water usage effectiveness, proposed by The Green Grid in its white paper number 35 in March 2011 as a companion to power usage effectiveness (PUE). WUE is the annual site water use, in litres, divided by the annual energy used by the IT equipment, in kilowatt-hours. The unit is litres per kilowatt-hour (L/kWh). A WUE of zero means no water used on site for the period measured.

Some published values give a sense of the range:

• US average, 2023: about 0.36 L/kWh. This is LBNL’s modelled average across all US data centers. LBNL expects it to rise to between 0.45 and 0.48 L/kWh by 2028, as more of the load moves into hyperscale and colocation sites and into liquid-cooled AI systems that still reject heat through evaporative plant.
• Microsoft fleet, fiscal year 2024: 0.30 L/kWh, down from 0.49 L/kWh in 2021, according to the company’s December 2024 announcement. At 0.30 L/kWh, a 100 MW load running for a full day would consume about 720,000 litres on site.
• Microsoft Asia Pacific, direct cooling: 1.65 L/kWh, more than three times the company’s then-global average of around 0.5 L/kWh. The IEA cites this figure to show how much climate matters: in a warm, humid region, the same design consumes far more water. At 1.65 L/kWh, the same 100 MW day would mean about 3,960,000 litres.
• Some operators’ global fleets: 1 to 1.5 L/kWh. The IEA notes that several operators have reported values in this range for their global operations, well above the US modelled average.
• Optimised hyperscale designs: 0.1 to 0.3 L/kWh. LBNL notes that some hyperscale facilities report values in this range for airside economiser systems with occasional adiabatic cooling.

Europe has set its first reference point. The Climate Neutral Data Centre Pact, signed by 97 operators and associations according to the IEA, targets a WUE below 0.4 L/kWh by 2040. Since 2024, EU Delegated Regulation 2024/1364 also requires operators of data centers with an installed IT power demand of at least 500 kW to report their energy and water consumption to a European database, which makes water the first metric beyond energy that large European sites must disclose in a common format.

WUE has two weaknesses worth knowing. It ignores indirect water, so a site can lower its WUE by switching to air-cooled chillers and buying more electricity, which may raise total water use if the grid is thermal. And disclosure is patchy. The IEA observes that WUE appears in company reports less often than comparable energy and emissions metrics, which increases the uncertainty in every industry-wide estimate, including its own.

How much water a data center uses per day

Per-day questions need a size and a design. The simplest way to answer them is arithmetic on WUE: power in kilowatts, times 24 hours, times litres per kilowatt-hour.

At the 2023 US average site WUE of 0.36 L/kWh:

• a 1 MW enterprise hall running flat out would consume about 8,640 litres a day on site;
• a 100 MW hyperscale campus would consume about 864,000 litres a day on site.

Add LBNL’s 2023 national average for indirect water, 4.52 L/kWh of electricity used by data centers, and the same 100 MW campus would carry about 10,848,000 litres a day of water consumed at power stations. That is far above the IEA’s benchmark of around 2 million litres a day in total for an average 100 MW US hyperscale site. The gap is not a contradiction to be averaged away. It shows how sensitive indirect estimates are to method. LBNL’s indirect figure is built from the actual grid mix around US data center locations, and its definition of consumption includes evaporation from hydroelectric reservoirs. The IEA’s figure is an average across the cooling technologies in use, and its box does not publish the electricity water intensity it assumes. Both numbers say the same thing in different units: for a typical US site, more water is consumed making the electricity than cooling the servers.

These per-day figures assume full load. Real data centers rarely run at their full rated capacity all year, and water use moves with the weather: an evaporative system can use most of its annual water in the hottest months and very little in winter. The arithmetic gives a ceiling and an order of magnitude, which is what most people asking the question want. For a specific site, the only reliable answer is the operator’s own metered figure, or the volume in its water supply agreement, where either has been published.

US data center water use in numbers

LBNL’s 2024 report for the US Department of Energy is the most detailed national estimate available. Its water findings:

• Direct consumption in 2014: 21.2 billion litres, with 64% in internal (enterprise) data centers.
• Direct consumption in 2023: about 66 billion litres (roughly 17 billion US gallons). Hyperscale and colocation sites accounted for 84% of it, while internal data centers fell to 12%, which LBNL attributes to water efficiency improvements in that segment.
• 2028: hyperscale data centers alone are expected to consume between 60 and 124 billion litres, and internal data centers are expected to fall to 2% of the total.
• Indirect consumption in 2023: nearly 800 billion litres, through the roughly 176 TWh of electricity US data centers used, at a national average of 4.52 L/kWh.

Two features of this series are worth noticing. The first is that the shift from enterprise rooms to large campuses raised average WUE rather than lowering it. Large sites are more energy-efficient (LBNL’s average PUE falls from 1.6 in 2014 to 1.4 in 2023), and many of them achieve that efficiency with evaporative cooling. The second is that indirect water is more than ten times direct water at the national level. Most of the water cost of a US data center is decided by the grid it connects to, not by its cooling plant.

For scale, the IEA notes that in the United States, withdrawals for data centers today equate to less than 10% of annual municipal water withdrawals. That national share hides concentrated local effects, which is why the debate happens county by county. Loudoun County in Northern Virginia, at the heart of the metro with the most facilities in the atlas, publishes its own figure: on an average day, data centers account for 9% of total potable water use in the county. The atlas records 95 facilities in the Northern Virginia metro alone; see them on the map.

Global data center water use

The IEA’s global estimate, published in April 2025 in its Energy and AI report, is around 560 billion litres of water consumption per year, rising to around 1,200 billion litres a year by 2030 in its base case. Withdrawals are much larger and, in the IEA’s words, show a similar steep increase to 2030.

Three details in the IEA’s analysis change how the global number should be read. First, the composition: in 2023 roughly two-thirds of consumption was tied to energy supply and electricity generation, a quarter to direct cooling and the rest to chip manufacturing. Second, the drivers diverge. Water tied to electricity grows more slowly than data center electricity demand, because more of that electricity comes from wind and solar. Water tied to chips grows faster: the IEA expects water consumption from chip manufacturing for data centers to grow more than 50% from 2023 levels to around 70 billion litres in 2030, because accelerated servers contain more chips than conventional ones. Third, geography: around half of consumption by 2030 is in Asia Pacific, where a typically warm and humid climate makes cooling more water-intensive.

The IEA also notes that a few operators, including AWS and Google, have pledged to be “water positive” by 2030, by combining recycling and replenishment projects with reductions in the direct WUE of their sites. Those pledges matter for corporate accounting, but replenishing water in one watershed does not reduce consumption in another. A county weighing a new campus is right to ask about the local draw, not the global balance.

Cooling design decides most of the water

Two data centers with the same megawatts can differ several times in water use because of the cooling system behind the racks. LBNL’s modelling compares the main designs, and the ranking is consistent:

• Water-cooled chillers without economisers have the highest site WUE, largely because of cooling tower water use.
• Waterside economisers (water-cooled chillers that switch off their compressors when the outdoor wet-bulb temperature allows) reduce water use somewhat, and more so in liquid-cooled halls that can run warmer coolant.
• Airside economisers, which cool with filtered outdoor air when the weather allows and shut down the chilled-water plant, save a large amount of water. Adding adiabatic cooling for hot days keeps water use low but not zero. LBNL notes that this combination is prevalent among hyperscale data centers.
• Dry coolers with adiabatic assist consume water mainly when their wet mode switches on in hot weather.
• Air-cooled chillers and other waterless systems consume almost no water on site, mostly for humidification, but use more energy.

The IEA describes the same contrast with two designs that dominate opposite ends of the market: direct expansion cooling, which accounts for around four-fifths of cooling in enterprise data centers, is many times less water-intensive than airside economiser and adiabatic cooling with water-cooled chillers, used in about half of hyperscale data centers.

The trade-off is explicit in LBNL’s report: low site WUE is not automatically good. Water-cooled chillers and other evaporative systems are generally more energy-efficient than air-cooled chillers, so cutting site water can raise electricity use and, on a thermal grid, raise indirect water and emissions. Microsoft states the trade-off in its own announcement of zero-water cooling: compared with evaporative designs, the new approach brings what it calls a nominal increase in annual energy use across its fleet.

The atlas’s own records show where new capacity is heading. Liquid cooling is recorded for 26.8% of the facilities under construction and 24.1% of announced projects, against 2.8% of operational facilities. The field is a floor, because many sources do not describe cooling at all, but the direction is clear. Liquid cooling by itself does not decide water use. Direct-to-chip loops still hand their heat to a facility system, which may be an evaporative tower or a dry cooler. What liquid cooling changes is the temperature of that heat. Warmer coolant leaves room for dry coolers and economisers for more hours of the year, which is how Microsoft’s design removes evaporation entirely and why LBNL’s model shows lower WUE for liquid-cooled halls on waterside economisers.

AI and water

AI did not invent the water question, but it changed its scale. Accelerator racks concentrate far more heat in the same floor area, and the campuses being built for them are measured in hundreds of megawatts and, in a growing number of announcements, gigawatts. More heat at a site means more water at that site unless the cooling design avoids evaporation, and more electricity means more indirect water unless the grid is clean.

The most cited academic estimate of AI’s water footprint is the paper “Making AI Less Thirsty”. In it, first posted in April 2023 and revised in March 2025, Pengfei Li, Jianyi Yang, Mohammad Islam and Shaolei Ren estimate that training GPT-3 in Microsoft’s US data centers directly evaporated about 700,000 litres of clean freshwater, and that global AI demand could account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027. Those are model estimates built on published efficiency figures, and the authors present them as such; they are useful mainly as an order of magnitude and as an argument for disclosure.

The IEA adds a less discussed channel: chips. Accelerated servers contain more chips than conventional ones, and almost 90% of the water consumed in chip manufacturing goes into producing ultra-pure water. That part of AI’s footprint is incurred where the chips are made, often far from the data center that runs them. In Chinese Taipei, the IEA recalls, semiconductor manufacturers were subject to water restrictions during a drought in 2021.

For a reader trying to judge one AI campus, the practical points are the same as for any data center. Ask for the cooling design and the expected WUE, the source of the water (potable, reclaimed, groundwater, surface water), and the grid mix. A gigawatt-scale campus with closed-loop liquid cooling and dry coolers can consume little water on site; a smaller campus with evaporative towers in a hot climate can consume much more. Size alone predicts electricity, not water.

Where water now decides permits

Water has moved from sustainability reports into planning law. The atlas tracks 61 moratoria, zoning rules, permit regimes and grid limits that apply to data centers; 6 of them, in 4 countries, cite water directly.

The pattern in these rules is consistent. Authorities rarely ban data centers outright on water grounds. They attach conditions: an alternative to drinking water, a conservation plan, a cap with surcharges, or a public review before a large user can connect.

• Johor, Malaysia. The state set up a committee in 2024 to vet data center applications against water and power criteria, and in November 2025 said it would stop approving Tier 1 and Tier 2 data centers because of their water consumption. New projects must source reclaimed water and pass several levels of vetting.
• Marana, Arizona. The town’s ordinance bars Marana Water from supplying drinking water for data center cooling or humidity control, so developers must show another source, alongside utility evidence of electrical capacity and a public rezoning.
• Tucson, Arizona. After the Project Blue data center was rejected, the city adopted rules in August 2025 that make customers expecting to use more than about 7.4 million gallons a month apply to the city with a conservation plan, with surcharges above 120% of the approved volume.
• Singapore and Thailand. Singapore’s 2019 pause on new data centers cited land, water and energy, and ended in 2022 with a pilot application scheme. Thailand’s data center policy board paused 166 projects in September 2026 while it writes standards that include water and electricity use.

The US metro table shows why the American debate concentrates in a few places. Phoenix, with 36 facilities in the atlas and 5 projects under construction or announced, sits in a desert where evaporative cooling is efficient but every litre is contested, which is why Arizona towns were among the first to write data-center-specific water rules. Northern Virginia has more facilities than any other metro in the atlas and handles water differently: through a reclaimed water network rather than prohibitions.

Salt water, reclaimed water and other alternatives

The question of why data centers cannot simply use salt water comes up often, and the answer is that some do, but not in the way people imagine. Seawater corrodes pipes, pumps and heat exchangers, fouls surfaces with marine growth, and leaves salt behind when it evaporates, so it cannot replace fresh make-up water in an ordinary cooling tower. What works is keeping seawater in a separate loop. Google’s data center in Hamina, Finland, built on the site of a paper mill the company bought in 2009 and opened in 2011, uses a cooling system with seawater from the Bay of Finland, according to Google’s own site page. When the system was announced in 2011, the seawater was described as arriving through the paper mill’s old intake tunnels, passing through heat exchangers that take up the servers’ heat, and being mixed with cold water before it went back to the sea so that it returned at close to the sea’s own temperature. That design depends on a cold sea next door, existing intake infrastructure and environmental permits for the return water, which is why it remains rare.

Reclaimed water is the more common substitute. Treated wastewater that is not safe to drink is perfectly usable in cooling towers after further treatment. Loudoun County reports that Loudoun Water’s reclaimed system delivered more than 750 million gallons of reclaimed water to roughly 40 customers in 2025, much of it to data centers. Johor now requires reclaimed water for new projects, and Marana’s ban on potable water pushes developers the same way. Reclaimed water does not reduce how much water is evaporated, but it moves the draw off the drinking-water supply, which is usually the local concern.

The other routes are design choices covered earlier: closed-loop liquid cooling with dry coolers, airside economisers in cool climates, and siting in places with low water stress. The IEA calls siting in areas of low water stress a straightforward way to meet sustainability goals. Corporate replenishment programmes sit alongside these. Google reports replenishing approximately 7.7 billion gallons of water in 2025, roughly 78% of its total freshwater consumption that year, through watershed projects. Replenishment is a real benefit to the watersheds involved, but it is accounting, not cooling: the water evaporated at a given campus is still evaporated there.

The pipeline, at today’s water intensities

The atlas tracks 581 data center projects under construction or announced. 307 of the 407 projects under construction publish a capacity, and together they add up to 51 GW. Announced projects that publish a figure add 32 GW. What would that mean for water if it were all cooled like today’s average US data center?

The table is arithmetic, and it deliberately overstates. It assumes every published megawatt runs at full load for a whole year, which no campus does, and it applies a US average to projects in every climate. Its purpose is to show scale. Run that way, the published capacity under construction alone would use about 443 TWh a year and, at 0.36 L/kWh, consume about 159.6 billion litres of water on site, more than double LBNL’s estimate of direct consumption by every US data center in 2023. Indirect water through electricity would be much larger still.

Two conclusions follow, and neither depends on the exact numbers. The first is that the cooling decisions being made now, in the projects listed on the atlas’s projects tracker, will set much of the industry’s water use for years to come, because cooling plant is expensive to change once a building is running. The second is that the electricity source matters as much as the cooling plant: a campus that pairs dry cooling with a thermal grid can end up with a larger total water footprint than one that evaporates water on site but runs on wind and solar.

How to read a water claim

When a press release, a planning document or a news story gives a water figure for a data center, six questions turn a number into information:

1. Consumption or withdrawal? Consumption is water lost, mostly to evaporation. Withdrawal includes water returned to its source. They can differ by several times.
2. Site, source or both? Site water is the data center’s own. Source water is consumed at the power stations that supply it. In the US average, indirect water is more than ten times direct water.
3. Per what? A WUE in L/kWh, a daily volume and an annual volume answer different questions. A daily figure in a heatwave can be many times the annual average per day.
4. At what load? A campus’s water volume scales with how much of its capacity is actually in use. Early phases of a large campus use a fraction of the eventual total.
5. What kind of water? Potable, reclaimed, groundwater and surface water have very different local consequences. A large draw of reclaimed water is a different proposition from the same volume of drinking water in a desert.
6. What is not published? Many operators do not disclose WUE or site volumes, and the IEA itself flags limited disclosure as a source of uncertainty. A missing number is not zero.

Water is the place where the data center debate is most local. A national average says little about a single campus, and a single campus says little about the industry. The numbers in this guide come from published estimates and from the atlas’s own records, and they will be updated as operators and regulators publish more. To follow where new capacity is being built, start with the projects tracker, the interactive map and the country pages; for the basics of what sits inside these buildings, see what is a data center. The documents behind every figure here are listed on the Sources page.