How Much Water Does a Data Center Actually Use? The Numbers, the Cooling Methods, and Where It Goes

There is no single answer, because a data center is not a standard-sized thing and cooling designs differ by climate and by choice. What exists now, and did not exist five years ago, is real disclosure. Google stopped claiming site-level water use as a trade secret nationwide after settling a public records suit it and the City of The Dalles lost the fight over, which we cover in our case page on that records fight. Every major operator now publishes annual water figures.

The national picture. Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report, prepared for the Department of Energy under the Energy Act of 2020, estimated that U.S. data centers consumed 21.2 billion liters of water directly in 2014 and about 66 billion liters in 2023, roughly 17 billion gallons. By 2023, hyperscale and colocation facilities accounted for 84 percent of that total. LBNL projected hyperscale data centers alone would consume between 60 and 124 billion liters in 2028.

The operators, in their own reports. These are the most recent figures each company published, with the year each covers.

Operator	Figure	Year	Source
Google, all owned and operated data centers	13,562 million gallons withdrawn, 3,039 million gallons discharged, 10,523 million gallons consumed	2025	Google 2026 Environmental Report
Google, company-wide including offices	14,689 million gallons withdrawn, 10,869 million gallons consumed, up 34 percent from 2024	2025	Google 2026 Environmental Report
Meta, data centers	4,145 megaliters withdrawn, about 1.1 billion gallons; 2,974 megaliters consumed	2024	Meta 2025 Environmental Data Index
Meta, water usage effectiveness	0.19 liters per kilowatt hour, down from 0.30 in 2020	2024	Meta 2025 Environmental Data Index
Microsoft, water usage effectiveness	0.27 liters per kilowatt hour globally, a 25 percent reduction from its 2022 baseline	FY25	Microsoft 2026 Environmental Sustainability Report
Amazon Web Services, water usage effectiveness	0.12 liters withdrawn per kilowatt hour of IT load, down from 0.15 in 2024 and 0.25 in 2021	2025	Amazon sustainability disclosures
U.S. average site WUE, all data centers	Just over 0.36 liters per kilowatt hour, projected to rise to 0.45 to 0.48 after 2023	2023	LBNL 2024 report

Site by site is where it gets local. Google publishes withdrawal, discharge, and consumption for each campus. The 2025 numbers, in millions of gallons consumed, include Council Bluffs, Iowa at 1,346.0; Mayes County, Oklahoma at 1,081.4; Berkeley County, South Carolina at 901.7; New Albany, Ohio at 835.7; Papillion, Nebraska at 548.2; The Dalles, Oregon at 469.0; Montgomery County, Tennessee at 424.7; Douglas County, Georgia at 367.1; Lockbourne, Ohio at 359.5; Lenoir, North Carolina at 344.3; Henderson, Nevada at 239.4; Midlothian, Texas at 220.2; Leesburg, Virginia at 190.7; Ashburn, Virginia at 53.2; and Mesa, Arizona at 9.7. The report footnotes eight of its listed sites as air cooled facilities, and five of them, Dublin, Montreal, Pflugerville in Texas, Phoenix, and San Bernardo in Chile, each consumed 0.1 million gallons for the entire year.

Servers turn essentially all the electricity they draw into heat, and that heat has to leave the building continuously or the equipment fails. There are only two ways to get rid of it: move it into the air, or move it into water and let the water carry it away by evaporating.

Evaporation is the cheap way, in energy terms. Turning liquid water into vapor absorbs a large amount of heat, and it happens without a compressor doing the work. A cooling tower sprays warm water over a surface, air moves across it, some of the water evaporates and carries the heat off with it, and the cooled remainder circulates back into the building. Microsoft describes the same mechanism in its 2026 report: servers generate heat that must be managed continuously, and doing so requires energy and, in many conventional cooling systems, water through evaporation.

Do the same job without evaporation and you have to pay for it with electricity. An air cooled chiller uses refrigerant and compressors to reject heat into the outside air, and compressors run on power. LBNL puts the tradeoff about as plainly as it can be put: water cooled chillers and other evaporation-based cooling systems are generally more energy efficient than air cooled chillers or other waterless systems, and while air cooled chillers use no water, they use more energy. Google's 2026 report makes the same claim from the other direction, describing water as a highly efficient means of cooling that requires less energy than air cooling technologies.

This is the tradeoff most coverage misses. A campus can cut its water use to nearly nothing by going air cooled, and it will then draw more electricity for the same amount of computing. A campus can cut its electricity overhead by evaporating water, and its water numbers go up. Neither choice is free, and a community that wins one fight has often, without being told, chosen the other cost. If a project's electricity draw is your concern, our guide on data centers and your electric bill covers that side.

Kansas has a concrete example of the choice being made in writing. In the development agreements Barber County commissioners approved on August 31, 2026, Google's roughly 1.7 million square foot Project Helium is air cooled and closed loop, with no new water wells, the existing well limited to domestic use, and all other water coming from and all wastewater going to a regulated municipal utility. Residents there raised many objections, but operational water withdrawal was contracted away. We cover the agreement, and what it does and does not leave residents, in our Barber County report.

Four designs cover most of what is being built. When you read a project's site plan or air permit, this is the vocabulary to look for.

Evaporative cooling, including cooling towers. Water is evaporated deliberately to carry heat away. Highest water use, lowest cooling energy. LBNL found that data centers using water cooled chillers without economizers show the highest WUE of any design, largely because of cooling tower water use. This is the design behind the big site numbers above.

Adiabatic or hybrid cooling. The system runs dry most of the year and sprays water only when outside air gets too warm to do the job alone. LBNL notes that this approach conserves water compared with evaporative cooling towers and is prevalent among hyperscale operators, whose optimized practices let them run on outside air most of the time with only sporadic adiabatic assist. Amazon describes exactly this pattern, saying it uses free air cooling about 90 percent of the time and turns to evaporative cooling during peak heat. LBNL also flags that this design is the hardest to model, and that some hyperscale operators report real WUE values of 0.1 to 0.3 L/kWh for similar systems, higher than the simulation suggested.

Closed loop chilled water. Water circulates in a sealed loop between the building and chillers and is not evaporated away. Once filled, it is topped up rather than continuously consumed. Water use is small, electricity use is higher, because the chillers do the work the evaporation would have done. This is what Barber County's agreement specifies.

Air cooling. Heat is rejected to outside air with no evaporative step. In a cool climate this is very efficient; in a hot one it costs real power. Google's site table is the clearest demonstration available: its facilities footnoted as air cooled in Dublin, Montreal, Pflugerville, Phoenix, and San Bernardo each consumed 0.1 million gallons in all of 2025, and the three others in that group, Sydney, Wilmer in Texas, and Storey County in Nevada, consumed 0.2, 0.3, and 6.7 million gallons. Council Bluffs consumed 1,346.0 million gallons.

Liquid cooling, or direct to chip. The design being built for dense AI racks, which produce more heat per square foot than air can practically remove. Coolant runs to the chip itself. Microsoft says the system design it introduced in 2024 circulates liquid in a closed loop between servers and chillers, avoids evaporative cooling altogether, and is expected to avoid the need for more than 125 million liters of water per year per data center. LBNL is more cautious about the category as a whole, projecting that the U.S. average site WUE rises after 2023, to between 0.45 and 0.48 L/kWh, partly because of the increased water consumption of liquid cooled systems. Both can be right: a closed loop at the chip can eliminate evaporation at that step while the facility's overall heat rejection still runs through a tower.

Water usage effectiveness, or WUE, is liters of water per kilowatt hour. It is an efficiency ratio, not a volume. Microsoft defines it as the ratio of the yearly amount of water used for cooling and humidification to the energy used, and notes that the lower the number, the better. Meta reported an annual data center WUE of 0.19 L/kWh for 2024, improved from 0.30 in 2020. Microsoft reported a global average of 0.27 L/kWh for FY25. Amazon reported 0.12 L/kWh for 2025. LBNL's estimate for the U.S. fleet as a whole was just over 0.36 L/kWh through 2023.

A low WUE can hide a very large total draw. That is the point most press releases are designed to obscure, and it is not a matter of opinion, it is arithmetic. WUE multiplied by the campus's electricity use gives you the water. A campus that doubles in size at a constant WUE doubles its water. Google's Council Bluffs figure and Google's fleet-wide efficiency claims are not in tension; a very efficient very large thing is still a very large thing. If you are evaluating a project, the number that matters to your utility is gallons per day, not liters per kilowatt hour.

Consumption and withdrawal are not the same number, and this is where public arguments break down. Withdrawal is how much water the facility takes out of the supply. Discharge is how much it puts back. Consumption is the difference, the part that evaporated and is gone from the local system. Google publishes all three and states its method directly: it calculates water consumption by subtracting water discharge from water withdrawal. For 2025, that was 14,689 million gallons withdrawn, 3,820 million gallons discharged, and 10,869 million gallons consumed company-wide.

Both numbers matter, for different reasons. Withdrawal is what competes with your town for supply and capacity on a hot afternoon. Consumption is what never comes back to the watershed. An operator quoting consumption and a resident quoting withdrawal can both be honest and still be describing numbers that differ by a factor of three or more. Microsoft acknowledged the gap when it changed its own replenishment target in FY25 to track total water withdrawals rather than consumption, saying the more rigorous approach aligns replenishment with the full volume of water withdrawn across its data centers, not just the portion evaporated for cooling.

The metrics can also be defined inconsistently between the sides of a comparison. Amazon's disclosure defines its WUE as liters of water withdrawn per kilowatt hour of IT load, and compares its 0.12 L/kWh to an industry average of 0.84 L/kWh that its own footnote attributes to consumption data from de Vries-Gao (2025) combined with a Department of Energy cooling tower methodology. A withdrawal-based numerator compared against a consumption-derived benchmark is worth knowing about before anyone quotes the ratio in a hearing.

Water leaving a data center goes to one of two places, and the split is the difference between withdrawal and consumption.

Into the atmosphere. The evaporated portion becomes water vapor, visible as the plume above a cooling tower on a cold day. It rejoins the water cycle, but not necessarily in the same watershed, and not on any schedule useful to the town it left. This is the consumption number.

Into a sewer, as blowdown. Cooling tower water circulates repeatedly, and each pass evaporates pure water while leaving behind everything dissolved in it. Minerals, salts, and treatment chemicals added to control scale and biological growth concentrate in the remaining water until it has to be dumped and replaced. That periodic purge is called blowdown, and it is the discharge line in the reports. Google reported discharging 3,039 million gallons from its data centers in 2025, about 22 percent of what it withdrew.

Why blowdown is the bridge to our contamination coverage. Blowdown is not the same water that went in. It is concentrated, and where it goes matters. Discharged to a municipal treatment plant, it becomes that plant's loading problem. Land applied, or routed to a shared industrial disposal system, it enters the ground. That is the mechanism at the center of the largest data center water payout on record: in a federal class action alleging that concentrated wastewater from Amazon's eastern Oregon data centers, routed to Port of Morrow disposal systems and applied to land, added to nitrate contamination in the Lower Umatilla Basin, Amazon Data Services agreed in March 2026 to a $20.5 million settlement. Amazon denies wrongdoing, and the money is earmarked for well and water infrastructure projects rather than individual checks. Our case page has the full picture, and our guide to data center water contamination claims covers the legal theories for both construction and operational water harm.

Not all of the intake is fresh water, either. Operators increasingly buy reclaimed or non-potable water where a utility will sell it. Google reported that 394.3 of the 440.6 million gallons withdrawn at its Douglas County, Georgia campus in 2025 was reclaimed wastewater, and that across all its data centers 918 million gallons were reclaimed and 1,695 million gallons non-potable, out of 13,562 million gallons total. Whether a facility is on potable municipal supply or reclaimed water is one of the most useful questions a resident can ask, because it changes who it competes with.

This is the question behind almost every search that leads here, and it deserves a careful answer rather than a reassuring or an alarming one. There are three separate mechanisms, they operate on different timelines, and they are not equally documented.

Municipal supply competition. If the facility buys from your utility, it is a customer on the same system you are, and its draw counts against the same treatment and pumping capacity, especially during a drought or a peak summer week. This is a public, checkable fact, not a theory: consumption records held by a public utility are generally reachable under state public records law, and after the settlement in The Dalles, Google stopped asserting site-level water use as a trade secret nationwide.

Aquifer drawdown. If the facility pumps its own wells from the same aquifer you do, the physics is the same as any large withdrawal. The U.S. Geological Survey describes it directly: withdrawing water faster than it is replenished lowers the water table and creates a cone of depression around the well, an effect that USGS says can be short-lived or last decades and can range from a small drop to many hundreds of feet, and that pumping your own well too hard can cause a neighbor's well on the same aquifer to run dry. Whether that is happening at a particular site is a hydrogeology question that requires monitoring data and experts, not inference from a timeline.

Construction phase impacts, which are separate from operational water use. Clearing, blasting, grading, and dewatering can disturb the shallow groundwater residential wells draw from, and this happens before the facility has cooled a single server. The volumes are real and usually invisible in the headline figures: Meta's 2025 Environmental Data Index notes that an additional 1,019 megaliters of water, roughly 269 million gallons, was withdrawn for construction of Meta data centers and is not included in its 2024 withdrawal numbers.

What residents have alleged, and what has been established. Be precise here, because the distinction matters legally. In Newton County, Georgia, four families sued Meta Platforms and the local development authority in late August 2026, alleging that clearing, blasting, and excavation for the Stanton Springs campus disturbed the aquifer and filled their wells with silt and sediment. Meta says a groundwater study it commissioned found that groundwater flows away from the plaintiffs' properties and that the facility was likely not the cause, and says it will defend against what it calls baseless lawsuits. In Louisa County, Virginia, a homeowner sued Amazon in late July 2026 over brown, contaminated well water, dust, and noise from data center construction, pleading nuisance, negligence, and property damage. Both cases are pending. No court or agency has found that either operator caused either household's well problems, and none of this establishes what is happening at any other site. Our tracker pages for Newton County and Louisa County carry the current status of each.

What you can check, starting today.

• Is the facility on municipal water or its own wells? Ask the utility and the county. The answer determines which of the three mechanisms above is even possible, and it is usually in the site plan or the development agreement.
• Get the water service agreement. If a utility is supplying the facility, there is a contract, and it is a public record. Read it for committed volumes, peak-day limits, rate structure, and any curtailment clause saying who gets cut first in a shortage.
• Get the permitted withdrawal limits. If the facility pumps its own groundwater, most states require a withdrawal permit or registration stating a maximum. Ask your state water agency for the permit and any required reporting.
• Pull the state well records. Most state geological surveys or environmental agencies maintain well logs and water level data. Your own well's drilling record, depth, and any monitoring wells in the area are the baseline you will wish you had.
• Read the permit conditions, not the press release. A special or conditional use permit can carry enforceable water conditions. Barber County put cooling design and water sourcing into a contract. Projections in a press release bind no one.
• Test your well now. A dated certified laboratory test taken before or early in construction is the single most valuable document in any future claim, and the Georgia plaintiffs' hardest problem is that almost nobody has one. Our evidence guide has the full checklist, and our guide to fighting a proposed project covers the approval process where water conditions get written.

Iowa is where several of these threads meet, with Google's Council Bluffs campus, Meta in Altoona, and Microsoft in the Des Moines metro drawing on the same regional systems; our Iowa page covers that state's water story, including the West Des Moines dispute, in detail. Our water issue center tracks the national picture.

If your water has already changed and a data center was built or is being built nearby, the useful order of operations is: test through a certified lab, report to your county health department and state environmental agency, request the facility's water records, and then get a legal read on your specific facts. A free case review connects you with an independent attorney in our network who handles data center cases in your state, at no cost and with no obligation. Whether your timeline and test results support a claim is their judgment, and it is a judgment that gets easier the earlier the record starts.