The Water and Electricity Costs of Keeping AI Data Centers Cool

On 3 November, voters in ten states decided 41 local ballot measures on data centres. In 40 of them the result backed limits on development, from outright bans and moratoriums to rejected rezonings. Ohio held 21 of the measures, and at least 14 of its municipalities approved bans or restrictions. Grove City now requires a public vote on any development using more than 50 acres, 20 megawatts of power or 500,000 gallons of water a day. The one result in the industry's favour came in Harwood, North Dakota, where voters defeated a review requirement for large industrial projects.

 

Residents' objections range from energy bills, pollution and noise to lost farmland and water. For investors, that makes resource use a permitting risk: local campaigns had already blocked $18 billion of US data centre projects and delayed another $46 billion between May 2024 and March 2025.

 

At an Equinix data centre in San Jose, California, eight racks of Nvidia Grace Blackwell servers each hold 72 GPUs. Under heavy workloads the chips run at about 80°C, and liquid flowing through copper cold plates carries that heat away. Once filled, the cooling loop around the chips needs little new water. The building around it still uses water and electricity to get the heat outside, and the way it does that decides most of its footprint.

 

Two stages of cooling

 

Cooling an AI data centre happens in two places. Inside the server, coolant runs through cold plates attached to the processors and carries their heat to a distribution unit. A heat exchanger passes that heat to a separate facility circuit. The coolant around the chips stays sealed in its own loop, so routine operation does not need a steady supply of replacement water. A given volume of water can absorb roughly 3,000 times more heat than the same volume of air, which is why liquid cooling lets operators pack far more computing into the same floor space.

 

The second stage, moving the heat out of the building, is where water and electricity use diverge. Operators choose between three broad approaches.

 

Heat-rejection method	How it works	Water use	Electricity use	Main drawback
Dry cooling	Fans push outside air across heat exchangers; chillers may be needed in hot weather	Lowest	Highest in hot weather	Adds load at summer peaks, when the grid is already strained
Evaporative cooling	Water passes through cooling towers or wet pads, and the part that evaporates carries heat away	Highest	Lowest	Evaporated water must be replaced, which matters most in dry regions
Hybrid	Dry cooling in cold months, evaporative cooling or misting when it is hot	Varies by season	Varies by season	Results depend heavily on local climate

 

Climate decides much of the outcome. Dry cooling works well in winter, but on hot days the smaller gap between coolant and outdoor temperature means more fan power and, in many designs, compressor-based chillers similar to household air conditioning. Sites in Texas and Arizona, where daytime highs approach 100°F, face a choice between heavy water use and heavy electricity use for much of the year.

 

Saving water can cost electricity

 

The operators that run the largest fleets describe the trade-off in their own numbers. Google says water cooling can cut a data centre's energy use by about 10% compared with air cooling in many locations. About two-thirds of its data centres use evaporative cooling, and the company argues that is the better choice where water supplies are healthy. Where a watershed is at high risk, it says it now chooses air cooling or recycled water instead.

 

Microsoft has gone the other way for its newest AI sites. Its closed-loop, direct-to-chip design avoids evaporating water for cooling, and the company has said the switch from evaporative to mechanical cooling brings a "nominal increase" in power use and a higher power usage effectiveness (PUE). Chips that tolerate warmer coolant help limit that penalty, because the chillers have less work to do.

 

The water saved on site can also reappear at the power station. Thermal power plants consume water for their own cooling, so a data centre that cuts its water use by drawing more electricity may shift part of its footprint to the grid. Estimates of how large that indirect share is vary widely. Lawrence Berkeley National Laboratory (LBNL) put US data centres' direct water consumption at about 17.4 billion gallons in 2023 and the water consumed generating their electricity at roughly 211 billion gallons, about 12 times as much. Rystad Energy puts the US indirect share at more than twice direct use. The answer depends on the local generation mix and on when the extra power is drawn, since wind and solar use little water and gas or coal plants use a great deal.

 

The demand behind the bills

 

Cooling is one part of a much larger rise in electricity demand. LBNL's latest update, published in June 2026, puts US data centre consumption at 192 terawatt-hours in 2024, or 4.7% of the country's electricity. Its reference case reaches 649 TWh by 2030, about 11.8% of US consumption, within a range of 521 to 843 TWh. Globally, the International Energy Agency expects data centre consumption to roughly double from 485 TWh in 2025 to about 950 TWh in 2030, around 3% of world electricity demand.

 

Households in parts of the US are already paying for that growth. In PJM, the grid operator serving 13 eastern states and Washington, DC, capacity prices rose from $28.92 per megawatt-day for 2024/25 to $269.92 for 2025/26 and $329.17 for 2026/27. PJM's independent market monitor attributed 63% of the first jump to data centres, about $9.3 billion recovered from all customers, and expects another $6.5 billion in data-centre-driven charges next year. In Ohio, Pennsylvania and Michigan, capacity charges have added as much as $18 a month to household bills, and power prices became an issue in the midterm campaigns.

 

Those increases come from data centre demand as a whole, mostly the computing load itself. There is no reliable estimate of how much of it comes from operators choosing water-saving cooling. What the figures do show is that where grid capacity is tight, any extra electricity a cooling design needs lands on a system that is already expensive to expand.

 

Water is a local issue

 

National water totals look modest. The 17.4 billion gallons US data centres consumed directly in 2023 is roughly triple their 2014 level, according to a Congressional Research Service report published in July, but still a small share of national use. The same report found that municipal systems supply an estimated 97% of data centres' on-site water, which puts the pressure on local utilities rather than on any national supply.

 

Virginia, home to the world's largest data centre market, shows how uneven that pressure is. A December 2024 study by the state's Joint Legislative Audit and Review Commission estimated the industry used about 2.1 billion gallons in 2023. It found that around 83% of data centres used no more water than an average large office building, that the industry accounted for less than 0.5% of state withdrawals, and that current use was sustainable. The commission still recommended tighter rules so that localities with limited water do not commit too much of it to data centres.

 

Location matters as much as design. About two-thirds of new data centres built or in development since 2022 sit in areas of high water stress, based on World Resources Institute and DC Byte data. More than 160 new AI data centres opened in high-stress US areas in the three years to early 2025, up 70% on the previous three years. Water is often the last factor weighed in siting because it is cheap compared with land and power.

 

Two measures are easily confused. Withdrawal is water taken from a source, some of which is returned after use. Consumption is water that is not returned, mostly because it has evaporated. Evaporative cooling drives consumption, and reclaimed wastewater can reduce demand for drinking water, though it needs its own treatment and pipes.

 

What Microsoft and Meta promise

 

At Microsoft's Build conference in June, chief executive Satya Nadella said the company's newest AI data centres fill their cooling loop once and then use no water for cooling during operation. Microsoft said in January that it had deployed the design at sites in Wisconsin and Georgia, but it has announced no comprehensive programme to retrofit its existing fleet of more than 500 data centres. The company says about 90% of its 2025 owned fleet already uses low- or zero-water cooling, and it reported a water usage effectiveness (WUE) of 0.27 litres per kilowatt-hour for fiscal 2025.

 

Meta's $10 billion, 1-gigawatt campus in Lebanon, Indiana, announced in February, will use a closed-loop liquid cooling system that Meta says will consume no water for most of the year. The company has committed more than $120 million to water infrastructure and other public works, says it will pay the full cost of the site's water and wastewater service, and plans to restore all the water it consumes to local watersheds, partly through an irrigation project expected to restore 200 million gallons a year for ten years.

 

How much water the campus will actually consume is not yet known. Meta's planning request to the city was for a phased supply of up to 3 million gallons a day between 2027 and 2031, a projection rather than a measured figure. The utility serving the area is building supply of up to 25 million gallons a day by 2031, but that covers the whole industrial district, including a large Eli Lilly plant.

 

Restoration pledges also work differently from the demand they offset. Saving irrigation water elsewhere in the state does not put water into Lebanon's pipes on the hottest day of the year. For the local utility, the questions are how much supply is needed, when, and whether its infrastructure can deliver it.

 

Disclosure is catching up

 

Site-level water data rarely reaches the public, and that is the main obstacle for investors trying to judge these claims. Equinix, which runs the San Jose site, does not publish consumption for individual facilities. Its company-wide figures show operations consumed about 1.2 billion gallons in 2025, up 10% on 2024, while its WUE improved from 0.95 to 0.91 litres per kilowatt-hour. The facilities became more water-efficient per unit of computing, but total use still rose because there was more computing. Equinix gives customers third-party-verified, site-level water reports, so the data exists but is not public.

 

Company WUE figures cannot simply be ranked. Amazon and Meta measure water withdrawn rather than consumed, and Amazon's own regional values range from 0.02 litres per kilowatt-hour in Stockholm to 2.85 in Jakarta. Most figures also exclude water consumed at power plants.

 

Reporting frameworks are starting to close the gap. The SASB standard for software and IT services asks for water withdrawn and consumed, and the share of each in regions of high or extremely high water stress. The TNFD's technology and communications guidance, finalised in September, asks for WUE by location rather than a fleet average. It also asks cloud and AI companies to report water at the data centres they lease, which would capture tenants such as those renting racks from Equinix.

 

Europe has gone furthest. Under the recast Energy Efficiency Directive, data centres of 500 kW or more must report PUE, WUE and related indicators each year, though only about 770 of roughly 2,000 eligible facilities submitted data in the first cycle. In September, the European Commission adopted a delegated regulation introducing A-to-G ratings for data centre energy and water efficiency. The measure is still undergoing scrutiny by the European Parliament and Council, with the first electronic labels planned for August 2027. In the US, site-level disclosure remains largely voluntary.

 

That leaves American communities to ask for the numbers themselves, and voters in Mandan, North Dakota, and Clarksville, Arkansas, face data centre measures in December. Before any approval, the figures that matter are site-specific: a campus's expected water withdrawal and consumption, its demand on the hottest days of summer, the electricity each cooling option would need, and what local utilities will spend on the pipes, treatment and grid connections it requires.

 

Sources: CNBC; Lawrence Berkeley National Laboratory; International Energy Agency; US Congressional Research Service; Virginia Joint Legislative Audit and Review Commission; IEEFA; Google; Microsoft; Meta; Equinix; TNFD; European Commission.