How data centers use water

Why servers need water at all

Racks of servers throw off heat, and most large data centers cool that heat with evaporative cooling towers — the same basic technology as a swamp cooler, spraying water into the air moving through the cooling system so its evaporation carries heat away. It's cheap and energy-efficient compared to running electricity-hungry air conditioning alone, which is exactly why it's so widely used, but it means the water doesn't come back: once it evaporates, it's gone from the local supply. Industry estimates put the loss at up to 85% of the water a data center uses for cooling. We cover the local-temperature and extreme-weather claims tied to that evaporated water on a separate page on evaporative cooling.

What's left behind when the water evaporates

The water going into a cooling tower is never pure H₂O — municipal and well water both carry dissolved minerals (calcium, magnesium, chloride, silica), and only the water itself evaporates. Everything dissolved in it stays behind and concentrates in what's left, a ratio operators call "cycles of concentration." The U.S. Department of Energy notes that most systems run at 2 to 4 cycles, and that pushing to 6 cuts both fresh-water intake and the volume of concentrated water needing disposal by roughly half — but concentration can't run forever, or the minerals start scaling equipment and corroding pipe.

So facilities periodically drain off ("blow down") a portion of that concentrated water and replace it with fresh makeup water. Blowdown carries the concentrated minerals, plus whatever operators added on purpose to keep the system running — biocides to stop algae and Legionella bacteria from growing in the warm water, and corrosion or scale inhibitors that have historically included metals like zinc and chromium. Supposedly, none of this is dumped untreated: it's either sent to a municipal sewer system under a local pretreatment program or discharged to surface water under a Clean Water Act permit that sets limits on dissolved solids, temperature, and specific metals. Ohio EPA, for instance, drafted exactly that kind of permit for a proposed data center in 2025, setting effluent and thermal discharge limits before construction could proceed — the kind of review a large industrial water user typically goes through, not a data center–specific carve-out. This has been seeing growing scrutiny by local residents near data centers since what the locals are seeing and what data centers have been reporting are not always in alignment.

How much, nationally

U.S. data centers directly consumed an estimated 17.4 billion gallons of water in 2023 — enough for roughly 160,000 American households for a year — up from about 5.6 billion gallons in 2014. That's still well under 0.2% of all U.S. water withdrawals nationally, but national totals hide local concentration: a large campus can be a meaningful share of water use in its specific county or river basin even while barely registering in a national figure.

The bigger, less visible number

Direct, on-site water use isn't the whole picture. Generating the electricity a data center runs on takes water too — mostly evaporative cooling at the power plant itself — and that "indirect" water use is actually larger: an estimated 211 billion gallons in 2023, over 80% of a data center's combined water footprint. It happens somewhere else on the grid, which is part of why it's easy to undercount a data center's real water use by looking only at what happens on its own property.

Efficiency varies enormously by design

The industry measures this with Water Usage Effectiveness (WUE) — liters of water per kilowatt-hour of computing, the water equivalent of the more familiar Power Usage Effectiveness (PUE). The typical facility runs around 1.8 liters per kWh, but design choices swing this dramatically: facilities in cooler, drier climates or ones built around air-cooling instead of evaporative towers can run far lower. Amazon has reported a fleet-wide WUE of just 0.19 L/kWh — roughly a tenth of the industry average — by leaning on outside air and reclaimed water instead of fresh-water evaporative cooling wherever the climate allows it.

What each company has disclosed

Water disclosure isn't standardized the way carbon reporting is, so these numbers come from each company's own sustainability report, on its own schedule, using its own boundaries — they aren't a clean apples-to-apples comparison, they are the "best" public figures available according to the companies and not an independent agency. Several major operators, including some of the largest cloud providers, don't publish facility-level water figures at all.

Why AI specifically is driving the increase

Google's own 2026 Environmental Report shows why its water use jumped: 10.9 billion gallons consumed in 2025, up 34% from the year before and more than double its 2021 level, even though Google replenished 78% of that through watershed restoration projects. Google attributes the jump directly to AI infrastructure — specifically, that AI data center construction and the power it draws are growing faster than the company's efficiency gains can offset. That's the same dynamic behind the 38–73 billion gallon range LBNL projects for 2028 in the chart above: it isn't data centers in general driving the growth, it's the AI-accelerator generation of them specifically.

Researchers at UC Riverside have tried to put a number on AI's water use directly. Their widely cited 2023 study estimated that training GPT-3 in Microsoft's U.S. data centers consumed roughly 700,000 liters of on-site freshwater over about two weeks of training — comparable to the water used to manufacture around 370 BMWs — and that the same training run would have used roughly three times as much water if it had happened in Microsoft's less-efficient Asia-based facilities instead. The same researchers estimated that every 10 to 50 medium-length chatbot responses collectively "drink" about one 500 mL bottle of water for cooling. OpenAI CEO Sam Altman has separately said a single ChatGPT query uses only about 0.3 mL of water — a gap of roughly three orders of magnitude from the outside estimates, largely because the two figures don't count the same things: which specific data centers, cooling systems, and electricity-generation water use get included changes the answer enormously. We're not in a position to referee whose methodology is right; both numbers are on the record, from named people, and the gap between them is itself part of the story.

What isn't in dispute is that the cooling technology a facility uses changes its water footprint dramatically for the exact same computing load. Direct-to-chip liquid cooling can cut water use by up to 95% compared to a conventional evaporative-cooling design, and full immersion cooling — submerging servers in a dielectric fluid instead of using water-based heat rejection at all — eliminates evaporative water use entirely. That's also exactly the technology AI accelerator racks are pushing the industry toward for thermal reasons having nothing to do with water, as we cover in detail on our page on data center cooling systems, including what share of U.S. facilities actually use each method today.

Nonpartisan, plainly

We don't take a position on whether data centers should be built near a given water source, what cooling technology they should use, or how local utilities should price water for large industrial users — those are real disagreements with legitimate arguments on more than one side, and it's not this page's job to settle them. These are the sourced numbers, plainly.

Talking points

These are the questions we think you should ask those who are running for office and will represent you. We don't give our opinion on the answer, but we DO think you should be talking about them.

1. Should data centers that heavily rely on water cooling be allowed in drought prone areas?
2. Should data center water use be publicly reported the same way electricity use often already is?
3. Should a data center be required to prioritize reclaimed or non-potable water over municipal drinking water where both are available?
4. Should communities facing water stress have the right to reject a new data center's water permit?

Read more

• LBNL: 2024 United States Data Center Energy Usage Report↗eta.lbl.gov

  The direct/indirect water-use figures behind the national numbers above.

• DOE FEMP: Best Management Practice #10 — Cooling Tower Management↗energy.gov

  The federal government's own explanation of cycles of concentration and blowdown, cited above.

• Congressional Research Service: Data Centers and Water (R49057)↗congress.gov

  A nonpartisan congressional primer on data center water use and the policy questions it raises.

• Google Environmental Report 2026↗sustainability.google

  Google's own disclosed 2025 water consumption figure and its AI-growth explanation, cited above.

• UC Riverside: AI programs consume large volumes of scarce water↗news.ucr.edu

  The "Making AI Less Thirsty" study (Li, Yang, Islam & Ren) behind the GPT-3 training and per-query figures below.

• Microsoft 2025 Environmental Sustainability Report↗microsoft.com

  Microsoft's own disclosed global water consumption figures.

• Brockovich AI Data Center Reporting↗brockovichdatacenter.com

  An environmental advocate's crowdsourced map and reports on data centers' local water and environmental impact — community reporting, not government data.