How much water does a data center use? The real answer spans 1,000x

A mid-size enterprise data center consumes roughly 300,000 to 500,000 gallons of water a day, a large hyperscale campus runs 1 to 5 million gallons a day (about what a town of 10,000 to 50,000 people uses), and a facility built on closed-loop or air cooling can operate on a few thousand gallons a day, essentially bathrooms and landscaping. All of these figures describe real facilities operating today, and the spread between them comes down to one design choice, the cooling system, plus one location fact, the grid and watershed the facility sits on. In turn, what this means is that a single average figure will misstate the impact of most individual facilities by a wide margin, so what is best is to look at a few key numbers, in addition to the reasoning behind each.

The national numbers and the per-MW math

US data centers directly consumed about 66 billion liters (17.4 billion gallons) of cooling water in 2023, according to Lawrence Berkeley National Laboratory's 2024 report to Congress, and current projections put direct consumption at 38 to 73 billion gallons by 2028, with hyperscale facilities expected to account for roughly half of it.

The standard efficiency metric, water usage effectiveness (WUE), measures liters of water consumed per kilowatt-hour of IT energy, and the industry average sits around 1.9 L/kWh. One MW of IT load draws about 24,000 kWh a day. What this comes out to: roughly 12,000 gallons per day per MW at the average WUE, which is over a million gallons a day for a fully evaporative 100 MW facility, with the more efficient hyperscale designs bringing that down to 400,000 to 550,000. The per-MW figure also moves with climate and season, and summer demand can triple at a given site, in the same weeks when municipal systems are under the most strain.

Google is one of the few operators publishing site-level water data, and sixteen of its twenty-four US and Canadian data centers withdrew under 1 million gallons a day in 2024 while three withdrew more than 2 million. The sites run on different cooling designs, which produces the gap. Its evaporatively cooled Council Bluffs, Iowa campus alone withdrew about 1.75 billion gallons over the year, the most of any Google site.

Where the water goes

Most public arguments about data centers tend to confuse two things:

1. Withdrawal, which is the amount taken from the pipe.

2. Consumption, which is the amount that evaporates and doesn't return.

Evaporative cooling remains the dominant design because it saves electricity, and it works by circulating water through the servers and cooling towers, where a large share of it evaporates. Of the 7.7 billion gallons Google's data centers consumed in 2024, most of it evaporated in the cooling towers, with the rest exiting as blowdown, wastewater concentrated with dissolved minerals that needs treatment before reuse. Evaporated water leaves that watershed until it falls somewhere as precipitation, and the local supply gets no credit for rain that falls in a different basin. Closed-loop systems recirculate the cooling water and cut consumption by up to 70%, while air-cooled designs consume close to zero on site and pay for it in higher electricity use, typically around 10% more.

The off-site water footprint through electricity

Lawrence Berkeley National Laboratory's same report estimated that the electricity US data centers used in 2023 consumed nearly 800 billion liters of water at power plants, around twelve times the onsite total (4.5 L/kWh indirect against 0.36 L/kWh direct). Thermoelectric generation from gas, coal, and nuclear plants evaporates cooling water through the same process the data center does, while wind and solar consume close to none. One caveat on the twelve-times ratio: LBNL's indirect figure also counts reservoir evaporation attributed to hydropower, which raises it in the West, so it works as an upper-bound reference rather than a precise multiplier.

As a result of this, WUE needs to be read together with PUE and the power source before it says anything about total water impact. When a county debates a facility's water permit, the onsite request is what appears in the docket... the electricity procurement behind it is frequently the larger draw.

Total water by cooling design and grid

For a 100 MW facility, onsite cooling water plus the water evaporated generating its electricity produces these totals in million gallons a day:

Cooling	Renewables	Gas CC	Coal	Nuclear
Evaporative	1.20	1.80	2.58	3.14
Closed-loop	0.35	0.97	1.78	2.36
Air-cooled	0.03	0.70	1.58	2.21

An air-cooled facility on coal (1.58) or nuclear (2.21) draws more water than an evaporative one on renewables (1.20), and air cooling on a gas grid (0.70) stays below it. On any single grid the drier cooling design uses less, and dry cooling stops beating a clean-grid evaporative build once the power source passes about 1.37 L/kWh, which coal (1.81) and nuclear (2.54) exceed and combined-cycle gas (0.78) does not.

Grid factors are Macknick et al. (2012) median consumption values for recirculating cooling: gas combined cycle 205 gal/MWh, coal 479, nuclear 672, wind and solar near zero.

Gasilov Group's water strategy practice works on this withdrawal/consumption and watershed-exposure analysis for operators, developers, and the utilities and municipalities that serve them.

Learn more about our Water Strategy services

What replenishment can fix

Google replenished about 7.7 billion gallons in 2025, roughly 78% of its freshwater consumption, up from 64% in 2024, and has committed to replenishing 120% of consumption by 2030, alongside air cooling in at-risk watersheds and reclaimed water where available. The replenished volume nearly doubled year over year while the percentage moved 14 points. In turn, what this means is that consumption is growing underneath the projects faster than the projects can scale, and the 2030 target depends on that growth slowing.

Replenishment credits accrue company-wide while the water leaves one specific watershed. In turn, what this means is that the pledges naturally fix the aggregate number without helping the county the facility sits in. In Newton County, Georgia, a single Meta facility uses 500,000 gallons a day, about 10% of the county's entire water consumption, and pending proposals would more than double what the whole county uses today. Northern Virginia's data centers drew close to 2 billion gallons in 2023, up 63% from 2019. Roughly two-thirds of data centers built since 2022 sit in water-stressed regions, so the distance between where credits accrue and where the strain occurs is widening.

As I told Al Jazeera during the June heat dome, a facility running evaporative cooling in a heatwave is drawing on the same supply residents are often already under restrictions for. I would take that a step further to say that peak cooling water demand and peak grid stress arrive in the same afternoon hours that municipal supplies are most constrained, and most siting studies still model annual averages.

What changes

Anyone evaluating a project, whether as a buyer, a regulator, or a host community, should ask for four numbers: consumption and withdrawal reported separately, WUE alongside PUE and the power source, the share of potable versus reclaimed supply, and the July draw, since annual figures average the peak away. Seven in ten Americans now oppose data center construction in their own communities, with resource use the leading stated reason, and a Georgia campus that drew 29 million gallons through unmetered connections while the county asked residents to stop watering lawns shows how one episode can turn a routine rezoning into a moratorium (Fayetteville has since banned new data centers across its zoning districts). A closed-loop or reclaimed-water design costs more upfront and increasingly comes out cheaper once permitting delay gets priced in, because contested hearings can run months in the counties where water is scarcest.

Fewer than a third of operators track their own water consumption, most publish nothing at the site level, and some local governments have signed NDAs covering the figures. Google's disclosure shift shows site-level reporting is commercially survivable, and its 120% replenishment target comes due in 2030, which gives the industry one measurable check on whether the numbers converge with the pledges.

If you're not sure where your exposure sits, the Water Exposure Diagnostic, a two-week screen, is the entry point. If you already know what you need, get in touch and we can scope it.

Contact Us

Arif Gasilov is partner for Natural Resources & Built Environment at Gasilov Group, where he works on energy regulatory analysis, water governance, utility rate cases, and the resource footprint of large development. Meet Our Partners →

FAQ

How much water does a single AI prompt use? Google's published methodology puts a median Gemini text prompt at about 0.26 mL per prompt, roughly 14,559 prompts per gallon, counting onsite cooling only. Independent estimates run higher once the electricity's water is included, and either way the per-prompt figure matters less than where the aggregate draw is concentrated.

Do data centers return the water they withdraw? In evaporative systems, roughly 70 to 80% evaporates and leaves the local water system; the rest returns as mineral-concentrated blowdown requiring treatment. Closed-loop and air-cooled designs return or never withdraw most of their water, which is why the cooling choice is the single biggest determinant of local impact.