The data center water figure everyone quotes measures 2014, and most of it is power plants

Four hundred and forty-nine million gallons a day.

That is the number. It appears in trade-group explainers, in advocacy briefs, in local newspaper coverage of proposed data centers, and, on the morning this piece was written, inside Google's own summary of what data centers do to the water supply. It is usually given as current, and it is usually given without a scope.

It describes 2014. About 85 percent of it is water that never went anywhere near a data center.

How much water do United States data centers actually use?

The current federal estimate is 66 billion liters consumed on site in 2023, which is 17.4 billion gallons, or 47.8 million gallons a day across every data center in the country. A second and much larger quantity, nearly 800 billion liters, is consumed at power plants generating the electricity those buildings use.

Both figures come from the 2024 United States Data Center Energy Usage Report, prepared by Lawrence Berkeley National Laboratory for the Department of Energy and published in December 2024. It is the third such report in nearly twenty years. Its electricity findings were covered widely: 176 TWh in 2023, 4.4 percent of national consumption, with a 2028 scenario range of roughly 325 to 580 TWh. Its water findings were covered almost not at all, and the reason is structural. The executive summary runs three pages and does not contain the word water in any quantity. The water section begins on printed page 55.

What that section says is worth reading slowly. In 2014 data centers consumed 21.2 billion liters directly, 64 percent of it in internal corporate facilities. By 2023 the total had reached 66 billion liters, with hyperscale and colocation accounting for 84 percent and internal data centers falling to 12 percent. Direct water use roughly tripled over the period. So did electricity, from about 60 TWh to 176 TWh. The industry did not become dramatically thirstier per unit of computing. It became much larger.

Where does the 449 million gallons a day figure come from?

It is a unit conversion of one sentence in the previous edition of the same report. The 2016 report estimated that approximately 626 billion liters of water was consumed in 2014 for data centers. That is 1.7 billion liters a day, or 449 million gallons. Every restatement in circulation traces back through a 2021 journal paper to that one line.

The paper is David Mytton's "Data centre water consumption," published in npj Clean Water in 2021, and it is careful about what it is doing. Its abstract puts United States data center water consumption at 1.7 billion liters a day. Its body sentence is explicit: in 2014, a total of 626 billion liters of water use was attributable to US data centers, footnoted to the 2016 Berkeley report.

What happens next is a chain of small, individually reasonable steps. The Environmental and Energy Study Institute, whose explainer currently ranks second on this search, cites the paper and writes that one report estimated US data centers consume 449 million gallons of water per day and 163.7 billion gallons annually, "as of 2021." The parenthetical is the publication year of the paper. The data year is 2014. From there the figure is quoted onward without either date, and the scope drops away entirely.

The scope is the part that matters most. The 2016 report's Figure 25 is titled "Direct vs. Indirect U.S. Data Center Water Consumption," so the 626 billion liters is both categories added together. The report states its own factors: 7.6 liters per kWh for the water consumed generating electricity, and 1.8 liters per kWh of site energy for water consumed at the building. Multiply its own 70 billion kWh estimate for 2014 by its own 7.6 liters per kWh and you get 532 billion liters of power plant water, leaving roughly 94 billion liters on site. Around 85 percent of the most-quoted data center water figure in America is the water footprint of the electricity grid.

That reproduction is arithmetic on the report's stated assumptions, not a number the report prints. But it is the only way to read the total honestly, and it explains why the newer report's on-site figure for the same year, 21.2 billion liters, is not a contradiction so much as a different question answered with better tools.

What is the difference between water withdrawal and water consumption?

Withdrawal is water taken out of a river, an aquifer or a municipal main. Consumption is the portion that does not come back, lost to evaporation or otherwise removed from the immediate water cycle. The gap between them is enormous and it varies by user, which is why the two words are not interchangeable and why almost every comparison in circulation mixes them.

The United States Geological Survey publishes the definitive numbers, and they make the point better than any argument. In 2015 thermoelectric power plants withdrew 133 billion gallons a day, 41 percent of all water withdrawn in the country for any purpose. They consumed 4.31 billion gallons a day. The Survey states the ratio itself: consumption was 3 percent of thermoelectric withdrawals. Nearly everything a power plant takes goes back to the river a few degrees warmer.

Irrigation is the opposite. It withdrew 118 billion gallons a day and consumed 73.2 billion, which the Survey puts at 62 percent of total use. Two sectors, similar withdrawals, and a twentyfold difference in what is actually removed.

Mytton's paper compares data center consumption of 1.7 billion liters a day to total United States water consumption of 1,218 billion liters a day. That second number is USGS's 322 billion gallons a day of total withdrawals, converted. So are its three components: 503 billion liters for thermoelectric power is the 133 billion gallon withdrawal figure, 446 is irrigation's 118, and 147 is public supply's 39.0. The comparison sets a consumption figure against a withdrawal figure, in a paper whose subject is the difficulty of measuring data center water consumption. If the distinction can slip there, it can slip anywhere, and downstream it slips constantly.

How much of a data center's water footprint is the building, and how much is the grid?

For 2023, indirect water was roughly twelve times direct water: nearly 800 billion liters consumed at power plants against 66 billion liters consumed on site. A data center water figure is therefore mostly an electricity figure in different units, and it moves when the generating mix moves, whatever the cooling towers do.

This is why the two Berkeley reports differ so much and why neither is lying. The 2016 edition used a single national factor of 7.6 liters per kWh, drawn from a 2003 National Renewable Energy Laboratory paper on consumptive water use in power production. The 2024 edition builds the factor from Energy Information Administration plant-level generation and water data, routed through the balancing authority that actually serves each county, and arrives at 4.52 liters per kWh for the grid mix data centers draw on, against 4.35 liters per kWh for United States electricity as a whole.

Between those two reports the assumed water intensity of American electricity fell by 41 percent. Some of that is real. Coal plants with once-through and evaporative cooling have closed, and gas and wind and solar consume far less water per unit of output. Some of it is simply better arithmetic replacing a two-decade-old average. The published numbers cannot separate the two effects, and the report does not claim to.

The practical consequence is counterintuitive. A data center that switches to air cooling and eliminates its on-site water use has removed the smaller number. If it does so by running the chillers harder, and the report notes exactly this tradeoff, it has increased the larger one. The report distinguishes site water usage effectiveness from source water usage effectiveness for this reason, observes that including both reflects the true water cost, and then says plainly that the calculation is complex and highly dependent on the source of electricity.

What do the companies that actually meter it report?

Google is the only operator publishing withdrawal, discharge and consumption together, by site, under limited third-party assurance. In 2024 it withdrew 11,011 million gallons globally, discharged 2,876 million and consumed 8,135 million. Its data centers alone withdrew 9,866 million gallons and consumed 7,787 million, which is 79 percent of what they took.

Set that against the national model and something uncomfortable appears. Google's global data center consumption of 7,787 million gallons is 29.5 billion liters, or about 45 percent of Berkeley's 66 billion liter estimate for every data center in the United States in 2023. The comparison is not exact. Google's figure is global and covers 2024, the Berkeley figure is domestic and covers 2023, and Google is one of the largest operators in the world. But a single company's disclosure landing at nearly half of a national bottom-up model is the kind of result that should be reported rather than smoothed over, and no page on this subject reports it.

The company's own methodology note is the clearest definition of consumption published anywhere in this literature: water consumption is calculated by subtracting water discharge from water withdrawal. Withdrawal is metered or invoiced where the data exists and estimated by engineering principles where it does not, which is a limit stated openly and rarely stated at all elsewhere.

Two further figures from that disclosure are worth carrying. Of the 9,866 million gallons its data centers withdrew, 7,699 million was potable, 78 percent. And its replenishment projects returned 4,590 million gallons in 2024 against freshwater consumption of 7,210 million, 64 percent, up from 18 percent a year earlier. Consumption itself rose 28 percent over the same year.

Which data centers consume the most water?

Not the ones in the desert. Google's largest 2024 consumer was Council Bluffs, Iowa, at 1,010.2 million gallons, followed by Mayes County, Oklahoma at 833.2 million and Berkeley County, South Carolina at 776.5 million. Its Phoenix, Arizona site consumed 0.03 million gallons, a factor of roughly 34,000 below Council Bluffs.

The pattern is a choice, not an accident, and the company describes the mechanism. Where its water risk framework flags high depletion or scarcity, it builds air-cooled. It names three such sites: Waltham Cross in the United Kingdom, Mesa in Arizona and Canelones in Uruguay. The result is that the facilities generating the most local anxiety are frequently the ones consuming almost nothing on site, while the large evaporative loads sit in Iowa, Oklahoma, Tennessee and the Carolinas, where water is cheap and abundant and the story is not being written.

The pattern is not uniform, and the disclosure shows that too. Henderson, Nevada withdrew 359.9 million gallons and consumed 207.4 million in the same year Phoenix consumed almost none. Two sites, one state, opposite answers. Anyone arguing from climate alone to water use is arguing from the wrong variable.

Why is there no measured national number?

Because nobody measures it. The Berkeley figures are outputs of a bottom-up model that multiplies an installed base of equipment by assumed power draws and then by modeled cooling efficiencies, and the report says its analysis is limited by the lack of direct data. No federal agency collects data center water use. The reporting that does exist is voluntary, corporate and partial.

The denominator is in worse shape than the numerator. USGS has estimated national water use every five years since 1950, and the last full compilation in that series is Circular 1441, covering 2015 and published in 2018. There is no 2020 Circular. The Survey has moved to model-based reanalyses delivered through a data companion and a data repository, thermoelectric power from 2009 to 2020, irrigation and public supply from 2000 to 2020, with the remaining five categories released separately. That is more granular in some ways and it is not the same thing. When a story compares data center water use to national water use, the national side of that comparison is eleven years old.

So the honest description of what is known is narrow. On-site consumption by American data centers in 2023 was modeled at 66 billion liters. Water consumed generating their electricity was modeled at nearly 800 billion. Both figures rest on assumptions the authors publish and defend, neither is a measurement, and the most recent year either covers is three years behind the present. Everything past that point, including every 2025 and 2026 total now circulating, is an extrapolation of an extrapolation.

The 2024 report does give one forward number worth holding onto. Under its scenarios, hyperscale data centers alone are expected to consume between 60 and 124 billion liters directly in 2028. The lower bound of that range is roughly what the entire industry consumed on site in 2023.

For the electricity side of the same measurement problem, and the same laboratory's account of how little of this is observed rather than estimated, see our earlier piece on how much electricity data centers use.

Questions

Is 449 million gallons a day wrong? It is not wrong, it is mislabeled. It is a correct conversion of the 2016 Berkeley report's estimate for 2014, covering water consumed both at data centers and at the power plants supplying them. Quoted as a current on-site figure it overstates by roughly a factor of nine.

What is the best single number to use now? For water consumed at the buildings, 66 billion liters in 2023, or 47.8 million gallons a day nationally. For the full footprint including electricity generation, about 866 billion liters, roughly 627 million gallons a day. Both come from the December 2024 Berkeley report and both describe 2023.

Does a data center consume all the water it takes in? No. Google's data centers consumed 79 percent of what they withdrew in 2024. Power plants are the opposite case: USGS puts thermoelectric consumption at 3 percent of thermoelectric withdrawals. Any comparison that mixes the two measures is not comparing like with like.

Do data centers in dry states use the most water? Not in the only fleet-wide disclosure available. Google's Phoenix site consumed 0.03 million gallons in 2024 while its Council Bluffs, Iowa site consumed 1,010.2 million, because the company builds air-cooled where its water risk framework flags scarcity. Cooling design, not climate, decides the number.