How Much Water Does a Data Center Use? Numbers & Benchmarks

Quick Answer

U.S. data centers directly consumed about 66 billion liters of water — roughly 17 billion gallons — in 2023, according to Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report, and LBNL projects hyperscale facilities alone will consume 16 to 33 billion gallons per year by 2028. At the facility level, Google reported its average data center used about 450,000 gallons per day in 2021, and the largest campuses can approach 5 million gallons per day. Most of that water is lost to evaporative cooling: at typical cooling tower settings, 75 to 83% of makeup water evaporates and never returns to the sewer. Count the water consumed generating a data center’s electricity and the footprint grows roughly 12-fold, to nearly 800 billion liters in 2023 per LBNL.

The Headline Numbers: US Data Center Water Usage

Data center water use is climbing fast because data center energy use is climbing fast. According to Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report, U.S. data centers consumed 176 terawatt-hours of electricity in 2023 — 4.4% of all U.S. electricity — and could reach 325 to 580 TWh by 2028, or 6.7% to 12% of the national total. Every one of those kilowatt-hours produces heat, and rejecting that heat is where the water goes.

The same LBNL report puts direct water consumption by U.S. data centers at 66 billion liters (about 17 billion gallons) in 2023, up from 21.2 billion liters in 2014 — a tripling in nine years. Hyperscale and colocation facilities account for 84% of that total. Looking ahead, LBNL projects hyperscale data centers alone will consume between 60 and 124 billion liters (16 to 33 billion gallons) annually by 2028.

“Direct” water means water piped into the facility itself, mostly for cooling. It does not include the much larger volume of water consumed at power plants to generate the electricity data centers buy — more on that below.

Company disclosures tell the same story from the inside. Google’s 2025 Environmental Report shows the company consumed approximately 8.1 billion gallons (31 billion liters) across its data centers and offices in 2024, a 28% increase over the prior year as its business grew.

How Much Water Does One Data Center Use?

National totals are abstract. Per-facility numbers are what site-selection analysts, utilities, and facility managers actually work with.

David Mytton’s peer-reviewed 2021 analysis in npj Clean Water found that a small 1 MW data center using traditional evaporative cooling can consume around 25.5 million liters (about 6.7 million gallons) of water per year — roughly 18,000 gallons per day per megawatt. A medium-sized 15 MW data center, Mytton calculated, uses as much water as three average-sized hospitals or more than two 18-hole golf courses.

Hyperscale operators run bigger. In a 2022 company blog post on climate-conscious data center cooling, Google disclosed that its average data center consumed approximately 450,000 gallons of water per day in 2021, with the global fleet consuming about 4.3 billion gallons that year. And a 2025 analysis by the Environmental and Energy Study Institute notes the largest facilities can consume up to 5 million gallons per day — comparable to the daily water use of a town of 10,000 to 50,000 people.

These figures vary enormously with climate and cooling design. A facility using air-cooled chillers or free-air cooling may use almost no water on site, while an identical IT load cooled with open cooling towers in a hot, dry climate sits at the top of the range. Our companion guide to data center cooling water compares the major cooling technologies and their water profiles in detail.

Hyperscale vs. Colocation vs. Enterprise

Facility type matters as much as size. LBNL’s 2024 report shows the market has inverted over a decade: in 2014, internal enterprise data centers accounted for 64% of direct water consumption, but by 2023 hyperscale and colocation facilities consumed 84% of the total, with enterprise server rooms falling to just 12% — a share LBNL projects will shrink to 2% by 2028.

Hyperscale campuses use the most water in absolute terms simply because they concentrate hundreds of megawatts on one site. Per unit of computing, though, they are usually far more efficient than the enterprise server rooms they replaced, thanks to optimized cooling plants and aggressive use of economizers. Colocation facilities sit in between — and colo tenants often never see the water bill their workloads generate, which is one reason water usage gets less scrutiny than power.

Where the Water Goes Inside a Data Center

In a typical water-cooled data center, the overwhelming majority of water goes to heat rejection — usually a cooling tower serving water-cooled chillers or a direct evaporative system. Smaller amounts go to humidification for the data halls and to ordinary domestic uses like restrooms, which together are usually a single-digit percentage of the total.

The U.S. EPA’s WaterSense at Work guidance on cooling towers (2023) identifies four ways water leaves a cooling tower system: evaporation, blowdown (also called bleed-off), drift, and leaks or overflows. Evaporation is the point of the machine — it is the mechanism that actually rejects heat. Blowdown is water deliberately drained to keep dissolved minerals from concentrating as pure water evaporates away. Drift — fine droplets carried out with the fan exhaust — is small, typically 0.05% to 0.2% of circulating water per EPA.

How the makeup water splits between evaporation and blowdown depends on cycles of concentration: the ratio of mineral concentration in the tower water to that in the incoming makeup water, which per EPA also approximately equals the ratio of makeup volume to blowdown volume. The math is simple and worth doing:

• Blowdown = Evaporation ÷ (Cycles − 1)
• Evaporated share of makeup ≈ (Cycles − 1) ÷ Cycles
• At 4 cycles: 75% of makeup water evaporates. At 5 cycles: 80%. At 6 cycles: 83%.

So at the 4 to 6 cycles common in well-run towers, roughly 75 to 83% of every gallon entering the cooling system leaves as vapor, not as wastewater. That single fact drives both the consumption statistics above and the billing problem covered later in this article. For the underlying mechanics, see our explainers on cooling tower blowdown and cycles of concentration.

One more distinction worth having straight: consumed is not the same as withdrawn. Withdrawal is water taken from a source; consumption is the portion that evaporates or is otherwise permanently removed from the local water cycle, as LBNL’s 2024 report defines it. Blowdown is withdrawn but not consumed — it goes back to the treatment plant. Evaporation is both. The distinction matters for permits, for drought planning, and, as we will see, for your sewer bill.

The quality of the water matters too. Cooling towers do not require drinking-quality water, yet Mytton’s 2021 analysis found some data center operators draw more than half of their water from potable sources — one large operator reported 57% to 65% potable in 2017 through 2019. That is why reclaimed or recycled water supplies are an increasingly common feature of new data center developments.

Direct vs. Indirect: The Water Hidden in Electricity

On-site consumption is only part of a data center’s water footprint. Thermoelectric power plants — coal, natural gas, nuclear — evaporate large volumes of cooling water for every megawatt-hour they generate, and that water is attributable to the electricity’s end user.

The reference dataset here is Macknick et al.’s 2012 study for the National Renewable Energy Laboratory, published in Environmental Research Letters, which compiled operational water consumption factors by generation technology:

Grid-average estimates vary by methodology, year, and regional fuel mix, so responsible analyses present a range. Mytton’s 2021 npj Clean Water paper used a 2015 U.S. average water intensity of 2.18 liters per kWh of electricity generated. LBNL’s 2024 report, using plant-level EIA data mapped to each balancing authority, calculated that data center electricity consumption in 2023 carried a national average of 4.52 liters of indirect water consumption per kWh — which is how 176 TWh of electricity translates into an indirect water footprint of nearly 800 billion liters (about 211 billion gallons), roughly 12 times the direct figure.

The indirect footprint also spreads the impact geographically. A 2021 study by Siddik, Shehabi, and Marston in Environmental Research Letters found that the U.S. data center industry directly or indirectly draws water from 90% of U.S. watersheds. The same study estimated the total 2018 operational water footprint of U.S. data centers at 513 million cubic meters — about 135 billion gallons — of which roughly three-quarters was indirect water at power plants.

The practical takeaway: a data center’s true water footprint depends on its electricity source as much as its cooling design. A water-frugal, air-cooled facility on a coal-heavy grid can consume more total water per kWh than an evaporatively cooled facility running on wind and solar, which consume almost no operational water per Macknick’s factors.

Water Usage Effectiveness (WUE): The Benchmark That Matters

The industry’s standard water metric is Water Usage Effectiveness, or WUE, created by The Green Grid in 2011. Site WUE divides a facility’s annual water consumption in liters by its IT equipment energy use in kilowatt-hours. A facility that consumes 100,000 liters while its IT gear uses 50,000 kWh has a WUE of 2.0 L/kWh. Source WUE additionally counts the water embedded in electricity generation, as LBNL’s 2024 report notes, but it is harder to calculate and rarely disclosed.

Benchmarks, from published sources:

A January 2025 guide from Data Center Knowledge pegs the average WUE across data centers at 1.8 L/kWh, while EESI’s 2025 analysis puts it at 1.9 L/kWh — call the industry average 1.8 to 1.9 L/kWh depending on the survey. The same Data Center Knowledge guide cites Amazon’s reported 0.19 L/kWh at the efficient end. Microsoft’s own datacenter efficiency disclosures report a global fleet WUE of 0.30 L/kWh for fiscal year 2024, improving to 0.27 in fiscal year 2025.

Two cautions. First, a low site WUE is not automatically green: LBNL’s 2024 report stresses the tradeoff that air-cooled systems use little water but more energy, shifting consumption to the power plant. Second, most operators cannot quote their WUE at all — a 2016 Uptime Institute analysis found fewer than one-third of data center operators track water usage or use the WUE metric. You cannot benchmark what you do not meter. Our deeper dive on water usage effectiveness (WUE) covers the formula, its limits, and how to measure it properly.

A worked example makes the scale concrete. Take a 20 MW IT load running at full utilization with a WUE of 1.8 L/kWh, the low end of the industry-average range:

• 20,000 kW × 8,760 hours = 175.2 million kWh of IT energy per year
• 175.2 million kWh × 1.8 L/kWh = 315 million liters ≈ 83 million gallons per year
• That is roughly 228,000 gallons per day — consistent with Google’s reported 450,000 gallons per day for its (larger) average facility

Does AI Increase Data Center Water Use?

Yes — mostly by increasing electricity use, which raises both direct cooling load and indirect water at the power plant. LBNL’s 2024 report traces the inflection point clearly: electricity use by GPU-accelerated servers, the hardware behind AI training and inference, grew from less than 2 TWh in 2017 to more than 40 TWh in 2023, and total server energy more than tripled from about 30 TWh in 2014 to nearly 100 TWh in 2023. Those TWh curves are the reason LBNL’s 2028 water projections run so much higher than its historical numbers.

AI also changes cooling design. High-density GPU racks increasingly require direct liquid cooling, which LBNL notes can pair with waterside economizers and higher coolant temperatures to improve both energy and water efficiency — or with evaporative heat rejection that consumes more water. Per-query estimates make headlines — UC Riverside researchers estimate a 100-word AI prompt consumes roughly 519 milliliters of water, about one bottle, as cited in EESI’s 2025 analysis — but for operators the actionable number is facility-level consumption, measured at the meter.

Regional Stress: Why Location Changes the Story

Averages hide the geography. The 2021 Siddik, Shehabi, and Marston study found that one-fifth of data centers’ direct water footprint comes from moderately to highly water-stressed watersheds, and nearly half of data center servers are fully or partially powered by power plants located in water-stressed regions.

Concentration compounds the issue. EESI’s 2025 analysis reports that data centers in Loudoun County, Virginia — one of the largest data center markets in the world — used around 900 million gallons of water in 2023. Operators are responding: Google’s 2025 Environmental Report states that 72% of its 2024 freshwater withdrawals came from sources at low risk of water depletion or scarcity, and Microsoft has driven its fleet WUE down by design. For any individual facility, though, the local watershed and the local utility’s rate structure matter more than any national average.

The Cost Angle: Sewer Charges on Water That Evaporated

Here is the part of the water story that rarely makes the sustainability reports: how the water is billed.

Most commercial water utilities calculate the sewer portion of the bill from the incoming water meter, on the assumption that what comes in goes back down the drain. For an office building, that assumption is roughly fair. For a data center running evaporative cooling, it fails badly — as shown above, 75 to 83% of cooling tower makeup water leaves as vapor and never reaches the sewer. Sewer treatment is also expensive: the sewer portion is often half or more of a combined water/sewer bill.

Run the numbers on the 20 MW example facility:

• Annual site water from the WUE math above: ~83 million gallons. WUE’s numerator is strictly water consumed, but to keep this estimate conservative, treat the 83 million gallons as total water purchased at the meter — nearly all of it cooling tower makeup
• Evaporated share of that makeup at 5 cycles of concentration: 80%, or about 66 million gallons that never enter the sewer
• Assumed sewer rate: $5 per 1,000 gallons (commonly published municipal sewer rates run roughly $3 to $12 per 1,000 gallons, but rates vary widely — check your utility’s rate schedule)
• Sewer charges on evaporated water: 66,000 × $5 = about $330,000 per year

That $330,000 is not a rounding error; it is a recurring annual overpayment for treatment service the utility never performed. Many utilities have a remedy on the books — usually called an evaporation credit, sewer credit, or cooling tower deduction — that adjusts sewer billing to reflect actual return flow, provided the facility can document the difference with submetering. We explain the mechanics, qualification requirements, and realistic savings in our guide to evaporation credits for data centers.

This is not a loophole. Sewer rates exist to recover the cost of treating wastewater, and evaporated water generates no treatment cost. Utilities grant these adjustments because accurate billing is the stated basis of their own tariffs — they simply require the customer to prove the evaporation with meter data rather than estimates. The scale of the correction grows with the facility: at hyperscale volumes of 1 million gallons per day and up, the same 80% evaporation math reaches 7 figures per year.

What Data Center Operators Can Do

Whether your goal is ESG reporting, drought resilience, or simply a smaller utility bill, the playbook starts with measurement and ends with documentation.

• Submeter the cooling loop. Separate meters on cooling tower makeup and blowdown turn “we think we evaporate a lot” into auditable gallons — the foundation for WUE reporting and sewer credits alike. Our practical guide to data center water submetering covers meter selection and placement.
• Optimize cycles of concentration. EPA WaterSense identifies blowdown reduction as the most significant water conservation opportunity in cooling tower operation. Raising cycles cuts both makeup and blowdown, within water chemistry limits.
• Claim sewer/evaporation credits. If your utility bills sewer on the incoming meter, documented evaporation is money on the table every month.
• Review the rate schedule. Some utilities offer industrial classifications, interruptible rates, or reclaimed water at lower cost. Our overview of data center water and sewer costs ranks the levers by effort and return.
• Monitor continuously. Real-time flow data catches stuck fill valves and leaks that a monthly utility bill hides for weeks.

None of this requires new cooling infrastructure. It requires meters, math, and paperwork filed with the right utility contact — a process RPM has repeated with nearly 200 water utilities across 36 states.

The Bottom Line

U.S. data centers consumed about 17 billion gallons of water directly in 2023, and LBNL’s projections point sharply upward through 2028 as AI workloads grow. The operators who come out ahead will be the ones who meter their water, know their WUE, and make sure they only pay for the sewer service they actually use. The gallons are getting bigger every year — and so is the value of measuring them.