AI Data Center Water: Why WUE Cannot Measure a Workload

Last updated: August 4, 2026

Water Usage Effectiveness, or WUE, divides a stated water quantity by the electricity a data center’s IT equipment consumed, reported in liters per kilowatt-hour. The water quantity is not defined the same way in every disclosure, and the facilities counted differ between publishers. No published WUE value isolates the water used by one AI training run or one inference request.

Verdict in one line: use a published WUE figure to interrogate a reporting boundary, and do not use it to rank providers or estimate a prompt until the boundaries have been matched.

WUE measures a facility, not a workload

The 2024 United States Data Center Energy Usage Report, prepared by Lawrence Berkeley National Laboratory for the Department of Energy, defines Water Usage Effectiveness as total data center water consumption divided by the electricity demand of the IT equipment, expressed in liters per kilowatt-hour.

Both terms in that ratio are facility quantities aggregated over a reporting year. The numerator covers whatever water the publisher counts as entering or leaving the site. The denominator covers all IT electricity in the building. Training and inference are not separated. AI workloads are not separated from storage, networking, database or conventional cloud workloads running on the same floor. In a colocation facility, tenants are not separated from each other.

Two arithmetic properties follow from the formula and matter in procurement. A falling WUE does not establish falling water use, because the ratio improves whenever IT energy grows faster than water. An annual figure also carries no seasonal shape, and seasonal shape is what a utility manages during a drought.

Berkeley Lab reaches the same limitation from the other direction. Its conclusions section states that comparing and benchmarking data center efficiency is hampered by the lack of sufficient metrics, and calls for energy and water performance metrics specific to different workloads or service segments so that more accurate comparisons become possible.

A per-prompt water figure therefore requires an allocation step from an annual facility ratio down to one request. That step connects a request to hardware energy, facility overhead, cooling conditions, water system operation, and, if indirect water is counted, the generation mix at that time and place. None of the disclosures reviewed in this briefing performs it.

Sources: Lawrence Berkeley National Laboratory, December 2024, pp. 39, 68 · Lawrence Berkeley National Laboratory, publication record

Direct water and electricity-related water

Berkeley Lab splits WUE in two. Site WUE covers water used on the premises, mostly for cooling. Source WUE covers water consumed generating the electricity the facility draws. The report keeps the two separate throughout and labels its water totals accordingly.

For 2023 the report estimates that US data centers consumed 66 billion liters of water directly. For the same country and year it estimates an indirect water footprint associated with the electricity those data centers used at close to 800 billion liters. Per unit, it gives a national average of 4.52 liters per kilowatt-hour indirect, against a modeled average site WUE that stays just above 0.36 liters per kilowatt-hour through 2023.

Three limits belong with those numbers, all stated in the report. The indirect estimate is modeled from regional grid mixes rather than metered at facilities. The method does not incorporate power purchase agreements between individual facilities and their electricity providers, or behind-the-meter generation, because facility-level data was unavailable. And the ratio between the two figures is a relationship between two national 2023 model outputs, not a constant that transfers to a given facility, provider, region or workload.

The report also states that a low site WUE is not automatically good. Evaporation-based cooling is generally more energy efficient than an air-cooled chiller or another waterless system, and a waterless system uses more electricity. A site metric can fall while electricity-related water rises.

Its forward scenario points the same way. The modeled US average site WUE rises from just above 0.36 through 2023 to a range of 0.45 to 0.48 liters per kilowatt-hour by 2028, and the report names increased water consumption of liquid-cooled systems among the drivers of that modeled rise. Within the 2023 modeled categories, the AI-specialized category records a site WUE of 0.61 against 0.32 for the general hyperscale category, with PUE running the other way at 1.14 against 1.22. These are modeled category averages built from cooling-system distributions and weather data. They are not measurements of any named operator.

Sources: Lawrence Berkeley National Laboratory, December 2024, pp. 39, 45, 47, 48, 55, 57 · The Green Grid, White Paper 35, 2011

What the four publishers actually publish

The four documents most often cited together do not contain four values of the same kind. Three publish a ratio in liters per kilowatt-hour. One publishes absolute volumes in million gallons.

Boundaries as stated in each publisher’s own methodology page or data index. Values are omitted deliberately. This table compares what is measured, not how well anyone performed.

Five differences in that table each block a direct comparison on their own.

Withdrawal and consumption are different quantities. Withdrawal is what enters the boundary. Consumption is withdrawal minus discharge. A site that returns most of its water reports a large withdrawal and a small consumption.

Cooling and humidification water is a third description, and the Microsoft page does not resolve it into either of the first two.

An owned-only estate and an owned-plus-leased estate cover different buildings. Workloads served from leased capacity may sit inside one publisher’s boundary and outside another’s.

A calendar-year value and a fiscal-year value can carry the same year label while covering different months.

An absolute volume in million gallons has no IT-energy denominator. It answers how much, not how much per unit of computing, and it cannot be placed in a column of ratios.

One further reading rule applies to the Alphabet schedule. Seawater returned to the sea is excluded from it, and Google separately identifies at least one seawater-cooled site. A low row for such a site cannot be read as the total water that site uses for cooling, because part of that water falls outside the reported category.

None of this shows that any publisher reported incorrectly. Each boundary above is a boundary its publisher stated, and the Alphabet schedule carries third-party assurance. Differences between publishers can also reflect climate, cooling architecture, utilization and water source, which are real physical differences rather than accounting ones. The public documents do not contain enough detail to separate the two, and this briefing does not claim to have separated them.

Sources: Amazon Web Services, sustainability · Microsoft, datacenter efficiency · Meta, 2025 Environmental Data Index · Alphabet, FY2025 environmental indicators assurance

How a seven-times-better benchmark is built

A published comparison shows the normalization work that usually stays out of sight.

AWS reports a 2025 withdrawal WUE of ‹0.12› liters per kilowatt-hour and describes it as more than seven times better than an industry average of ‹0.84›. The footnote states that the ‹0.84› figure was derived from a published consumption WUE of ‹0.56› by applying a cycles-of-concentration assumption of ‹three›.

The comparison may well be defensible. It is not a comparison of two identically reported values. It required converting a consumption figure into a withdrawal figure using an operating assumption about how many times cooling water is recirculated before it is discharged.

The useful part for a buyer is that the conversion is disclosed. A vendor that shows its normalization step is giving the reader what is needed to check the claim. A vendor that presents a ratio next to an industry average without one has left the same step unstated rather than unnecessary.

Source: Amazon Web Services, sustainability, water usage effectiveness and accompanying footnote

Absolute site totals are not efficiency

Alphabet publishes named-site water volumes, which most operators do not. Its 2026 Environmental Report gives withdrawal, discharge and consumption for each data center location for 2025.

Global operational water consumption for 2025 is 10,869 million gallons, against withdrawal of 14,689 million gallons and discharge of 3,820 million gallons. Data centers account for 10,523 million gallons of the consumption figure and offices for 346 million gallons. The report states that consumption rose 34% from 2024 to 2025. Across the reported series, consumption rose from 4,562 million gallons in 2021 to 10,869 million gallons in 2025.

The site rows span more than four orders of magnitude within one operator’s own estate. Council Bluffs, Iowa records 1,346.0 million gallons of consumption. Mayes County, Oklahoma records 1,081.4. Berkeley County, South Carolina records 901.7. The Dalles, Oregon records 469.0. At the other end, Dublin and Phoenix each record 0.1 and Hamina, Finland records 0.4.

The table does not attribute that spread. Facility load, facility scale, heat rejection technology and the source-water exclusion described in the previous section are all consistent with the rows as printed, and the table contains no IT-energy denominator with which to convert any row into an efficiency figure.

Three methodology points sit alongside the numbers, all stated in the report. Where metered discharge data is unavailable, the report applies a 90% discharge flow factor to domestic withdrawal and a 0% factor to irrigation withdrawal, and estimates the remainder using engineering principles. Because consumption is withdrawal minus discharge, the consumption figures are partly derived rather than metered. The report also gives two consumption denominators for 2025: total water consumption of 10,869 million gallons and freshwater consumption of 9,947 million gallons. Replenishment of 7,717 million gallons is published as 78% of the freshwater figure. Measured against the total water figure, the same volume is 71%, which is a calculation by Future Stack Reviews and not a company indicator.

Replenishment itself is an accounting result. The report describes it as volumetric water benefits assessed by a third party under the Volumetric Water Benefit Accounting 2.0 methodology, adjusted to Alphabet’s proportional financial contribution where a project has several funders. The 2030 figure of 19,722 million gallons is described as the capacity of the current portfolio once projects are fully implemented, which makes it a projection.

Sources: Google, 2026 Environmental Report, water sections and data appendix · Reporters Committee for Freedom of the Press, 15 December 2022 · City of The Dalles v. Advance Local Media, settlement and dismissal

Disclosure architecture by jurisdiction

Regulation is changing what facilities report. The route from a rule to a number a buyer can use runs through several separate layers, and the layers differ by jurisdiction.

Rows describe the instruments read for this briefing. Absence entries describe a documented search scope and do not assert that no such instrument exists elsewhere. Items in ‹guillemets› are reported by secondary retrieval and await direct verification.

Three points in that matrix are easy to compress incorrectly, so they are set out here in full.

The European Union operates two disclosure routes, not one. Under Article 12 of the recast Energy Efficiency Directive, Member States are to require owners and operators to make the Annex VII information about their data centers publicly available, and information protected by trade and business secret law is excluded from that duty. Separately, operators communicate the same information and key performance indicators to the European database, which is published in aggregated form at Member State and Union level while individual submissions are held confidential. A buyer seeking facility figures in the EU is asking under the first route, and national implementation of that route determines what arrives.

The delegated regulation also narrows the early filings. For the first reporting period an operator that cannot monitor certain listed indicators for technical reasons may omit them with an explanation. For the first two reporting periods a colocation operator that cannot calculate two specified indicators may instead estimate and state the share of computer room floor area its submission covers.

China’s 2024 action plan sets three power usage effectiveness figures and no water figure. For water it directs that conservation assessment for new and expanded projects be strengthened, and, under policy safeguards, that standards for data center water conservation assessment, water use quotas and water-saving enterprises be established. Those are targets and instructions in a policy document. Whether they carry the force of a binding legal instrument was not established here, and the position may have moved since: a national WUE method standard is reported to have entered into force in December 2025 as a recommended standard, and its full text has not been read for this briefing.

Sources: Council of the European Union, 15 March 2024, C(2024) 1639 final, recitals 3, 9 and 12 and Articles 1, 3 and 5 · Directive (EU) 2023/1791, Article 12 and Annex VII · National Development and Reform Commission and others, July 2024, action plan full text · Press Information Bureau, Government of India, 30 March 2026

The eight fields to request from a vendor

Every divergence documented above is a field a vendor can fill in one line. A response carrying all eight can be compared against another response carrying all eight. A single ratio cannot.

Fields derived from the divergences documented in sections 02 to 06 of this briefing.

Field seven is worth a sentence of its own because no reviewed disclosure answers it. Liquid cooling describes how heat moves from a chip into a loop. It does not describe how heat leaves the loop, and an evaporative tower and a dry cooler put very different quantities of water into the same reported number.

Sources: Lawrence Berkeley National Laboratory, December 2024, pp. 40 to 45 · The Green Grid, White Paper 35, 2011

What would make two figures comparable

This briefing argues that public water figures are not yet matched. That argument has a falsifier, and stating it is part of the argument.

Two published figures become comparable when the numerator, the direct or indirect boundary, the estate scope, the reporting period, the estimation policy, the water-source exclusions and the assurance basis are the same, or when a disclosed conversion brings one into the other’s terms. Section 04 shows what a disclosed conversion looks like. Nothing prevents an operator from publishing all eight fields alongside its ratio, and an operator that does removes this briefing’s objection for its own figure.

Three things this briefing does not establish belong here rather than in a footnote.

It does not establish that any operator performs better or worse than another. Boundary divergence blocks the comparison; it does not decide it.

It does not establish that the differences between published figures are mainly accounting rather than engineering. Climate, cooling architecture, water source and utilization produce real differences, and the public record does not contain enough facility detail to apportion the variance.

It does not establish that a buyer somewhere has changed a cloud region or provider on water grounds. Procurement interest in water metrics is documented in guidance. A completed procurement decision driven by a water figure was not found in the sources reviewed for this briefing.

Sources: Lawrence Berkeley National Laboratory, December 2024, p. 68 · Amazon Web Services, sustainability

FAQ

Methodology and evidence scope

Verdict