Data center water use explained: direct, indirect, WUE and reuse

US data centers consumed about 66 billion liters of water on site in 2023, and an estimated 800 billion liters more indirectly, at power plants that generate their electricity, according to Lawrence Berkeley National Laboratory. The gap between those two numbers, and the choices behind them, explain most of the debate.

This article walks through what counts as data center water use, how the standard metric (WUE) works, why operators trade water against electricity, and what reuse and closed-loop designs actually change. Every figure is labeled with its source and date. It is an explainer, not a verdict on any operator, community or policy.

Key numbers at a glance

• 21.2 billion liters: direct (on-site) water consumed by US data centers in 2014 (Lawrence Berkeley National Laboratory, LBNL, report dated December 2024, modeled estimate).
• 66 billion liters: the same measure for 2023, about 17.4 billion US gallons (LBNL, modeled estimate; gallon conversion added for US readers).
• 84 percent and 12 percent: share of 2023 direct water used by hyperscale plus large colocation sites versus internal (enterprise) data centers; internal sites were 64 percent in 2014 (LBNL).
• 60 to 124 billion liters: LBNL's range for hyperscale data centers alone in 2028 (projection, LBNL).
• Nearly 800 billion liters: indirect water consumed in 2023 at power plants supplying US data centers (LBNL, modeled estimate).
• 176 TWh: US data center electricity use in 2023, 4.4 percent of US electricity (LBNL).
• 4.52 liters per kWh: LBNL's indirect water intensity for data center electricity in 2023, versus 4.35 for US electricity overall.
• 0.12 liters per kWh: Amazon Web Services' reported global WUE for 2025, down from 0.15 in 2024 (AWS sustainability page, company statement, accessed 19 October 2026).
• 0.30 liters per kWh: Microsoft's reported FY2024 average WUE, versus 0.49 for 2021 (Microsoft blog, December 2024, company statement).
• 2.2 cubic metres per MWh (2.2 liters per kWh): median WUE in 2021 of Singapore data centers that are large water users, with an aim of 2.0 or lower within 10 years (Infocomm Media Development Authority, Green Data Centre Roadmap, 2024).
• Up to 97 percent: share of a data center's water use that cooling towers account for, per the same IMDA roadmap (government publication).

Three words that get mixed up

Much of the confusion in this topic comes from three terms that sound alike.

Water withdrawal is the water taken from a river, lake, aquifer or utility. Some of it is returned afterward.

Water consumption is the part that is not returned, because it evaporates or is otherwise removed from the local water cycle. LBNL defines consumption as the amount of withdrawn water that is "permanently removed from the immediate water cycle due to evaporation or other irreversible processes."

Water use, in loose headlines, can mean either. When a number is reported, check which one it is. A cooling tower might withdraw a large volume and consume only part of it. A figure for withdrawal can be larger than one for consumption for the same facility.

Direct water use: what happens on site

Direct (or site) water use is almost entirely about cooling. Computers turn electricity into heat, and one very effective way to remove heat is to evaporate water, which carries a lot of energy away with it. In a typical design, warm water from the building circulates through a cooling tower. A fraction of it evaporates, cooling the rest, and the cooled water returns to the loop. That evaporated fraction is consumed water. A little extra water is also drained to keep minerals from building up, a process called blowdown.

LBNL estimates that US data centers consumed 21.2 billion liters directly in 2014 and 66 billion liters in 2023. That is roughly a tripling in nine years, in line with the growth in electricity use over the same period. These are modeled estimates, not meter readings. LBNL built them from simulated cooling system performance across climates, paired with market data on which cooling systems are sold, and it warns that real facilities rarely perform as designed.

A striking detail is who accounts for the water. In 2014, internal data centers, the rooms and small facilities run inside ordinary companies, were 64 percent of the direct total. By 2023, hyperscale and large colocation facilities accounted for 84 percent of the 66 billion liters while internal data centers fell to 12 percent, which LBNL attributes to water efficiency improvements. In plain terms, computing moved into big, efficient facilities, and those facilities use less water per unit of computing even though they use more in total.

For 2028, LBNL projects hyperscale data centers alone could consume 60 to 124 billion liters. The wide range reflects uncertainty about how much AI hardware ships, how it is used, and which cooling systems operators choose.

Indirect water use: the water behind the electricity

Power plants use water too. Thermoelectric plants, which boil water to drive turbines, need large amounts of cooling, and hydroelectric reservoirs lose water to open-surface evaporation. A data center that draws electricity from the grid is, indirectly, drawing on that water.

LBNL puts the indirect figure for US data centers in 2023 at nearly 800 billion liters, about twelve times the direct figure. That works out to a national average of 4.52 liters of water consumed per kilowatt-hour of data center electricity, close to the 4.35 liters per kWh average for US electricity overall. LBNL calculates this using the electricity mix of each local grid operator and states that it does not account for power purchase agreements or on-site generation, either of which could change the estimate substantially.

Why does this matter? Because decisions about cooling push water between the two columns. Choose a cooling system that uses no water on site, and you generally need more electricity, which may mean more indirect water at the power plant. LBNL puts it plainly: low site water use is not necessarily good, because evaporation-based cooling is generally more energy efficient than waterless alternatives, which "use more energy." It describes site WUE and source WUE as two views of one picture, noting that including both reflects the true water cost but is complex and depends on the electricity source.

This is also why single-number comparisons can mislead. A facility that boasts of zero site water but runs on a water-hungry grid has moved the footprint, not erased it. The reverse also holds: a facility using some site water but powered largely by wind and solar, which consume little water, may have a lower total footprint. Public data are not detailed enough to rank real facilities on total footprint, so this article does not try.

WUE: the standard measure, and what it hides

Water usage effectiveness (WUE) is the standard metric. It is the water consumed by a data center divided by the electricity used by its computing equipment, expressed in liters per kilowatt-hour. Singapore's regulator uses the same ratio in cubic metres per megawatt hour, which gives the same number: 1 cubic metre per MWh equals 1 liter per kWh. Its 2024 roadmap gives a sense of scale: a WUE below 0.2 means a site consumes less than a glass of water for every MWh of IT energy.

A WUE of zero means no water consumed on site. At the other end, evaporative cooling in a hot, humid climate pushes WUE up. Singapore's IMDA reported a 2021 median of 2.2 for the city's large water using data centers, and AWS reports 2.85 for its Jakarta region in 2025.

Several reported values help calibrate the scale:

• Amazon Web Services reported a 2025 global WUE of 0.12 liters per kWh, down from 0.15 in 2024. It reports big regional differences: 0.05 in Europe, 0.08 in North America, 1.10 in Asia Pacific excluding China, and city values from 0.06 for its US East region (Northern Virginia) to 2.85 for Jakarta in 2025. AWS states that its figure is "more than 7 times better than the industry average," which it takes as 0.84 liters per kWh, derived from a consumption WUE of 0.56 and cycles of concentration of 3. That is a company calculation using its own assumptions.
• Microsoft reported a FY2024 average WUE of 0.30 liters per kWh, a 39 percent improvement from 0.49 reported for 2021. Note that Microsoft's footnote bases this on a withdrawal measure, and that definitions across companies may differ.
• LBNL's modeled median for hyperscale data centers is 0.32. It adds a caution: some hyperscale facilities report 0.1 to 0.3 liters per kWh for similar systems, higher than its simulations assumed for one cooling type, and if the true median for that type is 0.2, the hyperscale median would rise from 0.32 to 0.40.

Notice what these numbers do not tell you. WUE says nothing about where the water comes from, whether the area is short of water, whether the water is drinking-quality or recycled, or when the water is used. A small WUE at a site in a drought-stricken basin can matter more locally than a high WUE next to abundant supply. WUE also measures water per unit of computing energy. A big AI campus can have a low WUE and still draw a large total volume simply because it is big.

WUE also trades against energy. Evaporative cooling tends to raise WUE and lower PUE, while dry or mechanical cooling lowers WUE but raises PUE, which is the trade-off LBNL describes. PUE, power usage effectiveness, is total facility electricity divided by IT electricity. See the guide to air versus liquid cooling for how that plays out in design.

How operators cut or avoid water use

The main techniques, drawn from LBNL's cooling system list and company statements, are these.

• Free cooling with outside air. When the weather is cool and dry, outside air can do the work with no evaporation. LBNL's simulations show airside economizers allowing chilled water systems to shut off in favorable weather, "resulting in substantial water conservation."
• Dry coolers. Radiator-like units that reject heat to air with fans and no evaporation. LBNL notes dry coolers can meet cooling needs for much of the year in favorable climates, and that those with adiabatic assist, a water spray used only on the hottest days, use the most water among the remaining waterless-leaning options, mostly during wet mode.
• Warmer liquid loops. Liquid-cooled servers can run with warmer coolant, which lets facilities rely on dry coolers more often.
• Closed-loop chip cooling. Microsoft says its design in use since August 2024 "consumes zero water for cooling," with water filled once during construction and circulating between servers and chillers. The company says the design is expected to avoid more than 125 million liters of water per year per datacenter, a figure based on its FY2024 withdrawal WUE of 0.30, and that water is still used for restrooms and kitchens. It warns that replacing evaporative cooling with mechanical cooling raises energy use slightly.
• Better controls. Microsoft reports a 23 percent year-over-year WUE reduction in FY2025 from cooling updates in its Phoenix datacenters, achieved with finer real-time control of temperature and humidity.
• Cooling tower management. Singapore's IMDA points to recycling blowdown water and raising cycles of concentration, for example by treating the water so it can circulate longer, since towers account for up to 97 percent of a site's water use.
• Siting. Putting a facility where outside air is cool for more of the year. See why location changes cooling, water and power.

Reuse, reclaimed water and replenishment

These three words are not the same thing either.

Reuse means recycling water within the facility, for example cycling cooling tower water more times before draining it. The AWS page refers to cycles of concentration, the number of times water is reused in a tower before it must be discharged, which is a direct lever on consumption.

Reclaimed water, also called recycled or non-potable water, is treated wastewater supplied by a utility for uses like cooling, which leaves drinking-quality water for people. In Virginia, the state's Joint Legislative Audit and Review Commission (JLARC), in the data center study it was directed to carry out in 2023, concluded that most data centers use about the same amount of water as an average large office building or less, while a few use substantially more.

Replenishment is a corporate pledge to return as much water to the environment as the company consumes, usually through projects elsewhere in the same basin or country. Examples from company pages:

• AWS says it is committed to being water positive by 2030, and expects "22B+ liters of water" to be returned each year through its replenishment efforts.
• Meta says it aims to become water positive in 2030, in its 2025 Sustainability Report.
• Microsoft says it replenished over 14.2 million cubic meters of water in FY2025, more than it withdrew, and that this figure is company-wide, not specific to datacenters. It also describes rainwater harvesting in Ontario and Quebec and in Amsterdam, with expected, not measured, collection volumes.

Replenishment does not reduce the local draw. A replenishment project can be in a different watershed from the data center. It is a separate question whether the water is returned in the same place and at the same time of year as the use. There is no widely adopted independent audit method that settles this, so these remain company claims.

Rules and disclosure

Rules are emerging. Singapore's IMDA says in its 2024 Green Data Centre Roadmap that it will work with the national water agency, PUB, to help new and existing data centers reach a WUE of 2.0 cubic metres per MWh or lower over the next 10 years. That is a stated aim, not a hard cap. More detail is in the post on Singapore's data centre capacity rules. Water is also a recurring theme in local permitting, covered in how long it takes to build an AI data center.

Disclosure is uneven. Some companies publish site-level WUE, some only fleet averages, and some not at all. LBNL notes that its own estimates exist partly because facility-level data are not available. That is a reason to treat any single national figure as an order of magnitude, not a precise count.

Putting the numbers in context

A few honest comparisons help, with caveats.

• Direct water use of 66 billion liters is about 17.4 billion gallons for the whole US sector in 2023, in LBNL's model. Virginia's JLARC found most data centers use about as much water as a large office building or less. Both point to a modest footprint for the typical site, but averages say little about a given town or basin, where one large campus can be a big new customer.
• The indirect figure is far larger, but it is spread across hundreds of power plants and is a footprint of the grid, not a draw on the community near the data center.
• Growth is the live question. LBNL's 2028 range for hyperscale direct water is 60 to 124 billion liters, and its range for total US data center electricity is 325 to 580 TWh against 176 TWh in 2023. The ranges are wide because they depend on how much AI hardware ships and which cooling systems are chosen.

What is firm and what is uncertain

Firm enough to rely on:

• The definitions of consumption, WUE and PUE, and the direction of the energy and water trade-off, which LBNL states explicitly.
• LBNL's 2014 and 2023 direct water estimates and its indirect estimate, as modeled figures from a national laboratory.
• Company-reported WUE values from AWS and Microsoft, as company statements with their own boundaries.

Uncertain:

• Actual water use at any specific site. National figures are model outputs, and LBNL says real performance is usually worse than design.
• Whether company WUE figures are comparable. Definitions (withdrawal or consumption), boundaries and reporting years differ.
• The 2028 outlook, which depends on AI hardware shipments and cooling choices.
• Whether Singapore's 2.0 WUE aim is met within its 10 year window, and whether other regulators adopt similar targets.
• Whether replenishment claims match local conditions.

What to watch next

• Whether more operators publish site-level WUE and water source type, not just fleet averages.
• Results from Microsoft's planned 2026 pilot sites in Phoenix and Wisconsin, once online.
• Whether regulators move from WUE targets to caps on total withdrawals or requirements to use reclaimed water.
• Updates to the LBNL work, which was requested under the Energy Act of 2020 as an update to its 2016 report.
• How the liquid cooling shift changes heat rejection choices, since that is where the water is decided.

For the software that runs on this infrastructure, Safeguard offers SBOM and dependency scanning, with a free scan.

Quick check

Sources

• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (eScholarship), December 2024, read 19 October 2026
• Lawrence Berkeley National Laboratory, report publication page, accessed 19 October 2026
• Amazon sustainability, AWS cloud page, accessed 19 October 2026
• Microsoft, next-generation datacenters that consume zero water for cooling, accessed 19 October 2026
• Microsoft sustainability report page, accessed 19 October 2026
• Meta sustainability, data centers, accessed 19 October 2026
• Infocomm Media Development Authority, Green Data Centre Roadmap, 2024, accessed 19 October 2026
• Joint Legislative Audit and Review Commission, Data Centers in Virginia, accessed 19 October 2026

Last reviewed: 19 October 2026.