Data Center Water Usage: Figures, WUE, Cooling Choices and Reduction

Data center water usage is the water a data center consumes on its own site, mostly as evaporation from the cooling plant, together with the much larger volume consumed at the power plants that generate its electricity. Paul Industries designs, installs and maintains the cooling tower water treatment, makeup water treatment, reuse piping and closed-loop and dry cooler piping that decide how much water a site draws from its local supply. Below are the national figures, WUE with worked numbers, the effect of cooling design, where the water goes and how to cut it. For loops, treatment and piping in general, see the data center water treatment and cooling loops hub.

What data center water usage means: direct and indirect water

Two different quantities travel under one name. Direct, or site, water is what the facility itself consumes: makeup for cooling towers, adiabatic pads and evaporative coolers, humidifier feed and a small domestic load. Indirect, or source, water is consumed off site at the power plants that generate the electricity the facility buys, through evaporation in thermoelectric plant cooling and from hydroelectric reservoirs.

Data center water consumption, in either sense, means water that evaporates or is otherwise removed from the local water cycle. That is different from withdrawal. A once-through cooling system can withdraw a great deal of water and return most of it a few degrees warmer, while a cooling tower withdraws far less but evaporates most of what it takes.

Site and source water move in opposite directions as designs change. A plant that stops evaporating water usually runs more compressors or fans, and the extra electricity carries its own indirect water, so a fair comparison of two designs looks at both. Most facility metrics report only the site figure, the part an owner can directly control.

How much water do data centers use? The 2024 Berkeley Lab figures

The most complete national estimate comes from the Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report (Shehabi et al., December 2024), prepared for the Department of Energy. It puts direct water consumption by US data centers at about 66 billion liters in 2023, roughly 17.4 billion gallons, compared with 21.2 billion liters in 2014. That is the national data center water usage per year; divided by 365 it gives a national data center water usage per day of roughly 181 million liters, close to 48 million gallons.

The mix shifted too. Internal, or enterprise, data centers held 64 percent of direct water use in 2014 but only 12 percent in 2023, when hyperscale and colocation facilities accounted for 84 percent; the report expects hyperscale facilities alone to consume 60 to 124 billion liters in 2028.

So do data centers use a lot of water? On the direct measure the national total is small next to agriculture and power generation, as the comparison further down shows, but it is concentrated: one campus draws on one utility or one aquifer, often hardest in a few hot months, and that local draw is where the number matters. The indirect footprint is about twelve times larger and sits wherever the power is generated. Both are true at once, which is why answers differ.

Water usage effectiveness (WUE), with worked numbers

Data center water usage effectiveness is a facility’s annual site water consumption divided by the energy delivered to its IT equipment, stated in liters per kilowatt-hour (L/kWh). The Green Grid introduced the metric in 2011, and ISO/IEC 30134-9 now standardizes how it is measured and reported. Like PUE it is an annual figure; a summer week can run far above it. The Berkeley Lab report separates site WUE, which counts only the facility’s own water, from source WUE, which adds the water consumed generating its electricity.

The water usage effectiveness formula in practice

In formula form, WUE equals annual site water in liters divided by annual IT energy in kilowatt-hours. The usual way to answer how much water does a data center use is to start from its WUE. Multiply the average IT load in kilowatts by 8,760 hours to get annual IT energy, then multiply by WUE. A facility averaging 1 MW of IT load uses 8,760,000 kWh a year; at the national average of 0.36 L/kWh that comes to about 3.15 million liters a year, or 8,640 liters a day. Those are the data center water usage per MW figures that follow from the national average, and they scale in a straight line: a campus averaging 50 MW at the same WUE would consume about 158 million liters a year. Use the average IT load, not the nameplate.

How much water does a data center use per day, then? For a single facility the honest answer is the spread in that table, from a few thousand liters per megawatt of IT load for designs that seldom evaporate water to tens of thousands for designs that evaporate all year. Ask the designer for design WUE in the actual climate and the peak-day makeup flow.

Why do data centers need water?

Data centers need water because evaporating it is the most energy-efficient common way to push large amounts of heat into the outdoor air. Every kilowatt-hour a server draws ends up as heat that must leave the building, and evaporating one kilogram of water at typical cooling tower temperatures absorbs about 2,430 kilojoules, a little over 1,000 Btu per pound.

That property sets a physical scale. Rejecting one kilowatt-hour of heat entirely by evaporation takes roughly 1.5 liters of water, so one megawatt of heat rejected around the clock that way evaporates about 35,600 liters, near 9,400 gallons, per day. Real towers reject most, not all, heat by evaporation, and tower heat includes chiller and pump energy as well as IT load, so treat this as an order of magnitude.

How do data centers use water for cooling?

In a conventional water-cooled plant, chillers move heat from the chilled water loop that serves the white space into a condenser water loop, and pumps carry the condenser water to a cooling tower. In the tower, warm water falls across fill while fans pull air through it, and a small fraction evaporates. EPA’s WaterSense at Work guidance puts evaporation at about 1 percent of the recirculating flow for every 10°F of cooling the tower achieves. A tower circulating 10,000 gallons per minute through a 10°F range therefore evaporates about 100 gallons a minute, or around 144,000 gallons a day at full load.

Two other routes use water directly. Adiabatic systems wet pads or spray mist into the air ahead of a dry cooler coil or an air handler on hot days, lowering the air temperature by evaporation, and they draw water only while that mode is on. Humidification adds a small load. The chilled water loop and the rack-side liquid cooling loops are sealed and consume essentially no water in normal operation; how those loops are arranged is covered on data center cooling systems.

Data center water use by cooling design

Cooling design is the largest single lever on data center water use. The Berkeley Lab simulations, run more than a million times across cooling systems, space types and weather data from 965 US weather stations, placed water-cooled chillers without economizers at the top of the WUE range because of cooling tower water, waterside economizers below them, airside economizers lower again, dry coolers with adiabatic assist drawing water only in wet mode, and the remaining dry designs using little beyond occasional humidification.

The trade-off in that table runs one way. Designs that save water use more electricity on hot days, because dry heat rejection works against the outdoor dry-bulb temperature while a cooling tower can bring water within a few degrees of the lower wet-bulb temperature. The Berkeley Lab report makes the point directly: a low site WUE is not automatically better, since evaporative systems are generally more energy efficient than waterless ones. Equipment choices on the chilled water side are covered under chilled water systems.

AI data center water usage vs regular data center

AI data center water usage differs from that of a conventional facility mainly because of rack density and the cooling it forces, not because accelerators consume water themselves. Racks run at densities air cannot cool, so heat goes into liquid at the chip and then to facility water through a coolant distribution unit, and where that heat goes next decides the water figure.

That makes the effect cut both ways. Liquid cooling can return water to the facility loop far warmer than an air-cooled hall can, and the ASHRAE liquid cooling classes renamed in 2021 (W17, W27, W32, W40, W45 and W+, each named for its upper facility water supply temperature in degrees Celsius) let designs at W32 and above reject heat without chillers for most or all of the year. The Berkeley Lab report modeled liquid-cooled AI space at class W45 and notes that warmer facility water extends economizer hours and can lower WUE. The same report still projects the national average site WUE rising from just over 0.36 to between 0.45 and 0.48 L/kWh by 2028, citing higher WUE in hyperscale and colocation facilities and the water consumed by liquid-cooled systems, whose modeled plants include water-cooled chillers with waterside economizers. AI data center water consumption therefore turns on whether the warm water ends up at a dry cooler or at an evaporative tower.

AI data center water use also grows on the electricity side. The report puts 2023 data center electricity at 176 TWh and projects 325 to 580 TWh by 2028, between 6.7 and 12 percent of US consumption. At the 2023 average indirect intensity of 4.52 L/kWh, each additional terawatt-hour carries about 4.5 billion liters of water consumed at generating plants. The hardware is described on data center liquid cooling.

Is AI data center water usage a problem?

It can be a local issue and is rarely a national one. The national direct figure is small beside agriculture and power generation, but a large campus drawing evaporative makeup from a stressed aquifer or a small municipal system during peak summer demand can matter a great deal to that system. That depends on the cooling design, the makeup source, the cycles of concentration and the local supply. Each can be engineered, as the reduction section shows.

Closed loop vs open loop: is data center water recycled?

Is data center water recycled? It depends on the loop. Water in the closed loops, meaning chilled water, glycol and the technology cooling loop that feeds cold plates, circulates sealed for years and is topped up only to replace small leaks, sampling and occasional drain-downs for maintenance, so data center water usage closed loop figures are close to zero once a loop is filled and treated. The open loop is the condenser water circuit through a cooling tower: it recirculates too, but each pass evaporates some water and concentrates the minerals left behind, so it needs constant makeup and bleed.

Holding tower water in circulation longer before bleeding it off, raising the cycles of concentration, is the main form of recycling in an evaporative plant.

Can data center water be reused?

Yes, in several engineered ways. Tower blowdown can be filtered and run through reverse osmosis, with the permeate returned as makeup and a smaller, more concentrated reject stream sent to drain. Treated municipal wastewater and other non-potable supplies can replace drinking water as makeup, at the cost of heavier pretreatment and closer biological control. Air handler condensate is low in minerals and, as EPA notes, most plentiful when tower loads peak. The heat can be reused as well: warm return water from liquid-cooled halls can serve building or district heating through a plate heat exchanger, so that heat is used instead of being evaporated away at the tower.

Where does data center cooling water go?

Cooling water leaves a data center in the four ways EPA lists for any cooling tower: evaporation, blowdown, drift, and leaks or overflows. Evaporation is the largest and is the tower’s purpose; it leaves as vapor in the plume. Blowdown, also called bleed-off, is the smaller deliberate drain that keeps dissolved solids in check and normally goes to the sanitary sewer. Drift is fine droplets of tower water carried out by the fans. Leaks and overflows should be zero.

What is in blowdown? Mostly the same minerals that arrived in the makeup water, including calcium, magnesium, chloride, sulfate and silica, concentrated by the cycles of concentration: at five cycles, about five times the makeup level. It also carries residual inhibitors and biocide, plus solids scrubbed from the air. Discharge to a sanitary sewer is governed by the receiving utility’s permit and local limits under the federal pretreatment rules in 40 CFR Part 403, and some sites adjust pH or remove solids before discharge. That step falls under process wastewater and effluent treatment.

Drift deserves attention for a second reason: it is the route by which bacteria in tower water, Legionella among them, can leave the tower. ANSI/ASHRAE Standard 188 calls for a water management program for building water systems that include cooling towers, and sound drift eliminators, controlled biocide feed and routine cooling tower cleaning are all part of meeting it.

Data center water usage compared to other industries

Comparisons only hold when the units match, and many published ones mix three different quantities: withdrawal (water taken from a source, much of which may be returned), consumption (water evaporated or otherwise not returned), and applied water (irrigation put on land, part of which soaks back into groundwater). The Berkeley Lab figures for data centers are consumption. The US Geological Survey’s 2015 compilation, Circular 1441, reports consumption only for thermoelectric power and irrigation, so only those two compare like for like.

Read with those caveats, direct consumption by data centers in 2023 was around 1 percent of what thermoelectric plants consumed in 2015 and well under a tenth of a percent of irrigation consumption. The indirect figure is an order of magnitude larger and is consumed at power plants and reservoirs, so it overlaps the power sector’s own footprint rather than adding to it. The years differ, so treat the ratios as orders of magnitude.

Data center water usage vs golf course

The golf comparison needs care. The Golf Course Superintendents Association of America’s 2022 water survey estimated that US golf facilities applied 1.68 million acre-feet of water in 2020, about 2.07 trillion liters, with a median of 66.3 acre-feet, about 82 million liters, per facility. On national totals, data centers’ direct consumption in 2023 was about 3 percent of golf’s applied water. At the facility level, one median golf facility’s annual water equals the direct consumption of about 26 MW of average IT load at the 0.36 L/kWh national average, but only about 4.7 MW at an evaporation-heavy 2.0 L/kWh. Golf water is applied seasonally and partly returns to groundwater, while tower water leaves as vapor year-round, so the answer depends on the assumptions chosen.

Can data centers not use water?

Yes: a data center can be built to consume almost no water on site by rejecting all of its heat through air-cooled chillers or dry coolers with no adiabatic assist. The cost is electricity. On hot days dry heat rejection works against the dry-bulb temperature, so compressors run harder and longer, and the Berkeley Lab report states plainly that air-cooled chillers use no water but more energy. That extra electricity carries indirect water at the generating plant, 4.52 liters per kWh on average for data center electricity in 2023, so a zero-water site does not have a zero-water footprint. In a mild climate, or with warm-water liquid cooling, the penalty can be small.

The most persistent data center water usage myth runs in two opposite directions. One version holds that every data center drinks vast volumes; in fact a site on dry heat rejection consumes little beyond domestic water. The other holds that data centers use no water because their loops are closed; the server-side loops are closed, but an evaporative plant consumes water continuously at the tower, and every site consumes indirect water through its electricity. A reported WUE is also often read as the total, when it covers site water only.

How to reduce data center water usage

To reduce data center water usage without simply shifting the burden onto the electricity bill, work through the plant in order of payback: stop waste first, then raise cycles, then replace potable makeup, and only then change how the heat is rejected. The table summarizes each measure and what it trades away.

Cycles of concentration: the arithmetic

Cycles of concentration is the ratio of dissolved solids in the tower water to those in the makeup, which EPA notes is also roughly the ratio of makeup volume to blowdown volume. With drift neglected, blowdown equals evaporation divided by (cycles minus one), and makeup equals evaporation plus blowdown. Evaporation is fixed by the heat load, so raising cycles only shrinks the blowdown part of makeup, which is why the gains flatten quickly above about six.

EPA’s WaterSense at Work guidance notes that many systems run at two to four cycles and that six or more may be possible. The ceiling is set by makeup chemistry, chiefly hardness, alkalinity and silica, and by the treatment program. Softened or RO-treated makeup can lift that ceiling considerably, provided corrosion control is adjusted for the softer water. That is where commercial water softeners and reverse osmosis come in, and the trade-offs between purification methods are set out in RO vs DI water.

Design changes that take evaporation out of the plant

On new halls and major retrofits, the bigger savings come from changing what rejects the heat. Raising chilled water and facility water supply temperatures widens the hours in which dry coolers or economizers carry the load. Liquid cooling at W32 or warmer lets a dry cooler plant carry nearly all hours in many climates, and heat recovery through heat exchangers removes some heat from the tower altogether. Each changes the electrical load, so judge them on total water, site plus source, for the local grid. The installation side is covered on data center mechanical contractors, and proving the new plant hits its design setpoints and cycles at handover through data center commissioning.

What Paul Industries does to lower water use at data centers

Paul Industries works on the water side of the cooling plant. On existing evaporative plants, crews install and maintain conductivity controllers, chemical feed and sidestream filtration equipment, makeup and blowdown meters and the rest of the cooling tower water treatment equipment that lets a tower run at higher cycles, along with tower cleaning and mechanical repair. For makeup, Paul designs and installs softening, RO and other industrial water treatment and high-purity water systems, including blowdown recovery and piping for reclaimed or condensate makeup. On new and converted halls, Paul fabricates, installs, flushes, pressure-tests and starts up the closed-loop, dry cooler, heat exchanger and liquid cooling piping that moves heat rejection away from evaporation, then keeps it running under preventive maintenance, with water sampling and meter reviews that show whether the plant is holding its design cycles. To scope a project, request a quote.

Standards referenced: The Green Grid WUE metric · ISO/IEC 30134-9 · ASHRAE liquid cooling classes W17 to W+ · ANSI/ASHRAE Standard 188 · EPA WaterSense at Work · 40 CFR Part 403

Frequently asked questions

What is the difference between water withdrawal and water consumption for a data center?

Withdrawal is all the water a facility takes in; consumption is the part that does not go back to the local source, mainly evaporation at the tower. A data center with cooling towers consumes most of what it withdraws, with the blowdown fraction returned to the sewer system. Permits and utilities usually meter withdrawal, while national estimates such as the Berkeley Lab figures describe consumption, so the two numbers for the same site will not match.

What is a good WUE for a data center?

There is no single target, because WUE depends on climate and on the cooling design, and a low value can come at the price of higher electricity use. A more useful test is whether a site meets the design WUE its engineers predicted for its own climate, year after year. Reporting site and source WUE side by side shows whether a water saving was real or simply moved to the power plant.

Does WUE count the water used to generate electricity?

Site WUE does not; it counts only water consumed at the facility. Source WUE adds the water consumed generating the electricity, and because data center electricity averaged 4.52 liters of indirect water per kWh in 2023 against a site average just over 0.36, the source term is usually the larger of the two. Which figure a report uses should be stated next to the number.

How is WUE measured at a real site?

Every water input serving cooling and humidification is metered, usually makeup to each tower and adiabatic system plus any humidifier feed, and IT energy is metered at the UPS output or the power distribution units. Both are totalled over twelve months and divided. ISO/IEC 30134-9 sets out measurement categories and reporting rules, so a quoted WUE should say which category it follows.

Why does data center water use peak in summer?

Hot, humid weather cuts economizer hours, so chillers and towers carry more of the load, and adiabatic systems switch to wet mode on the hottest days. A plant can draw several times its annual average on a design-day afternoon. Makeup supply, storage and treatment equipment have to be sized for that peak, not for the yearly mean.

How many cycles of concentration can a data center cooling tower run?

It depends on the makeup water. The limit is usually whichever constituent first reaches its scaling threshold, often calcium carbonate, calcium phosphate or silica, and the treatment program sets how close to that threshold the tower can safely run. Softened or RO-treated makeup removes the limiting constituents, which is why sites with hard or high-silica water gain the most from makeup pretreatment.

Can reclaimed water be used for data center cooling towers?

Yes, and it is one of the most effective ways to cut potable demand. Reclaimed water typically carries more nutrients, ammonia, phosphate and suspended solids than drinking water, so it needs filtration, closer biocide control and a treatment program written for that supply. It also needs separate, clearly identified piping and backflow protection so it can never cross into potable lines.

Is drinking water used to cool data centers?

Often, yes, because municipal potable water is reliable, consistent and already at the site. It is not required. Reclaimed municipal water, recovered blowdown, captured condensate and other non-potable supplies can all serve as makeup when the treatment is designed for them, and switching even part of the makeup to those sources reduces demand on the drinking water system.

Does liquid cooling reduce water use?

Not by itself. Liquid cooling moves heat from the chips into water more efficiently, but the facility still has to reject that heat outdoors. If the warm water goes to dry coolers, site water use can fall to near zero; if it goes to an evaporative tower, the tower still evaporates water. The saving comes from the warmer water temperatures liquid cooling allows, which make dry heat rejection practical for more of the year.

Do closed cooling loops ever need makeup water?

Yes, in small amounts: sampling, valve and pump seal weeps, expansion tank adjustments and drain-downs for maintenance all remove water. A meter on the closed-loop makeup line is one of the best early warnings of a leak, because steady makeup on a sealed loop always means water is escaping. On glycol loops, plain water makeup also dilutes the glycol, so concentration has to be checked after any top-up.

Does cooling tower drift contain chemicals?

Drift has the same composition as the water in the tower basin, so it carries the concentrated minerals, residual treatment chemicals and any bacteria present in that water. That is why drift eliminators matter even though the volume is small. Damaged, missing or poorly seated eliminator sections should be replaced promptly, and inspecting them is a routine part of tower maintenance.

Why not use dry coolers everywhere?

Dry coolers need more fan power, more coil area and more plot space for the same heat, and on hot days they cannot reach the water temperatures a tower can, so chillers have to make up the difference. In a hot climate that can raise electricity use enough to outweigh the water saved once the indirect water at the power plant is counted. The choice is a site-specific balance of water availability, power cost, climate and space.

How much water does adiabatic cooling use compared with a cooling tower?

Far less over a year in most climates, because an adiabatic system wets its pads or sprays only during the hottest hours, while a tower evaporates water whenever it rejects heat. During those peak hours the adiabatic system's flow can be significant, so supply sizing still matters. Spray systems usually need softened or RO-treated water to avoid scale on pads and coils.

How do I estimate water use for a planned data center?

Start with the cooling design and run an hour-by-hour simulation against typical weather for the site, the same approach the Berkeley Lab report used across 965 weather stations. Combine the evaporation that results with the planned cycles of concentration and the expected average IT load. Ask the design team for annual consumption, peak-day makeup flow and peak-day blowdown flow, since the utility will need all three.

What is the water intensity of electricity?

It is the volume of water consumed per kilowatt-hour generated, mostly through evaporation at thermoelectric plants and from hydroelectric reservoirs. The Berkeley Lab report put the US average at 4.35 L/kWh in 2023 and found that the grid serving data center locations averaged 4.52 L/kWh. The figure varies widely by region, so the same facility can carry a very different indirect footprint in different places.

Does heat reuse save water?

Yes, during the hours when a heat user actually takes the heat, because that heat no longer has to be evaporated at the tower. The savings depend on a steady nearby demand such as building heating, district heating or a process, and on supply temperatures high enough to be useful, sometimes raised with heat pumps. Outside those hours the tower or dry cooler still carries the load.

What happens to the minerals removed by reverse osmosis on makeup water?

They leave in the RO concentrate, a reject stream that goes to drain or, at some sites, is blended into tower water where its chemistry allows. That reject is a water cost, so RO on makeup saves water overall only when the reduction in tower blowdown exceeds the concentrate volume. The balance is usually favorable on hard or high-silica supplies and should be calculated for each site.

Does data center water go back to rivers and lakes?

The evaporated part does not return locally; it leaves as vapor and falls elsewhere as precipitation. Blowdown goes to the sanitary sewer, through the municipal treatment plant, and is then discharged to a receiving water, so that fraction does return. Water held in closed loops stays in the building until a loop is drained for maintenance.

Related: Data Centers · Data Center Cooling Systems · Cooling Tower Water Treatment · Industrial Water Treatment · Process Wastewater, Effluent & Chemical Sanitization · Cooling Tower Cleaning · Request a quote · Data Center Cooling Towers: Installation, Redundancy and Repair