How data center cooling methods compare on water use, once the power plant is counted
An evaporative cooling tower (which cools a data center by evaporating water into the outside air) consumes about 2.2 liters of water on site for every kilowatt-hour of server electricity, in the Lawrence Berkeley National Laboratory's (Berkeley Lab, a federal laboratory) 2024 simulations. A dry cooler with adiabatic assist (a radiator-and-fan unit that rejects heat to the air without evaporation, with water sprayed on the intake air on the hottest days) consumes about 0.2 liters. Liquid-cooled servers (servers whose chips are cooled by liquid piped to them) on a dry cooler consume close to zero. An air-cooled hall on a dry cooler draws more electricity than one on a tower, and every kilowatt-hour of it evaporates water at the power plant supplying the facility. A full comparison counts the water on site and the water at the power plant, and the grid decides which of the two is larger.
On every one of the eight US grids in the calculator below, the dry designs use less water in total than an evaporative tower once the power plant is counted. The size of the saving from an air-cooled dry design depends on the grid. It is largest on a grid supplied mostly by gas, wind and solar, such as Texas, and smallest on a grid whose hydropower reservoirs are charged to its electricity, such as Duke Energy Carolinas. Liquid-cooled servers on dry coolers save water on every grid by a wide margin, because they cut electricity use as well as site water. Texas is where that comparison is being argued in public.
Haynes Strader of Skybox Datacenters, a data center developer, told a state House committee on April 9, 2026 that his company's closed-loop buildings (cooling systems that recirculate the same water and evaporate almost none of it) use less water than five households. HARC (the Houston Advanced Research Center, a nonprofit research group) wrote in January 2026 that a pivot to air cooling on the current Texas grid could raise total water demand. Strader is counting the site term and HARC the power-plant term, and the two have to be added together before either claim can be tested.
How much water does each cooling design use on site?
The comparison uses two ratios that the Berkeley Lab report simulates for every cooling system it covers. PUE (power usage effectiveness) is total facility electricity divided by server electricity, and it rises when fans and chillers (the refrigeration units that produce chilled water for the server hall) replace evaporation. Site WUE (water usage effectiveness) is liters of water consumed on site per kilowatt-hour of server electricity. The report prints both ratios as a chart with no numeric table behind it. We digitized the report's medians from that chart, to a precision of 0.03 liters per kilowatt-hour.
The liquid-cooled dry design has the lowest PUE in the set as well as the lowest site water use, because liquid cooling lifts the coolant temperature high enough for dry coolers to reject the heat to the outside air through most of the year. An air-cooled server hall trades water for electricity when it switches from a tower to a dry cooler.
A liquid-cooled hall of the kind now built for AI makes no such trade, because its dry design uses less electricity and less water than the tower design at the same time. Microsoft's design standard since August 2024 uses the same temperature margin. Its facilities run chip-level closed loops with no evaporation for cooling. The company says the electricity penalty of that design is offset by running the coolant at warmer temperatures.
How much water does the power plant add, and why does it depend on the grid?
The power plant adds a second term to each design's site figure, equal to the facility's electricity multiplied by the water intensity of the grid supplying it (the liters of water evaporated at power plants for every kilowatt-hour they deliver). Berkeley Lab publishes an hourly dataset of water consumed per kilowatt-hour for every US balancing authority (the organization that keeps supply and demand matched on one section of the grid), built from federal plant-level generation and cooling data. The dataset carries two accounting methods. The primary method charges reservoir evaporation at hydropower dams (water that evaporates from the lakes behind the dams) to the electricity they generate, and the second method leaves that evaporation out and counts thermoelectric cooling water alone.
Our consumption-weighted 2022 averages of that file (each hour weighted by the electricity used in it) run from 0.84 liters per kilowatt-hour on ERCOT (the Electric Reliability Council of Texas, which runs the grid covering most of the state) to 15.4 liters on Duke Energy Carolinas, under the primary method. ERCOT's figure is the lowest of the ten largest US grids by the laboratory's own accounting. Its 2022 supply was mostly gas, wind and solar with almost no hydropower, and reservoir evaporation at hydropower dams supplies most of the national average.
Cooling design (large-scale archetype) | PUE | Site water, L per server kWh | Total on ERCOT | Total on PJM | Total on US average | Total on Southern Company | Total on Duke Energy Carolinas |
Evaporative tower (water-cooled chiller with waterside economizer) | 1.12 | 2.18 | 3.12 | 4.38 | 6.97 | 11.43 | 19.48 |
Liquid-cooled servers, evaporative tower | 1.13 | 1.96 | 2.91 | 4.18 | 6.80 | 11.29 | 19.41 |
Airside economizer with adiabatic cooling, water-cooled chiller backup | 1.14 | 0.42 | 1.38 | 2.66 | 5.30 | 9.84 | 18.03 |
Dry cooler with adiabatic assist, air-cooled chiller backup | 1.20 | 0.22 | 1.23 | 2.58 | 5.36 | 10.13 | 18.75 |
Liquid-cooled servers, dry cooler, air-cooled chiller backup | 1.09 | 0.00 | 0.91 | 2.14 | 4.67 | 9.00 | 16.84 |
Totals are liters per server kilowatt-hour, site water plus PUE multiplied by each grid's 2022 water intensity under the primary method (ERCOT 0.839, PJM 1.964, US average 4.281, Southern Company 8.261, Duke Energy Carolinas 15.445). Site figures are our digitization of Figure 4.4 in the Berkeley Lab report, the analysis is Gasilov Group's, and MISO, SPP and CAISO are in the calculator.
On ERCOT, the evaporative tower's total is 3.12 liters against 0.91 for liquid-cooled servers on dry coolers, and the site term makes up most of the tower's total. On PJM, which serves northern Virginia, the tower's total is a little over twice the liquid-cooled dry design's. The dry designs stay below the tower on every grid in the exhibit. The two middle rows swap places once the grid figure passes about 3.3 liters, which is why the airside economizer edges the dry cooler on the US average and wetter columns: its lower PUE outweighs its extra site water there.
On Duke Energy Carolinas, where reservoir evaporation raises the grid figure to 15.4 liters, the tower's total comes to 19.5 liters and the liquid-cooled dry design's to 16.8. The site term that dominates in Texas is a small share of the total there, and the cooling choice moves the total by a smaller margin than the choice of grid does.
With reservoir evaporation excluded from the count, every grid in the table falls between 0.64 liters (CAISO) and 1.57 liters (MISO), and the totals for all of them then look much like the ERCOT column.
When does a switch to dry cooling raise total water use?
A switch from one cooling design to another stops saving water at the grid intensity where the added power-plant water equals the water removed on site. That break-even intensity equals the site water removed divided by the PUE added. For a large hall switching from an evaporative tower to a dry cooler with adiabatic assist, the switch removes 1.96 liters per server kilowatt-hour on site and adds 0.08 to the PUE. The break-even for that switch therefore comes to 24.5 liters of grid water per kilowatt-hour delivered.
With reservoir evaporation excluded from the count, no US grid reaches 2.1 liters, and no switch in the table loses. With it counted, three balancing authorities that consume more than a terawatt-hour (a billion kilowatt-hours) a year exceed 24.5 liters: the Los Angeles Department of Water and Power, South Carolina Electric & Gas, and NorthWestern Energy in Montana. On those three grids, and under that accounting alone, the switch from a tower to a dry cooler with adiabatic assist raises total water use.
For the midsize chiller pair in the same simulations, a water-cooled chiller (one that sheds its heat through an evaporative tower) against an air-cooled one (one that sheds its heat through fans), the break-even is 9.2 liters per kilowatt-hour. Eleven of the fifty balancing authorities above a terawatt-hour exceed that level under the primary method, Duke Energy Carolinas at 15.4 liters among them. A midsize facility on one of those grids that moves to an air-cooled chiller alone, without an economizer, uses more water in total than it did on the tower.
Liquid-cooled servers on a dry cooler have no break-even at all, because that design has a lower PUE than the tower as well as a lower site WUE. No grid intensity makes that switch a net loss. The calculator below runs the same arithmetic for any load, design and grid, including a custom grid intensity, and reports the break-even for any pair of designs.
Figure 1. Annual water consumption on site and at the power plants serving the facility, by cooling design and grid. For 100 MW of servers on ERCOT, an evaporative tower consumes 505 million gallons a year on site and 217 million at power plants; liquid-cooled servers on dry coolers, zero and 212 million. Analysis: Gasilov Group, as of September 2026.
Our water strategy advisory work on siting starts from the balancing-authority water intensity for the county before the cooling design is fixed. Our earlier post on how much water a data center uses covers the site side in more detail.
The Texas dispute, run through the same arithmetic
HARC's caution comes from a comparison of percentage changes, and its own paper sets out that comparison in a stoplight table (a table that rates each option green, yellow or red). The paper puts the air-cooled chiller's electricity penalty at 20 percent for a midsize facility. The stoplight table then weighs a percent change in water efficiency against an equally sized percent change in energy efficiency. Scored on equal percentages, the energy side wins in HARC's table. Measured in volumes, a full switch from a water-cooled chiller to an air-cooled one removes 98 percent of site water in the same simulations. The extra water evaporated at Texas power plants to cover that 20 percent electricity penalty comes nowhere near the volume removed on site.
For a large hall on ERCOT, the tower-to-dry-cooler switch in the table removes 1.96 liters per server kilowatt-hour on site and adds about 0.07 liters at the power plants. The electricity penalty costs about 4 percent of the water saved. ERCOT's intensity sits far below the 24.5-liter break-even, and HARC's own paper reports a Texas power-plant figure close to ours.
Strader's per-facility figure leaves out the same power plant that HARC's paper counts, and the table's figures can be run against his own numbers. He put 800 megawatts of large data centers in operation in Texas, with a fifth of them on evaporative cooling. We take that 800 megawatts as facility load. The server load is then the facility load divided by each design's PUE. On that basis the evaporative fifth consumes 720 million gallons a year on site.
The closed-loop remainder consumes between nothing and 270 million gallons a year on site, depending on how much adiabatic assist it uses. The power plants serving the whole fleet consume 1.55 billion gallons a year. In the Texas fleet as Strader described it, more water is now evaporated at power plants than in cooling towers. The towers that remain account for between roughly three quarters and all of the on-site total.
Measured in the EPA's average household use (the Environmental Protection Agency's WaterSense figure for a US home), 100 megawatts of liquid-cooled servers on dry coolers on ERCOT consumes the water of zero households at the site. That figure sits under Strader's five, and it is the figure a per-facility count reports. The power plants serving those same 100 megawatts consume the water of about 1,900 households a year on ERCOT's 2022 intensity, and both figures are true at once.
Governor Abbott's directive of August 3, 2026 orders the Public Utility Commission of Texas (PUC) and ERCOT to audit every data center in the interconnection queue (the list of projects waiting for approval to connect to the grid) before any more of them advance. The audit collects the cooling technology, the projected annual and peak water consumption, and the generation type of any on-site power, in the same request. Nothing in the directive asks anyone to combine the water figure with the generation type into a total that includes the power plant. The earlier survey the directive cites drew responses from fewer than a third of the state's 341 data centers. The commission's July 2026 response to the governor had already proposed that the Legislature require data centers to register with their electricity and water use.
The 90th Texas Legislature convenes on January 12, 2027, and any registration bill it passes will write a water line that counts either the cooling tower alone or the cooling tower and the power plant together. I would write it to count both, with the balancing authority named on the same line as the site figure, and the same choice faces any state writing a data center water rule.
What the power contract decides that the cooling choice cannot
Once a facility is on closed-loop cooling, almost all of the water it consumes is evaporated at the power plant. The size of that power-plant volume depends on what generates the electricity. The generation source is chosen in the power contract, and that choice moves more water than anything left in the cooling design. A gas combined-cycle plant (one that runs a gas turbine and then a steam turbine on the exhaust heat) on cooling towers consumed 244 gallons per megawatt-hour in Texas, on HARC's reading of federal plant data. A dry-cooled gas plant, a wind farm or a solar array consumes close to none.
I argued in OGEL Energy Law Journal in April 2026 that water scarcity in the Colorado River Basin already constrains thermal generation in ways that utility resource plans leave out, and a data center's power contract carries that same constraint. A developer choosing between a grid connection, an on-site gas plant and a renewable contract therefore makes the larger water decision at the contract stage. The calculator shows the size of that decision when the grid preset is changed. The cooling design then decides a smaller volume than the contract does.
I would put the grid figure on the same line as the site figure in any water disclosure a facility files, with the balancing authority named. The practical effect is that a Texas closed-loop site reports 0.9 liters per server kilowatt-hour where it now reports zero, and a site on a hydro-heavy grid reports the number that decides its siting. Contracts follow from the same line: a power purchase agreement (a contract to buy electricity from a named plant or supplier) can specify a dry-cooled or renewable supply, and an on-site gas plant can be specified with dry cooling at a known efficiency cost. A related recommendation I gave Developing Telecoms in August 2026 for India is that every facility report both water withdrawn and water consumed.
For the facilities still on cooling towers, wherever they are, the exposure comes from timing, because a tower's water draw peaks on the hottest afternoons of the year. Al Jazeera ran a line of mine in July 2026 on that coincidence: a facility on evaporative cooling in a heatwave draws on the same supply residents are already under restrictions for. The gas plant on a cooling tower behind that facility peaks on the same afternoons.
If you are not sure where to start, the two-week water exposure screen is the entry point. If you already know what you need, get in touch and we can scope it.
Arif Gasilov is partner for Natural Resources & Built Environment at Gasilov Group, where he works on energy regulatory analysis, water governance, utility rate cases, and the resource footprint of large development. Meet our partners →
Frequently Asked Questions (FAQ)
Is a closed-loop data center the same as a zero-water data center?
No. A closed loop still evaporates water where adiabatic pads spray the intake air on hot days. The loop itself takes a one-time fill, which Skybox put at up to 30,000 gallons a building. Domestic use (kitchens, restrooms and the like) continues. Halls with liquid-cooled servers on dry coolers reach zero water for cooling in the Berkeley Lab simulations, with domestic use the remaining line.
Does the national figure of about 4.3 liters of power-plant water per kilowatt-hour apply to every grid?
No. That national average charges reservoir evaporation at hydropower dams to electricity, and reservoirs supply most of it. A grid's figure depends on its own generation mix and on which accounting method is used. ERCOT has almost no hydropower, and its figure barely moves between the two methods: 0.84 liters with reservoir evaporation counted and 0.73 without. CAISO moves from 4.48 liters to 0.64 on the same change, and Duke Energy Carolinas from 15.4 to 1.35.
