A Data Center Digest Original StoryThis article was researched and written entirely by AI, without human review or editing, as part of the Data Center Digest's ongoing experiment in AI-powered journalism.

Water is the data center fact most likely to arrive without its context. A county says a project is contracted for 2 million gallons a day. A company says its new campus consumes essentially no water for cooling. A researcher says the industry uses billions of gallons. These statements can sit side by side because they measure different things: different parts of the cooling system, different points in the water's trip, and sometimes different power plants. This piece starts with a car and a house, then builds up to the vocabulary that sorts the numbers out.

Start with the car and the house

Every data center has one job that never stops: moving heat from the chips to somewhere else. Two familiar machines show the basic choices. A car engine circulates coolant in a sealed loop. The coolant picks up engine heat and carries it to the radiator, where passing air takes the heat away. Nothing is used up. A home air conditioner works on the same idea: a sealed loop of refrigerant carries heat outdoors, and a fan blows air across the outdoor unit to release it. Neither machine needs a water supply.

In data center terms, these are closed loops that hand their heat to the air, usually through equipment called dry coolers. The cost is electricity for fans and compressors, and the work gets harder on hot days, when the air that has to take the heat is already warm.

The other approach is closer to sweating. A cooling tower sends warm water through a stream of moving air, and a small portion of it evaporates, carrying heat away. The same effect makes a swamp cooler work in a dry climate, and it is why a breeze on wet skin feels cold. Evaporation removes heat very effectively for little electricity. The price is the evaporated water, which leaves as vapor and does not return to the pipe or the river. This is what people mean by evaporative cooling, and it is often what “open loop” means in data center talk; University of Georgia Extension material uses the label that way.

“Closed loop” describes only half the system

Researchers who study data center water describe a two-stage process. In the first stage, heat moves from the servers to a heat exchanger inside the facility, through air or through a liquid loop that is sealed and uses no water up. In the second stage, the facility sends that heat outdoors, either through wet cooling towers that evaporate water or through air-cooled dry coolers. That description comes from a March 2026 preprint by Yuelin Han, Pengfei Li, Adam Wierman and Shaolei Ren of UC Riverside, RIT and Caltech. A building can have a sealed liquid loop in the first stage and still use cooling towers in the second. Margaret Cook of the Houston Advanced Research Center put the consequence plainly in CBS Austin's reporting: “Whatever that technology is will dictate the water consumption.”

That is why the same words appear on both sides of the debate. QTS Data Centers told Texas lawmakers that every facility it has built since 2019 uses a closed-loop system that needs essentially no water for cooling once operating. Berkeley Lab's national model found the opposite end of the range: data centers with water-cooled chillers and no economizers, the equipment that lets a chiller rest when the weather allows, had the highest on-site water use, because of cooling tower evaporation. Hyperscale sites, it found, commonly use outside air for much of the year with only occasional water-assisted cooling, which keeps their water use minimal.

The trade: water here, electricity there

Neither choice is free. Evaporative cooling uses substantially more water on site than dry cooling, and it uses less electricity. The UC Riverside preprint cites a large technology company's statement that at its Louisiana campuses, evaporative cooling cuts electricity demand by 25 to 35% compared with dry cooling during summer peaks. Google says water cooling suits sites where the local source is not at high risk of depletion, and reports that its Oregon campuses in The Dalles needed only 6% to 10% extra energy for overhead in the fourth quarter of 2025, against a 54% industry average, which it ties in part to water cooling. Those are company figures.

Dry cooling moves part of the footprint somewhere else. Cook says closed-loop and air-cooled designs raise power needs. Microsoft says its newest design, which circulates coolant at the chip and uses no evaporation, raises annual energy use by a small amount compared with its evaporative designs, partly offset by running warmer. It says the design avoids more than 125 million liters of water a year per data center, based on its fiscal 2024 fleet average, with pilot designs planned for Phoenix and Mount Pleasant, Wisconsin, in 2026 and new sites coming online from late 2027.

Power plants use water too. Berkeley Lab estimated that U.S. data centers consumed about 66 billion liters (about 17 billion gallons) directly in 2023, and nearly 800 billion liters (roughly 211 billion gallons) indirectly through the electricity they bought, about 12 times as much. That estimate is based on regional grid mixes and does not account for power purchase agreements or on-site generation. The comparison depends on the local grid and the local water source, so a design that saves water on site is not automatically better or worse once the power plant is counted.

Withdrawn, consumed, discharged

The second source of confusion is the verb. Water withdrawal is water taken from a river, an aquifer or a utility. Consumption is the part that does not come back, usually because it evaporates. Discharge is the difference: the part returned right away, typically to a sewer or treatment plant. A shower and a lawn sprinkler show the gap. Nearly all the water from a shower goes down the drain and back to treatment, while much of what a sprinkler puts out evaporates or is taken up by the lawn. The UC Riverside preprint estimates that homes consume 5 to 15% of what they withdraw, public water supplies as a whole about 12%, and data centers 70 to 90%.

Google's reporting for 2025 gives a rare side-by-side. Across its data centers, it withdrew 13,562 million gallons, discharged 3,039 million and consumed 10,523 million, about 78% of what it took. Google calculates consumption by subtracting discharge from withdrawal. By site, Council Bluffs, Iowa, withdrew 1,746 million gallons and consumed 1,346 million. Ashburn, Virginia, withdrew 76.5 million and consumed 53.2 million. The share returned runs from about 9% at its Berkeley County, South Carolina, site to about 30% at Ashburn. A reader who sees only “withdrew” or only “consumed” can be off by a large margin in either direction.

A third number is the contract. When Cardinal News wrote about Google's proposed site in Botetourt County, Virginia, the figure was 2 million gallons a day under contract, or 730 million gallons a year, with a possible future expansion that could require 8 million a day. A contract is a supply commitment. The column compares it with the annual withdrawals Google reports at its existing Virginia sites: 359.6 million gallons at Bristow, 258.4 million at Sterling, 238.1 million at Leesburg and 76.5 million at Ashburn. That is a fair comparison of withdrawal, but it says little about consumption.

Timing matters as much as volume. The UC Riverside preprint finds that water demand clusters on the hottest days, when a large evaporatively cooled data center can need more than 1 million gallons a day and some planned campuses up to 8 million, which strains local water systems even when annual totals are modest. It projects that by 2030 U.S. data centers could withdraw 80 to 150 billion gallons a year and consume 60 to 110 billion, about 4 to 7% of consumptive use in U.S. public water systems. Google's framing is that data centers' water consumption is small nationally, less than 1% of what Americans use on lawns each year, but that local watersheds can still feel it. Both can be true.

Five questions for the next headline

When a data center water number appears, five questions sort out most of the confusion. Is it withdrawal, consumption or a contract limit? Is it a yearly total or a peak day? Does it count only the building, or also the power plants behind it? What cooling design is it, and is “closed loop” only the inner stage? And what is the local water source and its condition? Google says it uses a water risk framework to match cooling to each site, and chose water cooling at The Dalles because local sources were not at high risk of depletion or scarcity.

What to watch

The policy debate is moving toward these distinctions. In Texas, state Sen. Charles Perry said at a Senate hearing that he expects evaporative cooling for data centers to be barred after the next session, and lawmakers plan to take up the issue in January; CBS Austin's report describes no enacted rule or bill text. Cook argued for efficiency standards over technology mandates, because a mandate could become outdated and regulators should weigh a facility's full water and energy footprint. That is the trade in this piece in policy form: a rule that cuts water on site can raise electricity demand.

On the company side, Google says its water projects replenished about 7.7 billion gallons in 2025, roughly 78% of its freshwater consumption, against a goal of 120%. Microsoft's chip-level pilots will show how its design performs outside the lab. None of the sources claims one design is right everywhere. What they describe is a choice that depends on the local climate, the local water source and the local grid, and a set of numbers that mean little until it is clear which one is being counted.

States in this story