The Water Crisis Big Tech Refuses To Bill For

A data center in Oregon drank 355 million gallons of water in a single year. The company that owns it spent thirteen months in court trying to keep that number secret. The number is now public because a newspaper refused to drop the case. Every part of that sentence should bother you.

The town is The Dalles, Oregon. Population around 16,000. The company is Google. In 2021, Google's data centers in The Dalles consumed 355.1 million gallons of water, which worked out to 29 percent of the entire city's water use. Not 29 percent of industrial use. Twenty-nine percent of everything. Every home, every business, every school, every fire hydrant in The Dalles, on one side. A single company's server cooling on the other. Close to a third of the water went to the servers.

The number tripled in four years. In 2017 the same facilities used 124.2 million gallons. By 2021 it was 355 million. Google opened a third data center on the campus and increased the density of the others. Two more are planned. The Dalles sits in a meteorologically dry region in the middle of a multiyear drought. A local resident named Dawn Rasmussen, whose well water level dropped year after year, told The Oregonian the company had become a water vampire.

The public only knows any of this because of a lawsuit. When The Oregonian filed a public records request to see Google's water deal with the city, the city of The Dalles sued its own newspaper to block the disclosure. The legal fight ran thirteen months. Google funded it, providing more than $100,000 toward the city's legal costs. The city dropped the case in 2022. The water numbers came out. The pattern in that story, a company spending money to keep a public resource number hidden, is the entire subject of this entry.

What Cooling A Server Actually Costs

Here is the physical reality nobody puts on a marketing slide. A large data center can consume up to 5 million gallons of water per day. A medium-sized one consumes around 110 million gallons per year, which the Environmental and Energy Study Institute notes is equivalent to the annual water use of about 1,000 households. The cooling method that uses the most water is evaporative cooling, where water passes through a tower, absorbs heat, and evaporates away. That water does not go back into the system. It is gone. And in most facilities it is not greywater or reclaimed water. It is potable water. Drinking water. The same water that comes out of a kitchen tap, run through a cooling tower and released as vapor.

The national numbers are large and they are climbing fast. According to the EPA and the 2024 Lawrence Berkeley National Laboratory report, US data centers directly consumed 17.4 billion gallons of water in 2023. Direct consumption is forecast to reach between 38 and 73 billion gallons by 2028. That is only the water consumed on site. The same federal analysis estimated that data centers consumed roughly 211 billion gallons indirectly in 2023 through the power plants that generate their electricity, because fossil and nuclear generation boils water to make steam. Roughly 80 percent of a data center's true water footprint is hidden inside the power bill.

Texas alone shows the trajectory. A study by the Houston Advanced Research Center estimated Texas data centers would consume about 49 billion gallons in 2025 and as much as 399 billion gallons by 2030. The high end of that range is the equivalent of drawing down Lake Mead, the largest reservoir in the United States, by more than sixteen feet in a single year. The Ceres research group projected that water use tied to data center cooling will rise 870 percent as more facilities come online, and that data center growth could increase water stress in already strained basins by up to 17 percent annually.

Now the part that should make every reader sit up. In May 2025, the Fayette County water system in Georgia discovered that a QTS data center had consumed more than 29 million gallons of water without being billed for it. Twenty-nine million gallons. The equivalent of 44 Olympic swimming pools. The county found the gap only because it was upgrading its 33,000 residential meters to smart meters and noticed one data center meter still sitting on the old system, untracked. Residents had been complaining about low water pressure. They were unknowingly the ones who tipped off the regulators. The data center was drinking the town's water and the bill was not even being sent.

The Pattern Repeats In Every Drought State

Memphis sits on the Memphis Sand Aquifer, one of the cleanest natural drinking water sources in the country. Elon Musk's xAI built its Colossus supercomputer there. Engineers estimated Colossus 1 alone needs more than one million gallons of water per day for cooling. xAI promised to build an $80 million wastewater recycling plant so it could cool the servers with treated greywater instead of aquifer drinking water. The company broke ground on the plant in October 2025. In April 2026, local media reported the project had been quietly put on hold. Musk said on X that the company needed to finish building Colossus 2 first. The recycling plant, the one designed to protect billions of gallons of aquifer water a year, was the part that could wait. The Memphis mayor, the local utility CEO, a US congressman, and a community group called Protect Our Aquifer all had to publicly pressure the company to honor a commitment it had already made a ceremony out of.

Tucson said no before the pattern could start. In August 2025, the Tucson City Council voted 7 to 0 to reject Project Blue, a 290-acre data center development tied to Amazon. The first two phases of the project would have required almost 2,000 acre-feet of water every year, which would have made the data center the single largest water customer in the entire Tucson Water system, in a desert city, during a drought. The developer promised the project would be net water positive. Tucson residents packed three public meetings, up to a thousand people at a time, and the council listened. The developer found a path around the rejection anyway, switching to air cooling and getting power approved through the state corporation commission. Then in April 2026, the city discovered the project's contractor had drawn two acre-feet of city water through a construction meter the council had explicitly refused to authorize. The city revoked the meter and demanded the water be paid back.

Three states. Three versions of the same story. A company needs water. The community is not asked in any way that counts. The water gets used. And in case after case, it is the residents, noticing low pressure or dropping wells or a contractor's hose, who find out what happened and force it into the open.

The Machine That Does Not Need The Aquifer

There is a version of computing that does not drink the town's drinking water. It is the one already sitting in your house.

A phone does not have a cooling tower. A laptop does not evaporate the municipal supply. A gaming PC is cooled by the air in the room and the air conditioning that the household already runs and already pays for. The thermal load of consumer-device computing is handled by ambient air and by hardware the owner bought for other reasons. The incremental potable water cost of running an inference workload on a distributed network of consumer devices is, for practical purposes, zero. There is no cooling tower because there is no concentrated heat. The heat is spread across billions of devices in billions of rooms, each one shedding a few watts into air that was going to be cooled anyway.

This is not a small detail. It is a structural difference between two ways of building the same capability. Centralized AI compute concentrates heat into a building, and concentrated heat demands concentrated cooling, and concentrated cooling in most of the United States means evaporating millions of gallons of drinking water per facility per year. Distributed compute never concentrates the heat in the first place. The physics that makes a data center thirsty is the physics of concentration. Remove the concentration and the thirst goes with it.

The honest caveat belongs here. Distributed compute is not free of environmental cost. It still uses electricity, and electricity still carries an indirect water footprint at the power plant. The point is narrower and it is correct. The single most visible, most local, most community-destroying cost of the AI buildout, the draining of a specific town's specific water supply, is a cost that distributed architecture does not impose. The Dalles lost 355 million gallons because the computing was concentrated into a campus on the Columbia River. Spread that same computing across the devices already in the town's homes and there is no campus, no cooling tower, no 29 percent, no lawsuit against the newspaper.

The water going into these data centers is drinking water. It is being used in drought regions. It is being used faster every year. In at least one county it was being used without a bill. The companies doing it have, in multiple documented cases, fought in court to keep the public from learning the size of the number. The Oregonian had to win a thirteen-month lawsuit to publish a figure that should have been public the day the water meter was installed.

The water crisis is not coming. It arrived. It has a meter reading. The only open question is whether the next decade of computing keeps getting built in a way that drains a town, or in a way that never touches the aquifer at all.