Shaolei Ren has been studying data center water consumption since before most people knew what a GPU cluster looked like. The associate professor at UC Riverside's Bourns College of Engineering has published multiple papers tracking how much water AI infrastructure drinks. His latest findings, released in collaboration with Caltech, land harder than anything before: cooling America's growing fleet of data centers could require $10 billion to $58 billion in new water infrastructure. Not new servers. Not new power plants. Pipes, pumps, reservoirs, and treatment plants.
The paper, titled "Small Bottle, Big Pipe: Quantifying and Addressing the Impact of Data Centers on Public Water Systems," was co-authored by UCR doctoral student Yuelin Han, Rochester Institute of Technology assistant professor Pengfei Li, and Caltech professor Adam Wierman. Its central argument is simple and devastating. The industry has been measuring the wrong thing.
Data center operators love reporting annual water consumption. The numbers sound manageable when spread across 365 days. But Ren and his team focused on something different: peak daily demand, the actual volume a facility pulls from the local water system on the hottest days of summer, when servers run hardest and evaporative cooling towers are gulping at maximum capacity.
The gap between average and peak is staggering. Daily water demand from evaporative cooling can spike 6 to 10 times above average usage. For some planned facilities, that ratio exceeds 30. A large data center can withdraw more than 1 million gallons per day during heat waves. Some facilities under construction have been allocated up to 8 million gallons daily, enough to serve multiple small towns.
Without new efficiencies, the researchers project that U.S. data centers could collectively require 697 million to 1.45 billion gallons of additional peak water capacity per day by 2030. For context, New York City's average daily water supply runs about 1 billion gallons. The AI boom is building a thirst that rivals America's largest city.
Here is where the physics matter. Most large data centers use evaporative cooling systems because they are dramatically more energy efficient than conventional mechanical chillers. The principle is straightforward: hot outside air gets pulled through water-saturated media or spray systems, and the phase change from liquid water to vapor absorbs massive amounts of heat, dropping the air temperature toward the wet-bulb point. The cooled air then enters the server halls. In dry climates with wide wet-bulb depressions (the gap between actual air temperature and the theoretical limit of evaporative cooling), these systems perform brilliantly. A typical 100 MW IT load facility requires roughly 0.5 to 2.5 million gallons per day depending on local climate and cooling design. But here is the catch. Water consumption scales with outside temperature. On a 75-degree day in March, a facility might use a fraction of what it needs in August when ambient temps push past 100 degrees. Municipal water systems are sized to handle predictable demand curves. Data centers warp those curves violently.
Ren's team used a conservative peak-to-average daily water use ratio of just 4.5 in their analysis. The real numbers for many facilities run far higher. That means even their $10 billion floor estimate is likely optimistic.
You can see the future by looking at The Dalles, Oregon, a city of about 15,000 people along the Columbia River where Google has operated data centers since 2006. In 2012, Google consumed 12 percent of the city's water supply. By 2024, that figure hit 33 percent. One company, one third of a city's water.
Google has invested nearly $29 million into The Dalles' water infrastructure, building new groundwater wells, storage tanks, and pump stations. The company also transferred a $28 million aquifer storage and recovery system to the city that captures rainwater runoff and banks it underground for dry months. The city's 2024 water master plan estimates it needs $260 million in infrastructure improvements over 20 years, with Google's revenue expected to cover roughly a quarter of that.
That still leaves nearly $200 million on The Dalles' residents. For a city of 15,000.
The kicker is that American water infrastructure was already in crisis before data centers showed up demanding billions of gallons. The EPA estimates the national water infrastructure funding deficit exceeds $1.2 trillion over the next two decades. Over 2 trillion gallons of treated drinking water leak out of aging pipes every year. Some mains still in service predate the Civil War.
Now layer data center peak demand on top of that. The United States has roughly 50,000 community water systems, and about 40,000 of them serve populations of 3,300 or fewer. These small systems have minimal engineering staff, thin budgets, and no experience negotiating with hyperscalers. When a company like Meta announces a $10 billion campus in a rural Indiana county, the local water utility suddenly needs to plan for peak demand volumes it has never imagined. In February 2026 alone, three major tech companies announced water infrastructure commitments approaching $1 billion combined for projects in Virginia, Louisiana, and Indiana.
The paper's recommendations are pointed. First, data center developers should be required to report peak water use, not just annual averages. The annual number obscures the real strain on local systems. Second, data center companies should partner directly with communities to fund water infrastructure upgrades with verifiable outcomes, so that expansion costs do not fall entirely on local ratepayers. Third, operators should dynamically adjust cooling methods based on grid stress and water system capacity, throttling evaporative cooling during peak periods even if it costs more in electricity.
"People recognize power as a constraint for data center growth," Ren told UC Riverside News, "but most haven't realized water is a hidden and even more binding constraint."
He is right. And the constraint gets worse. "Even if you have money, the water source is another challenge," Ren said. "Money can't buy more snowpack."
The hyperscalers are betting they can build fast enough to capture the AI training and inference market while backfilling water infrastructure as they go. Google's Dalles investments suggest at least some awareness. Meta's $75 million earmark for public water projects alongside its Indiana campus shows the same. But these are reactive moves, company by company, town by town, with no coordinated national framework for how data center water demand gets allocated, priced, or capped.
My read: within 18 months, at least one major data center project in the American West will face a construction delay or permit denial specifically because the local water system cannot guarantee peak supply. Water will become a site selection constraint on par with power availability. The companies that figured out liquid cooling and closed-loop systems early will have a real competitive advantage over those still dependent on evaporative towers sucking from municipal mains.
Ren's paper puts a dollar figure on what communities have been feeling for years. The water bill for AI is coming due. And the people paying it are not the ones training the models.