AI is incredibly thirsty. The data centers that run these models already use massive amounts of water, and by 2030, those in the U.S. could require enough additional water capacity to rival New York City’s daily supply.
That’s according to a new study led by Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside. The findings—which have not yet been peer reviewed but are publicly available on the preprint server arXiv—show that limited public water capacity is emerging as a critical bottleneck to data center growth.
To avoid burdening local ratepayers, tech companies are partnering with communities to fund water infrastructure upgrades, often spending hundreds of millions of dollars. “Those companies are profit driven, right? So I think clearly there is something wrong,” Ren told Gizmodo.
Why so thirsty?
Data centers operate continuously, generating lots of heat from dense concentrations of servers, networking equipment, and other forms of IT infrastructure. Liquid cooling techniques are the most efficient way to prevent overheating and system failure, but they tend to be highly water intensive.
Tech companies will often argue that by using “closed-loop” cooling systems, their data centers recycle most of the water they use and minimize consumption. But even these systems can consume huge amounts of water because many rely on evaporative cooling towers to transfer heat outside of the facility.
For example, peak daily water demand—the amount required during the hottest days of the year—for a large state-of-the-art data center using evaporative cooling can often exceed 1 million gallons per day, and for some planned facilities it may reach 8 million gallons per day, according to the study.
The water bottleneck
Public water systems are engineered to reliably meet maximum demand at all times, so a data center’s peak water usage is a critical factor in infrastructure planning, system resilience, and operational reliability. Despite this, most operators only disclose their total annual water use. To assess the peak water demand of U.S. data centers, Ren and his colleagues analyzed a wealth of data from public sources, including government records and water utility databases.
This revealed that if the current water use intensity persists, U.S. data centers will require between 697 million and 1.45 billion gallons per day of new peak water capacity by 2030. That’s comparable to the typical daily water supply of New York City. Building this additional capacity could cost between $10 billion and $58 billion, with much of the financial burden falling on the communities hosting data centers.
And that’s a “very conservative” estimate, Ren said. His team’s calculations assume a peak-to-average daily water use ratio of just 4.5, which is at the low end of the spectrum.
This presents numerous problems for the tech sector. Insufficient water capacity could directly impact the feasibility and efficiency of data center projects, leading to increased costs, delays, and scalebacks. It could also lead to operational inefficiencies, as data centers often must switch to dry cooling—using air instead of water—when water becomes unavailable. This is far less efficient and increases electricity demand, further straining the grid during summer peaks.
Ren and his colleagues do have some ideas about how to address the growing water capacity demand of U.S. data centers. Firstly, they emphasize the importance of requiring data centers to report their peak demand, not just total annual usage. They also recommend developing corporate-community partnerships to fund infrastructure upgrades so that residents don’t shoulder all of the burden.
“I don’t see any ways for them to afford this type of upgrade,” Ren said. “We need corporate funding and support.”
As data centers continue to proliferate across the country, the tech sector will be forced to contend with this often overlooked bottleneck. If nothing changes, these companies will face the consequences alongside the communities they’re impacting.
