In recent years, there’s been increasing concern about how the U.S. electrical grid is going to provide the power required by the skyrocketing number of data centers being constructed to train and run artificial-intelligence models.
A new academic paper suggests policymakers need to be devoting similar concern to how those data centers are going to affect the country’s water systems.
The computers inside data centers generate ample amounts of heat. To keep temperatures within their operating ranges, the data centers typically rely on cooling systems that incorporate water-evaporation towers. Those systems can consume millions of gallons of water per day, particularly during peak periods, a group of researchers from UC Riverside, the California Institute of Technology and the Rochester Institute of Technology point out in their new paper, which they released in draft form prior to peer review.
As has already been seen with particular data-center projects around the country, that kind of demand can easily exceed available water supply in the areas where the computing facilities are being proposed or built, necessitating the construction of new water infrastructure, operational delays or a switch to other cooling methods that use less water but require more electricity, the researchers note.
“Many data centers want to use water” for cooling, said Shaolei Ren, an associate professor of electrical and computer engineering at UC Riverside who is one of the authors of the study. “The challenge right now, today, is many communities are not able to support [the facilities’] water capacity needs.”
How data-center operators address the water-capacity questions could be crucial for the artificial-intelligence industry, which is concentrated in San Francisco and the wider Bay Area.
The sky-high valuations and the tens of billions of dollars raised by OpenAI and Anthropic — both of which are based in The City — are predicated on them continuing to grow their numbers of users and develop their models, both of which depend on them building or having access to a growing number of data centers to train and run their AI models.
Meanwhile, Meta and Alphabet, both based on the Peninsula, are investing hundreds of billions of dollars in data centers to power their own AI models, which they see as crucial to the future of their businesses.
“Conceptually, this [study] is very interesting, because this is definitely a very novel angle” of research, said Alex de Vries-Gao, a Ph.D. candidate at Vrije Universiteit Amsterdam who himself has looked closely at the environmental impacts of AI.
Previous research has focused on the annual water use of data centers, typically in aggregate across the nation. But there’s no nationwide reservoir that water systems around the country draw on, Ren noted. Unlike with electricity, there’s generally no way to quickly transfer water from one place to another across long distances in times of need, he said.
And what’s really important for the local providers is not how much water is used annually, but peak demand and their own available capacities, Ren said.
Ex // Top Stories
At their peak usage, the gigawatt-size data centers being built to run AI workloads can require anywhere from 500,000 gallons of water per day to more than 5 million gallons per day in hot climates, according to the study. Numbers that get into the millions of gallons a day can easily exceed the available capacity of most water systems, according to the study.
That’s because out of the approximately 50,000 water systems in the United States, about 40,000 are small, serving no more than 3,300 customers, according to the report, which cites data from the Environmental Protection Agency. Another 9,000 serve fewer than 100,000 customers. The smaller systems typically have peak capacities of fewer than 2 million gallons, according to the EPA.
The New York Times reported that as of last summer, nine companies had applied to build data centers in Newton County, Ga. One of them was seeking to use up to 6 million gallons of water — more than the entire county used daily, according to the report.
Because local water systems are typically small, they are frequently under-resourced and can’t easily bear the cost of upgrading their infrastructures, according to the study. Even before the strain expected to be put on the system by data centers, those systems were expected to need at least $1.3 trillion worth of upgrades over the next 20 years, according to the EPA.
To accommodate their needs, some data-center developers are themselves investing in new water infrastructure, the study notes. But that can be costly — three recent data-center projects are spending a combined total $1 billion to upgrade the water systems in their areas, the report said.
Building new water infrastructure can also result in significant delays, the report noted. Meta’s new data center in Lebanon, Ind., might not be able to operate at full capacity until 2031 because the water-system upgrades needed for it won’t be ready until then, according to the report.
And such upgrades might not be easily accomplished in other areas. Some parts of the country are facing water stress and might not be able to provide additional water. Others face challenges with water or land rights to expand their existing capacities, Ren said.
As one example, he pointed to The Dalles, Ore., where Google is looking to expand its data centers — and water use. The water reservoir for The Dalles is in a national forest, he said. Expanding it would require federal review or a transfer of the land to the city. Discussion of doing that is already sparking concerns among the public, he said.
“This is a very controversial issue,” Ren said.
Other data-center developers are turning to cooling systems that don’t involve water evaporation. But those systems typically require significantly more electricity — which, ironically, can lead to even more water use at the power plants that provide that energy.
Indeed, one of the recommendations Ren and his colleagues make is that communities and data-center developers should be considering both the power and water needs of such facilities in tandem. Data centers can conserve electricity by using water for cooling, but that requires sufficient water supplies, he said. Alternatively, they can save water, but that requires more electricity.
“There are two bottlenecks” to data-center development, he said.
If you have a tip about tech, startups or the venture industry, contact Troy Wolverton at twolverton@sfexaminer.com or via text or Signal at (415) 515-5594.