There's a power/water trade-off in data center resource allocation
Landon Marston, who studies sustainable water resources management, finds himself in the middle of a growing challenge when it comes to data centers.
Note to readers: This series of articles focuses on researchers whose work improves efficiency, addresses concerns, or offers alternative solutions to some of the pressing issues created by data centers.
There’s a lot of conflicting information floating around about just how much water data centers really use.
When it comes to the hyperscaled, artificial intelligence (AI)-driven data centers, there’s water for cooling the components, especially during intense model training. There’s some nominal water use for each query. But how that water — and, at least as crucially, power — is allocated is a lot more complicated than may appear on the surface.
For Associate Professor of civil and environmental engineering Landon Marston, who specializes in environmental and water resources engineering, it’s not the water use that so often gets highlighted that concerns him.
“Energy is a bigger deal in my mind than water,” he said. “Water gets more attention, probably because people understand that more. It’s obviously fundamental to life. You know what a gallon of water is — most people don’t know what a kilowatt-hour of energy is.”
The resource allocation equation is, essentially, a trade-off between power and water. But there are often complimentary costs within each trade-off. Server racks can be cooled through other technologies, such as air-side cooling, but that will, in turn, require more power. And the water use from generating that power can be a lot more than people realize.
“From a water perspective, that could potentially have an impact on local water resources because many forms of energy generation — particularly when we’re talking about natural gas, coal, and nuclear — will withdraw and consume a notable amount of water,” said Marston.
So while initiatives like the Ratepayer Protection Pledge require AI companies to be responsible for their own power, mega projects will also put multiple layers of strain on local water resources. In a 2021 study, Marston and his team estimated that about three-fourths of the total water footprint of data centers actually came from offsite power sources, and recent data suggests that number may now be closer to 90 percent.
“You’re solving one problem, but perhaps you’re creating new ones,” said Marston.
In water-abundant areas, this could actually be a benefit, serving to reduce some of the energy load. But in drought-stricken areas where many major data center projects are planned, like southern Arizona, the Colorado River Basin, and Texas, already strapped municipalities could face compounding competition for both water and power resources. There’s also the broader issue of water loss in places like Lake Powell, which could diminish the Hoover Dam’s ability to create hydroelectric power for the region.
“I think we’re going to see more recognition of this beyond pure academics because it’s starting to impact people’s lives when there’s simply not enough water available to meet direct water requirements, but energy demands as well,” said Marston.
In the other direction, it also takes a lot of energy to purify and move water from one place to another, with about 80 percent of municipal water processing and distribution costs spent on electricity. For major new projects, like the Google data center in Botetourt County, Virginia, these resource costs can be accounted for, so long as the public-private partnership helps support the cost of all this additional infrastructure capacity. Marston’s larger worry is if, due to broader forces, that partnership breaks down.
“What happens if we’re in an AI bubble?” he asked. “These data centers commit to purchasing all this water, investing in all this infrastructure, the utility moves forward with that, makes these really large investments that have these 30-year payback periods. Then, three years down the road, the data center industry dips, and many of these data centers they’re serving go under. Who’s on the hook for all those stranded assets?”
All articles in this series
Power supply or grid stress? Both need to be solved to meet data center demand
Reducing power consumption, protecting data through algorithms
There's a power/water trade-off in data center resource allocation
Thinking small: How small language models could lessen the AI energy burden
