Powering AI Could Soon Use as Much Water as 10 Million Americans

The rise of AI is changing the world in more ways than one. As this booming new industry revolutionizes the way we live and work, scientists are racing to quantify its environmental impact as well.

Projections for AI’s resource consumption and carbon emissions already paint a grim picture. A new study published Monday in the journal Nature Sustainability shows that AI server deployment across the U.S. could produce 26 million to 48 million tons (24 million to 44 million metric tons) of carbon dioxide equivalent per year between 2024 and 2030—equal to putting 5 to 10 million new cars on the road.

On top of that, the researchers expect AI to consume roughly 193 billion to 297 billion gallons (731 million to 1,125 million cubic meters) of water per year, as much as the annual household water usage of 6 to 10 million Americans.

“The magnitude is striking,†lead author Fengqi You, an endowed professor of energy systems engineering at Cornell University, told Gizmodo in an email. Though his findings underscore the significant environmental consequences of AI’s rapid growth, they also point to effective ways to make the industry more sustainable.

“Equally striking is that better facility location, clean-power procurement, and efficient cooling can cut those impacts by roughly 70–85%,†You added. “So the footprint is significant but also manageable with coordinated action.â€

Quantifying AI’s environmental footprint

As AI data centers proliferate across the U.S., communities are already feeling the impact. Skyrocketing electric bills, air pollution, and power grid strain are just some of the challenges Americans face when these facilities come to their hometowns, and the consequences are most severe for low-income areas and communities of color.

But the impacts aren’t just local. Through their analysis, You and his colleagues show how much the U.S. AI industry is contributing to global climate change and national water scarcity.

To quantify these larger impacts, they linked projected AI electricity demand to each state’s grid carbon and water characteristics, You explained. His team then compared baseline impact scenarios to mitigation scenarios that use strategies such as smart siting, grid decarbonization, and advanced cooling techniques to reduce data centers’ impact.

Charting a sustainable path forward

The findings point to data center siting as one of the most significant factors influencing AI’s environmental footprint.

States differ considerably in their renewable energy and water resources, local climate, power grid carbon intensity, and the concentration of existing data centers. As such, the same AI workload can have a two- to five-fold difference in emissions or water use depending on where it operates, You said.

Many data center clusters are currently being constructed in water-scarce states, such as California, Nevada, and Arizona, according to the study. And in states that have already become data center hubs—such as Virginia—new construction can compound existing strain on local infrastructure and resources.

You and his colleagues therefore advocate for siting new facilities in regions with low water stress and plentiful renewable energy resources. They highlight the Midwest and “wind belt†states—particularly Texas, Montana, Nebraska, and South Dakota—as ideal siting locations for reducing data centers’ carbon and water footprint. New York also stands out for its clean electricity mix of nuclear power, hydropower, and renewable energy.

That said, making a meaningful difference in AI’s environmental impact will require more efficient cooling technology and grid decarbonization as well, according to the study.

“Location, power source, and cooling technology together determine whether AI infrastructure becomes part of the sustainability solution,†You said. “AI may look virtual, but it runs on very real systems, such as grids, water, and cooling. The footprint depends on where we build and how we power and cool it.â€