Servers hum like a distant river, cool air pushed through grates, lights blinking in neat rows. It looks orderly. The impacts are not. A new Nature Sustainability study from Cornell University translates the boom in artificial intelligence into the everyday currency of carbon, water, and where to put the next building.
The team compiled industrial, power system, and climate data to project what today’s AI growth means by 2030. Their mid-range scenarios point to 24 to 44 million metric tons of CO2 emissions a year, roughly the same as adding 5 to 10 million cars to U.S. roads. Water is the other meter that spins fast: 731 to 1,125 million cubic meters annually, on par with the household use of 6 to 10 million Americans. In other words, the environmental toll is knowable, and large.
The authors are not doomsayers. They are system modelers, and their models also test solutions. Three levers dominate: siting, grid decarbonization, and operational efficiency. Choose locations with lower water stress and cleaner power, accelerate renewables where AI grows, and squeeze more useful work from each server while cooling with less water. In combination, the roadmap trims carbon by about 73 percent and water by about 86 percent compared with the worst case. That is not a footnote. It is the difference between AI as a climate accelerant and AI as a manageable load.
“Artificial intelligence is changing every sector of society, but its rapid growth comes with a real footprint in energy, water and carbon.”
Location emerges as the quiet protagonist. Many data clusters now rise in water scarce basins such as Nevada and Arizona, or in already congested hubs like northern Virginia. The analysis shows that shifting capacity toward the Midwest and windbelt states, especially Texas, Montana, Nebraska, and South Dakota, delivers the best combined carbon and water profile. That is partly about abundant wind and solar, and partly about avoiding evaporative losses tied to hydropower heavy grids in some coastal regions. It is also about easing pressure on local water systems that serve people first.
The Grid Can Help, But It Cannot Do Everything
Even if the grid keeps decarbonizing, demand matters. If AI load grows faster than power gets cleaner, emissions can still rise. In the study’s ambitious high renewables scenario, carbon falls only about 15 percent versus the baseline, leaving roughly 11 million tons of CO2 in 2030 that would still need to be neutralized. The paper translates that remainder into concrete builds: about 28 gigawatts of wind or 43 gigawatts of solar. That is a lot of steel in the ground, and it must be sited, permitted, and interconnected in the same places where compute demand is taking off.
Operations add another layer. Advanced liquid cooling and better server utilization shave a few more percentage points from both energy and water. Combine those with improved power usage effectiveness and water usage effectiveness and the gains compound. But efficiency is not infinite. The authors caution that they are already pushing the physical limits of what modern data centers can do inside the fence line.
Their modeling also treats the rebound effect with clear eyes. Faster chips and clever software can cut energy per computation, then unlock more applications that raise total demand. If that sounds familiar, it is. Efficiency is essential, but it rarely closes the loop by itself.
Siting, Transparency, And The Politics Of Scale
Practical challenges remain. Texas could be pivotal yet may need to support an additional 74 to 178 terawatt hours of AI demand by 2030, a scale that stresses transmission and interconnection queues already bulging with queued renewables. Northern Great Plains states would need upgrades in connectivity and security to host hyperscale AI, which carry their own costs and environmental footprints. None of this is a single company’s problem. It invites regional planning among utilities, regulators, and the firms racing to build compute.
Public trust will hinge on water. Communities will ask whether data centers take priority over farms and households when drought arrives. The paper’s answer is not rhetorical. It is quantitative, and it points to avoiding scarce basins, adopting water efficient cooling, and pairing new load with new, local clean power. Credible offsets and water restoration may still be required, but the authors emphasize that verification should be third party and transparent. In a sector that loves opacity, that is a cultural change as much as a technical one.
“New York state remains a low-carbon, climate-friendly option thanks to its clean electricity mix of nuclear, hydropower and growing renewables, although prioritizing water-efficient cooling and additional clean power is key.”
There is a memorable simplicity to the roadmap. Put AI where the grid can be clean and the water can be spared. Accelerate the clean energy transition in those same places. Run the machines more efficiently and cool them with less water. Then measure, report, and verify what remains. The decade’s build out is happening now. Whether AI lightens the climate load or adds to it will be decided by siting maps, interconnection queues, and cooling diagrams as much as by breakthroughs in the models themselves.
Nature Sustainability: 10.1038/s41893-025-01681-y
ScienceBlog.com has no paywalls, no sponsored content, and no agenda beyond getting the science right. Every story here is written to inform, not to impress an advertiser or push a point of view.
Good science journalism takes time — reading the papers, checking the claims, finding researchers who can put findings in context. We do that work because we think it matters.
If you find this site useful, consider supporting it with a donation. Even a few dollars a month helps keep the coverage independent and free for everyone.