When most people think about artificial intelligence, they picture chatbots, voice assistants or startlingly realistic fake photos. Fengqi You, a systems engineer at Cornell University, sees something else: a vast, physical machine spreading across the United States, pulling electricity from the grid, sucking up water for cooling and quietly rewriting the country’s climate math. His team’s new “roadmap” study, published in Nature Sustainability, is one of the first to put hard numbers on that footprint—and the results are a blaring warning siren, not just for tech companies, but for climate policymakers.

The Cornell researchers modeled how fast AI-heavy data centers are likely to multiply through 2030 and asked how much extra energy, water and carbon pollution they would add under different growth scenarios. Their answer is sobering. If AI keeps booming, U.S. data centers alone could end up consuming enough water each year to meet the needs of roughly 10 million people and generating as much climate pollution annually as 10 million gasoline cars. In total, the new AI infrastructure could demand as much electricity, water and emissions space as an entire large U.S. state such as New York. That is the scale of what is now on the line.

Crucially, the damage is not evenly spread. The study breaks the country into a patchwork of risk, showing that the same AI workload can have a two- to fivefold difference in emissions or water use depending on where it runs. In some states, grids are still dominated by coal and gas, and rivers and aquifers are already stressed by drought. In others, there is room to expand wind, solar and transmission lines, and water is relatively abundant. AI is not an abstract cloud; it is rooted in specific places with specific limits.

Right now, the industry is pouring concrete in some of the worst possible locations. For years, the densest cluster of data centers on Earth has been Northern Virginia’s so-called “data center alley” outside Washington, D.C., which grew up around tax breaks, fiber-optic cables and proximity to federal agencies. That hub is set to grow even faster as AI companies race to build new “hyperscale” facilities. But You warns that Northern Virginia is already hitting a wall: the region does not have enough clean power or water to host wave after wave of new AI complexes without pushing the local environment—and the grid—to breaking point. Yet unless policy intervenes, the Cornell team projects that it will remain the dominant destination for new builds through at least 2030.

The study’s most hopeful message is that it does not have to be this way. Because many of the biggest AI campuses have not yet broken ground, there is still time to steer them toward better sites. The team highlights Midwestern and Plains states with strong wind and solar resources, more stable water supplies and lower overall stress on local ecosystems. States such as Texas, Montana and Nebraska, for example, combine abundant land with high renewable-energy potential and, in many regions, more robust water availability. Redirecting AI projects there could cut the sector’s emissions and water use dramatically without requiring a slowdown in digital innovation.

Other analysts have arrived at similar conclusions. A recent Wired article based on the same Cornell research describes how smart siting could slash the climate cost of AI by aligning new server farms with clean-energy build-out instead of fossil-fuel pipelines. The International Energy Agency, meanwhile, estimates that global electricity demand from data centers will more than double to around 945 terawatt-hours by 2030, with AI responsible for most of that surge. In the United States alone, data centers already account for roughly 4 percent of electricity use, and their demand is expected to more than double again by the end of the decade. Without strong guardrails, those new terawatt-hours will overwhelmingly come from gas- and coal-fired plants.

That is where the second alarm bell rings. A report released last month by the Center for Biological Diversity warns that a massive, largely unregulated data center construction boom—powered mostly by fracked gas—could blow up America’s remaining carbon budget for the power sector. If current trends continue, the report estimates, U.S. data centers could eat up nearly half of the power-sector emissions allowed under the country’s 2035 climate target, and around a tenth of the total emissions the entire economy can emit while still meeting that goal. To stay on track, all other electricity-using sectors would have to slash their emissions an extra 60 percent beyond what was already planned, just to make room for AI’s appetite.

For Jean Su, one of the report’s authors, that trade-off is unacceptable. In comments to Inside Climate News, she argues that “unfettered data center growth is not an inevitability” and pushes back against techno-optimists promising that AI will magically solve the climate crisis. The science is clear on what actually reduces emissions: phasing out fossil fuels, building renewables and using less energy overall. AI may help in some specific tasks, such as optimizing grid operations, but Su warns that it is being used as cover for a dangerous expansion of gas plants and pipelines.

Those political choices are already playing out. In Inside Climate News’ coverage of a second Trump administration, the outlet describes how President Donald Trump has moved quickly to gut renewable-energy programs, fast-track new drilling and publicly champion coal and gas. In that scenario, utilities see AI data centers not as a reason to accelerate the clean-energy transition, but as a justification to lock in decades of new gas infrastructure. Projections cited by Su suggest that, under current policies, data centers will be overwhelmingly powered by fracked gas through at least 2035—directly undermining U.S. attempts to hit net-zero targets.

At the same time, communities living near proposed AI mega-campuses are starting to push back. In towns from Tennessee to upstate New York, residents have packed meetings to demand basic information: How much water will this facility withdraw from local rivers and aquifers? Will it raise their electricity bills? What happens when the next heatwave hits and the grid is strained? Many are frustrated to discover that companies and regulators treat key details as trade secrets, even though the consequences—higher pollution, stressed water supplies, more frequent outages—are very public.

You’s study argues that this lack of transparency is itself a barrier to good policy. The researchers call for AI-specific benchmarks for energy and water use, similar to fuel-efficiency standards for vehicles, alongside stricter reporting requirements. Without those rules, neither regulators nor citizens can tell whether a new project is genuinely efficient and aligned with climate goals, or simply another fossil-fuel-hungry server warehouse wrapped in green marketing. The International Energy Agency has made similar recommendations, urging governments to demand better data from tech firms and to tie new data center approvals to clear decarbonisation plans.

There is also a deeper question about who AI is really for. Supporters often frame it as a universal public good, promising breakthroughs on everything from climate modeling to cancer treatment. But the infrastructure is being planned largely by a handful of billionaire-led corporations and utilities whose first obligation is to their shareholders, not to vulnerable communities facing rising bills, droughts and heatwaves. As Su puts it, the risk is that “billionaire corporations dictate policy while the rest of us pay the price”—in the form of polluted air, dried-up rivers and blown climate targets.

Despite the grim numbers, neither the Cornell team nor the Center for Biological Diversity believe catastrophe is inevitable. The Cornell paper stresses that the Biden administration’s original 2035 climate goals—cleaning up the power sector and cutting economy-wide emissions—remain technically achievable if AI infrastructure is planned properly. That would mean rapidly decarbonizing the grid, prioritizing the most efficient cooling and computing technologies, steering data centers toward regions with abundant renewables and water, and enforcing strict limits on gas-fired expansion. It would also mean giving communities a real say in whether and how these projects are built.

The timeline, however, is brutally tight. Data centers are long-lived, capital-intensive assets. Every siting decision today locks in a pattern of energy and water use for decades. If AI campuses are wired into fossil-fuel systems now, they will still be emitting in the 2040s and 2050s, long after scientists say rich countries must have reached net-zero. Conversely, if governments force tech giants to pair their growth with genuine renewable build-out, grid upgrades and strong efficiency rules, AI could be one of the first major industries designed from the ground up with climate limits in mind.

When Fengqi You says “it’s not too late,” he is not offering comfort but a challenge. The AI boom is still in its early stages, which means there is a brief window to decide whether it becomes a climate accelerant or part of the solution. That choice will be made not in glossy AI demos, but in utility commission hearings, zoning boards and federal rulemakings over the next few years. The numbers in the new studies are a warning: without urgent action, the invisible servers behind our AI tools could end up doing far more damage to the climate than they ever repair. The question now is whether policymakers are willing to confront that reality—and act fast enough to keep this new wave of technology from crashing the planet’s remaining climate budget.