America’s biggest data centers used about 82 terawatt-hours of electricity in a single year, and more than half of it likely came from gas and coal. That’s the main finding of a new study in the journal Nature Sustainability, The Verge AI reports. That’s roughly 1.8 percent of all US electricity, which is more than about 148 countries each use in a year.
The study looks at hyperscale facilities. These are the huge campuses spanning millions of square feet that developers are racing to build, largely to run generative AI.
📊 The numbers
Here’s what the researchers estimated for May 2024 through April 2025:
- Facilities tracked: 403 hyperscale data centers in the US
- Electricity used: about 82 TWh, or 1.8% of US consumption
- CO2 emissions: 25.7 million metric tons, about the same as Denmark’s yearly footprint
- Share from fossil fuels: about 54% from gas and coal
- Countries that use less electricity: about 148, including Austria, Israel and Venezuela
- Countries with lower total emissions: about 113
The country comparisons come from John Sterman, director of the MIT System Dynamics group. He wasn’t involved in the study. He said he used Claude to help match the paper’s findings against national electricity and emissions data.
🛰️ How the researchers did it
Data center operators don’t publish their energy use, so the team had to estimate it. Here’s how:
- They used satellite imagery to find 403 hyperscale facilities across the US.
- They estimated each site’s power use from its electrical capacity.
- They used Environmental Protection Agency data to work out how much carbon each site would produce on its local power grid.
The interesting part is what that method shows about location. Grid electricity in coal-heavy Wyoming produced about seven times more carbon per kilowatt-hour than in Oregon, which runs largely on hydropower.
“Two data centers consuming exactly the same amount of electricity can have dramatically different carbon footprints simply because of where they’re located,” lead author Gianluca Guidi wrote. He’s a postdoctoral fellow at Harvard T.H. Chan School of Public Health.
💡 Why this matters
To me, Guidi’s framing is the most useful takeaway. People tend to talk about AI’s environmental cost as if it’s built into the technology. The study suggests it depends mostly on the grid behind the servers.
“I actually find that encouraging, because it means these impacts aren’t inevitable,” Guidi told The Verge. “We have choices about where we build and how we power this infrastructure.”
The catch is timing. Fengqi You, an energy systems engineering professor at Cornell, warns about “long-term carbon lock-in.” Gas plants built today to meet data center demand could keep running for decades. “Decisions made today could therefore shape emissions well into the future,” You said.
Sterman wants policymakers to make data center operators bring new clean energy online. “Once you allow these data centers to emit the CO2, you can never undo that,” he said.
🛠️ What you can do with this
If you buy cloud compute or run AI workloads, here’s how to use these findings:
- Ask your cloud provider which region you’re in. The same job can have a very different carbon footprint depending on where it runs.
- Factor grid mix into vendor choices. If you report on sustainability, region selection is one of the few levers you control directly.
- Expect more disclosure rules. A bipartisan bill in Congress could require data centers to report their electricity use nationwide.
⚠️ Limitations
The researchers and outside experts flagged several gaps:
- The figures are estimates based on electrical capacity, not measured use. Operators still don’t disclose that data.
- The study leaves out local air pollution from new gas infrastructure and from the on-site gas turbines some AI companies use to run off the grid.
- Hyperscale centers run more than just AI work, and they’re only a small share of the thousands of US data centers. New ones open every year.
In other words, the real total is probably higher.
🔭 What comes next
Guidi’s team is building a website to track the environmental impact of individual data centers. If it launches with good coverage, it could become the independent data source that operators haven’t provided themselves. With AI buildouts speeding up, where these facilities go and what powers them will matter a lot. You can find the full story at The Verge AI.