400 US data centers use as much power as Michigan
Most conversations about AI focus on the models. The electricity bill behind them is just as striking. A study by Harvard researchers published in Nature Sustainability found that 403 hyperscale data centers in the US consumed as much electricity in a single year as the state of Michigan, according to the report. Most of that power came from fossil fuels, and data center emissions are now three times higher than before the AI boom.
There is a reason the numbers stay so unclear: data center electricity use is not tracked separately from the rest of the power sector. Figures for individual facilities are often treated as proprietary business information, and neither operators nor utilities are required to disclose them publicly. To fill that gap, the team built a list of facilities operating between May 2024 and April 2025 using industry research and satellite data, then modeled their energy use and estimated emissions from the fuel mix of regional grids.
The study also shows that consumption is unevenly spread. Data centers in Virginia, Oregon, Ohio and Iowa accounted for more than half of the energy use examined, the article says, with Virginia home to the cluster known as Data Center Alley. Not every hyperscale facility in the dataset is used for AI training or inference, the researchers note.
The findings may date quickly. The average facility in the study is 40 megawatts, while campuses of hundreds or even thousands of megawatts are now being proposed. The research also leaves out on-site, behind-the-meter power plants, a factor one outside researcher calls a wild card in judging total data center demand.