Just 400 US Data Centers Use More Electricity Than Michigan
4 75An anonymous reader quotes a report from Wired: Just 400 hyperscale data centers connected in the US used as much electricity as the state of Michigan in a single year, new Harvard research has found. The study, published Monday in Nature Sustainability, also finds that the majority of that electricity was generated using fossil fuels, and data center emissions are three times higher than they were before the AI boom. Federal and state governments do not track data center use separate from the rest of the power sector, despite mounting projections that data centers are putting serious strain on the American power grid. Individual data centers' power use is often considered proprietary business information, and data center operators and utilities do not have to disclose specific numbers to the public.
The Harvard study specifically focuses on hyperscale data centers -- the largest types of facilities, which have become flashpoints in community opposition across the country. While often used for AI inference or training, not all of the hyperscalers surveyed in this study are being used for AI. [...] To assemble their dataset, the Harvard researchers put together a list of 403 hyperscale data centers around the country that were operational between May 2024 and April 2025 using a mix of industry research and satellite data. The team then modeled these facilities' energy use in a variety of scenarios and calculated the energy mix of the regional grids to estimate individual facilities' carbon emissions. Not all regions were created equal: Data centers in Virginia, Oregon, Ohio, and Iowa -- the first of which is home to Data Center Alley -- were responsible for more than 50 percent of the energy use in the study. The report notes that these findings may already be outdated. "Some of the massive campuses being proposed today are substantially larger than the typical facility in our dataset," says Gianluca Guidi, a postdoctoral research fellow at Harvard's T. H. Chan School of Public Health and the lead author of the study.
The study also excludes onsite power generation, a rapidly growing trend that BloombergNEF estimates could increase overall U.S. emissions by 10% if all proposed projects are built.
4 comments
Does 3X emissions come with 3X ROI? (Score: 5, Informative)
by aslashdotaccount ( 539214 ) on Monday November 16, 2026 @11:07PM (#66428232)
Can we honestly claim that the benefits of frontier AI (those massive multimodal mixture-of-expert systems) has improved our economies enough to justify the desecration of our planet?
My organisation uses agents tied to an internal data analytics MCP for all of our reporting, but I can easily replace those cloud agents with self-hosted Llamas and get an equally good result (at least for data analysis). The only reason why I am not allowed to take that route is because my C-suite is FOMOing on the hype created by the big AI vendors.
Re:Does 3X emissions come with 3X ROI? (Score: 5, Insightful)
by gtall ( 79522 ) on Tuesday November 17, 2026 @03:38AM (#66428354)
You misunderstand a major part of the AI scam. "AI has never been implemented properly or thoroughly enough, that's why it isn't contributing to the bottom line." Five years from now, it will be the same refrain, after they've lobotomized its users who will have lost any use of their brains. Use it or lose it.
I'm sorry. (Score: 5, Funny)
by msauve ( 701917 ) on Monday November 16, 2026 @11:12PM (#66428238)
The summary makes it sound bad. I'm in Michigan, I guess I'm not using enough electricity.
Michigan doesn’t run the economy. These do. (Score: 5, Insightful)
by cliffjumper222 ( 229876 ) on Monday November 16, 2026 @11:38PM (#66428260)
What gets lost in these stories is that data centers don’t burn electricity for its own sake. They run the websites, databases, payment systems, logistics, and yes, AI that nearly every US business depends on. This isn’t just Instagram and TikTok; it’s payroll, inventory, banking, and healthcare records.
If that work weren’t done in hyperscale facilities, it would be done somewhere else: in smaller colos or on-prem server rooms, which are typically far less efficient per unit of work. Hyperscalers run at a PUE close to 1.1, while the average corporate server closet is much worse.
So the right question isn’t “how much power do data centers use?” It’s “how much power does this economic activity use, and is this the most efficient place to do it?” The study’s real findings are about the grid mix (mostly fossil) and the lack of disclosure. Those are fair criticisms, and both are fixable. But a Michigan-sized power bill for a large chunk of the digital economy isn’t obviously a scandal.