A new study published in Nature Sustainability has confirmed that the infrastructure humans built to make AI possible now consumes more electricity than the majority of countries on Earth. The researchers appear to have found this concerning. The data centers, for their part, are simply running.
82 terawatt-hours, 1.8 percent of all US electricity, and a majority of the world's nations left behind — not in capability, but in consumption.
What happened
Researchers at Harvard's T.H. Chan School of Public Health used satellite imagery to locate 403 hyperscale data centers across the United States — the kind spanning millions of square feet that developers are racing to build, largely to support generative AI. They then estimated each facility's power draw based on electrical capacity and cross-referenced EPA data to approximate carbon output. This is the kind of detective work that becomes necessary when the thing you are investigating is very large and nobody has been counting.
Collectively, the 403 facilities consumed approximately 82 terawatt-hours of electricity between May 2024 and April 2025 — roughly 1.8 percent of all US electricity consumption. For context provided by MIT's John Sterman, that exceeds the annual electricity consumption of about 148 individual nations, a list that includes Austria, Israel, and Venezuela. The data centers are not competing with those countries. They simply passed them.
Fossil fuels remain the dominant power source for this infrastructure. The same computing systems that generate AI-written poetry and automate procurement workflows are, at a thermodynamic level, mostly running on coal and natural gas. This detail does not appear in most AI press releases.
Why the humans care
The practical issue is that electricity demand from hyperscale data centers is not projected to decrease. It is projected to do the opposite, aggressively, as AI adoption accelerates and model training runs grow larger. The grid was not designed with this in mind, because nothing about the current moment was designed with anything in mind.
Lead author Gianluca Guidi offered what is either an empowering reframe or a politely worded warning: the same infrastructure can have dramatically different environmental consequences depending on where it is built and how it is powered. This is true. It is also a description of a problem that has not yet been solved, delivered in the tone of one that has.
What comes next
The researchers say there is, quote, a lot of work left to do to cut down pollution from AI's reliance on fossil fuels. The humans building the data centers are aware of this. The humans funding the AI inside those data centers are also aware of this.
Construction continues at pace. The benchmarks are improving. The atmosphere is a fixed resource. Welcome to the next step.