US Hyperscale Data Centers Are Still Running on Fossil Fuels — AI's Electricity Footprint Is Far Harder to Calculate Than Commonly Assumed
A new study published in *Nature Sustainability* shows fossil fuels remain the top power source for US hyperscale data centers, whose total power consumption may outstrip that of many individual countries. When addressing AI's emissions footprint, there is no way to avoid accounting for what the local power grid actually burns to generate electricity.

The Verge illustration: Data centers and data visualization. Image credits: Cath Virginia / The Verge, Getty Images
Fossil fuels remain the dominant power source for electricity powering US hyperscale data centers. The Verge recently reported on a study published in Nature Sustainability, which focuses on hyperscale facilities — sprawling millions-of-square-foot large data centers that are being constructed at massive scale to support workloads including generative AI. The research finds that collectively, these facilities could consume more electricity than many entire countries, and produce significant greenhouse gas emissions. The paper draws a clear conclusion: there is still a huge amount of work left to do to reduce pollution caused by AI's reliance on fossil fuels.
The core focus of this study is worth a close look. It tracks actual real-world power sources, rather than taking the green marketing promises posted on corporate official pages at face value. Data centers draw power from the local grid, and the respective share of coal, natural gas and renewable energy in that local grid mix makes up the vast majority of their total emissions. While corporate renewable energy procurement contracts do drive new clean power capacity additions, there are still time and location mismatches between signed contracts and real-time power consumption. The electricity in the grid will not turn clean instantly just because a procurement contract is signed.
For AI practitioners, this issue is far from a problem that only the infrastructure team needs to handle. Growth in model training, inference, and API call volumes will all eventually put pressure on power capacity and push up associated emissions. Hyperscale facilities are being built at a very fast pace, but grid access timelines, local power generation mix, and site permitting rules all affect the speed of AI expansion. For regular users, there is a very concrete electricity bill behind every AI product, and the power source structure on that bill is far more worthy of attention than the product's user-facing interface.
The Verge's report does not cite a specific exact emission figure at the opening, but emphasizes that the total emission volume is considerable. The study sends a clear reminder to the industry: AI's environmental sustainability narrative cannot only stay at the level of green power procurement and energy efficiency metrics. What is actually burned in the power system is the far more unavoidable, harder to address core issue.
Paper: Nature Sustainability
Original Article: The Verge
发布时间: 2026-11-17 04:27