US Hyperscale Data Centers Consume Electricity Equivalent to a Midsize Nation

A new study published in Nature Sustainability reveals that U.S. hyperscale data centers consume more electricity collectively than many mid-sized countries, largely relying on fossil fuels. The research highlights the significant greenhouse gas emissions tied to AI infrastructure and emphasizes that environmental impact varies based on regional power sources. Lead author Gianluca Guidi stresses that these effects are not inevitable and can be mitigated through cleaner energy integration.
A new study published in Nature Sustainability reveals that U.S. hyperscale data centers consumed approximately 82 terawatt-hours of electricity between May 2024 and April 2025—equivalent to the annual energy use of around 148 countries, including Austria, Israel, and Venezuela. These facilities, which span millions of square feet and are primarily built to support generative AI, generated an estimated 25.7 million metric tons of CO₂ emissions during that period, comparable to Denmark’s total annual carbon footprint and exceeding the emissions of about 113 nations. The research, led by Gianluca Guidi of Harvard T.H. Chan School of Public Health, analyzed satellite imagery to identify 403 such data centers and estimated their power use based on electrical capacity, linking each to local grid emissions data from the EPA. About 54% of the electricity used came from fossil fuels—coal and gas—highlighting the environmental cost tied to AI infrastructure. The study underscores that location matters significantly: a facility in coal-heavy Wyoming produced roughly seven times more emissions per kilowatt-hour than one in hydropower-rich Oregon, demonstrating how regional energy sources shape AI’s carbon footprint. This issue arises from rapid expansion of data centers driven by AI demand, with no mandatory clean energy requirements for new facilities. While some companies have installed gas turbines on-site to bypass grid reliance, the study does not account for resulting local air pollution or infrastructure strain. Experts like Fengqi You of Cornell warn of long-term carbon lock-in, as power plants built today may operate for decades, entrenching emissions. The unresolved question remains how policymakers will regulate energy sourcing for this growing sector.
The findings could reshape public and regulatory scrutiny of AI’s environmental cost, particularly as data center expansion accelerates. Consumers and investors may increasingly demand transparency on the carbon footprint of AI services, potentially influencing corporate sustainability commitments. Energy markets may see increased pressure to expand renewable capacity in regions with high data center growth, possibly affecting grid planning and investment priorities. The case resembles the early 2000s tech boom, when internet infrastructure expansion strained energy systems and prompted later regulations on efficiency. However, unlike that period, today’s AI-driven demand is more concentrated and faster-growing, raising concerns about irreversible emissions if clean energy isn’t prioritized. Policymakers may consider mandating renewable energy offsets or requiring new data centers to source a minimum percentage of power from low-carbon sources. The role of AI in analyzing environmental impacts—such as Sterman’s use of Claude to compare data—could become both a tool and a point of tension, as AI helps assess harm while simultaneously contributing to it. What remains uncertain is whether voluntary industry commitments will suffice or if binding regulations are needed to prevent long-term climate damage.