TL;DR — Key Takeaways
- A Harvard University study estimates that 403 major U.S. data centers produce 25.7 million metric tons of carbon dioxide annually, comparable to the emissions of some countries.
- The facilities consumed 82 terawatt-hours of electricity in one year, with fossil fuels accounting for 54% of their power consumption.
- Researchers warn that limited transparency and rapid AI infrastructure expansion make it difficult to assess the industry’s full environmental impact.
A new Harvard University study reveals that 403 of the largest U.S. data centers generate approximately 25.7 million metric tons of carbon dioxide annually, a level comparable to the emissions of entire countries such as Denmark. The research raises concerns about the environmental costs of the massive AI data center expansion.
Published in Nature Sustainability, the research examined data center electricity consumption between May 2024 and April 2025. The facilities consumed 82 terawatt-hours of electricity, representing nearly 2% of total U.S. electricity use. That’s enough electricity to power approximately 7 million American homes for a year.
The study also found that 54% of the electricity consumed by these facilities was generated using fossil fuels, primarily natural gas and coal. The resulting carbon emissions represented approximately 0.5% of total U.S. carbon dioxide emissions.
The research indicates that the environmental impact of AI infrastructure is far greater than previously understood. The emissions estimate is two to three times higher than earlier research indicated, according to Harvard biostatistics professor Francesca Dominici, a senior author of the study. Furthermore, the research captures only a portion of the data center industry, which is expanding rapidly as hyperscalers invest billions in AI computing infrastructure.
Challenging Environmental Oversight
A big challenge facing researchers is the lack of publicly available information about data center operations. To calculate electricity consumption, Harvard researchers examined commercial property records, satellite imagery and building dimensions. They then estimated power requirements and assessed the energy sources supplying regional electrical grids.
This process was necessary because data center operators generally don’t disclose detailed information about their facilities’ electricity consumption or carbon emissions.
The study identified Virginia, Iowa and Ohio as major contributors to data center emissions, collectively accounting for approximately half of the total. Virginia alone represented more than one-quarter of estimated emissions.
Geography plays a key role in determining environmental impact. In Virginia’s PJM electricity market, which includes Virginia and several other states, fossil fuels generate approximately 60% of the power used by data centers. Facilities operating in regions with greater access to renewable or other carbon-free electricity produce fewer emissions for comparable power consumption.
The industry’s environmental footprint could be substantially larger than the study indicates. According to energy data provider Cleanview, at least 250 additional data centers began operating after the research period ended. Additionally, the study excluded facilities whose locations could not be independently verified and data centers powered by on-site generation.
The findings arrive as governments face pressure to address AI’s energy requirements. At the COP31 climate conference in Antalya, Turkey, organizers introduced the voluntary Antalya Artificial Intelligence Pledge, encouraging data center operators to adopt cleaner electricity, improve energy efficiency and disclose environmental information. The initiative also promotes more efficient water consumption and responsible AI development.
Beyond cleaner electricity, researchers are examining ways to reuse waste heat generated by computing equipment. The International Energy Agency estimates that data center heat recovery could potentially supply heating for approximately 10% of European homes by 2030 if fully integrated into district heating systems.

