A new analysis by researchers from Georgia Tech and the University of Waterloo has quantified how artificial intelligence adoption is shaping the country’s energy use and carbon output.
The study, published in Environmental Research Letters by IOP Publishing, estimates that widespread AI integration across U.S. industries could add roughly 896,000 tons of carbon dioxide emissions each year. The rise represents about 0.02 percent of national emissions, small in proportion but large enough to draw attention to AI’s energy implications.
The authors, Anthony R. Harding and Juan Moreno-Cruz, modeled AI’s impact across 55 industries, using data on occupational exposure, productivity, and energy intensity. They estimated that AI-related productivity gains translate into an additional 28 petajoules of energy use annually, equal to about 7.8 terawatt hours. In practical terms, that is roughly the annual electricity demand of 300,000 U.S. homes. The research captures not only the direct energy cost of training and running AI systems but also the indirect energy demand created when industries grow more productive.
The analysis shows that energy and emissions impacts vary sharply by sector. Industries with heavy infrastructure or high energy dependence (such as education, healthcare, construction, and retail) see the largest rises. The education sector alone could account for around 12 petajoules of new energy demand and more than 51,000 tons of added CO₂ each year. Publishing activities, by contrast, show only marginal changes, with about 0.003 petajoules of added energy use and less than one-tenth of a kiloton of CO₂. These differences stem from how energy-intensive each industry is and the mix of fuels that power its operations.
At the national level, the study’s estimate of 28 petajoules aligns closely with hardware-based projections from other sources. For instance, analysts have suggested that NVIDIA’s AI servers could consume between 5.7 and 8.9 terawatt hours of electricity a year if operated at full capacity. The new model’s estimate falls within that range, pointing to consistency between top-down economic modeling and bottom-up hardware estimates.
The researchers note that even though AI’s added footprint appears modest, it signals a trend that could accelerate as adoption spreads. As of early 2024, about 5.4 percent of U.S. firms were already using AI tools, a number that had risen by more than one percentage point in just six months. If that rate continues, AI’s share in the broader economy could double within a few years, expanding both productivity and energy demand.
The paper also acknowledges several uncertainties. It assumes fixed energy and emissions intensities for each industry and does not account for future gains in efficiency or changes in the energy mix. Because much of the underlying data dates from 2014, the real-world numbers today might be slightly lower thanks to cleaner power generation and more efficient computing. When the authors recalculated using projected 2023 data, the total impact dropped to 24 petajoules of added energy use and 790,000 tons of CO₂. The change, though small, shows how faster decarbonization could help offset AI’s indirect emissions.
Harding and Moreno-Cruz caution that while AI can also be used to optimize energy systems and improve industrial efficiency, the near-term effect is an overall rise in energy consumption. They describe AI as a productivity driver that amplifies economic activity -- and, by extension, the resources required to sustain it. Their findings suggest that even small efficiency gains or cleaner power sources could have an outsized effect in balancing AI’s carbon cost as adoption accelerates.
The study concludes that AI’s footprint, though currently limited in scale, deserves early attention from policymakers and industry planners. If firms integrate energy efficiency and renewable sourcing into their AI deployment plans, the technology’s economic advantages can continue without deepening the carbon problem. For now, the message is straightforward: as AI spreads, America’s energy bill grows with it.
Notes: This post was edited/created using GenAI tools. Image: DIW-Aigen.
Read next: Hackers Hand the Wheel to AI: Google Cloud Flags Automation Surge in Cybercrime
The study, published in Environmental Research Letters by IOP Publishing, estimates that widespread AI integration across U.S. industries could add roughly 896,000 tons of carbon dioxide emissions each year. The rise represents about 0.02 percent of national emissions, small in proportion but large enough to draw attention to AI’s energy implications.
The authors, Anthony R. Harding and Juan Moreno-Cruz, modeled AI’s impact across 55 industries, using data on occupational exposure, productivity, and energy intensity. They estimated that AI-related productivity gains translate into an additional 28 petajoules of energy use annually, equal to about 7.8 terawatt hours. In practical terms, that is roughly the annual electricity demand of 300,000 U.S. homes. The research captures not only the direct energy cost of training and running AI systems but also the indirect energy demand created when industries grow more productive.
The analysis shows that energy and emissions impacts vary sharply by sector. Industries with heavy infrastructure or high energy dependence (such as education, healthcare, construction, and retail) see the largest rises. The education sector alone could account for around 12 petajoules of new energy demand and more than 51,000 tons of added CO₂ each year. Publishing activities, by contrast, show only marginal changes, with about 0.003 petajoules of added energy use and less than one-tenth of a kiloton of CO₂. These differences stem from how energy-intensive each industry is and the mix of fuels that power its operations.
At the national level, the study’s estimate of 28 petajoules aligns closely with hardware-based projections from other sources. For instance, analysts have suggested that NVIDIA’s AI servers could consume between 5.7 and 8.9 terawatt hours of electricity a year if operated at full capacity. The new model’s estimate falls within that range, pointing to consistency between top-down economic modeling and bottom-up hardware estimates.
The researchers note that even though AI’s added footprint appears modest, it signals a trend that could accelerate as adoption spreads. As of early 2024, about 5.4 percent of U.S. firms were already using AI tools, a number that had risen by more than one percentage point in just six months. If that rate continues, AI’s share in the broader economy could double within a few years, expanding both productivity and energy demand.
The paper also acknowledges several uncertainties. It assumes fixed energy and emissions intensities for each industry and does not account for future gains in efficiency or changes in the energy mix. Because much of the underlying data dates from 2014, the real-world numbers today might be slightly lower thanks to cleaner power generation and more efficient computing. When the authors recalculated using projected 2023 data, the total impact dropped to 24 petajoules of added energy use and 790,000 tons of CO₂. The change, though small, shows how faster decarbonization could help offset AI’s indirect emissions.
Harding and Moreno-Cruz caution that while AI can also be used to optimize energy systems and improve industrial efficiency, the near-term effect is an overall rise in energy consumption. They describe AI as a productivity driver that amplifies economic activity -- and, by extension, the resources required to sustain it. Their findings suggest that even small efficiency gains or cleaner power sources could have an outsized effect in balancing AI’s carbon cost as adoption accelerates.
The study concludes that AI’s footprint, though currently limited in scale, deserves early attention from policymakers and industry planners. If firms integrate energy efficiency and renewable sourcing into their AI deployment plans, the technology’s economic advantages can continue without deepening the carbon problem. For now, the message is straightforward: as AI spreads, America’s energy bill grows with it.
Notes: This post was edited/created using GenAI tools. Image: DIW-Aigen.
Read next: Hackers Hand the Wheel to AI: Google Cloud Flags Automation Surge in Cybercrime
