A study finds that heavy adoption of complex generative systems could outpace gains from more efficient hardware and cleaner electricity.
The study presents a method for estimating the environmental footprint of a company’s artificial intelligence portfolio, including electricity use, greenhouse gas emissions, water consumption and resource depletion. It combines estimates for different uses of artificial intelligence with the life-cycle impacts of the hardware and infrastructure involved, from manufacturing through operation and end of life.
The researchers modeled several possible paths to 2030, including changes in usage, hardware efficiency and electricity sources. In the scenario with widespread use of increasingly complex generative systems and software agents, electricity consumption was projected to be 24.4 times higher than the study’s starting point.
How use changes the footprint
Generative artificial intelligence systems used substantially more energy per request than traditional systems in the study’s estimates. Smaller systems used about 25 times more energy than traditional language-processing applications, while a larger system used up to 4,600 times more.
Energy use also increased with the complexity of the task. A basic chat request was estimated to use 1.55 watt-hours for a medium-sized system, compared with 2.64 watt-hours for a system that retrieves information and 8.54 watt-hours for an agent-based workflow. A large model used 186 times more energy for one chat request than a smaller model in the analysis.
The 2030 projections combined growth in artificial intelligence use with expected improvements in computing equipment and changes in electricity sources based on climate scenarios. Under high adoption of complex generative systems and agents, electricity use was projected to rise by a factor of 24.4. The researchers found that improvements in hardware, software and electricity sources would need to work together; any one of these measures alone would not be enough to offset the projected impact.
Why the projections matter
The findings suggest that companies cannot assess the environmental cost of artificial intelligence by looking only at the electricity used during operation. The study also includes hardware manufacturing, water use, resource depletion and end-of-life processes, which can make a substantial contribution to the overall footprint.
The researchers argue that companies need more transparent information from technology providers and common methods for reporting environmental impacts. They also propose a “Return on Environment” measure to help compare artificial intelligence development with sustainability goals.
Models, scenarios and limits
This is a modeling study, not a measurement of future worldwide electricity use. The researchers built a life-cycle assessment that estimates impacts from hardware production, operation and end of life, then combined it with clusters of artificial intelligence use cases and a fictitious portfolio for a large company. They tested several 2030 scenarios involving adoption and efficiency.
The projections depend on assumptions about how quickly use grows, how complex systems become, how efficient computing equipment gets and how electricity sources change. The company portfolio is modeled rather than observed, and the study notes that results for particular systems can differ with the hardware and operating conditions used. More detailed comparisons between model sizes were identified as a need for future research.
// Source
Scientific Reports · 2026 · DOI: 10.1038/s41598-026-77711-2
Authors: Clément Desroches, Louis Ladan, Caroline Vateau, Simon Gosset, Philippe Cordier
Institutions: Capgemini (France)
