Smaller AI models cut energy use without sacrificing performance
Specialised AI models use 14 to 50 times less energy than large general-purpose models for targeted tasks, Mathias Abitbol writes in Banque de France's Eco Notepad. He argues that efficient models and industrial AI can support decarbonisation.
Bigger models strain electricity grids
The International Energy Agency (IEA) puts data-centre electricity use at 415 TWh in 2024, nearly 1.5 percent of global demand, rising to 945 TWh by 2030 in its base scenario.
Indirect emissions rise from 180 to 320 Mt of CO₂; faster AI adoption raises the increase to 282 Mt.
Local grid constraints prolong fossil-fuel generation.
Yet specialised small models can match larger models on targeted tasks, cutting summarisation energy use tenfold, according to UNESCO.
Its comparison covers Llama 3.1 with 8 billion parameters and smaller models with 61–570 million.
Arcep reports that 84 percent of surveyed large models disclose no environmental information; only 2 percent publish directly measured carbon footprints.
Abitbol proposes taxes, standards or disclosure requirements.
Control systems offer carbon savings
Abitbol et al. estimate $940 billion in welfare gains from avoided industrial emissions, or $1,800 billion including energy savings.
This assumes a 6 percent carbon-intensity reduction across proven applications covering 5 percent of global emissions.
The IEA estimates existing AI applications can cut emissions by three to four times data-centre emissions by 2035.
Metcalfe et al. report 42 percent lower household peak electricity demand with AI-managed charging.
Eurostat puts European business adoption of AI for data-analysis optimisation at just over 5 percent in 2025.
Efficiency alone is not enough
Abitbol makes the net climate benefit conditional on limiting rebound effects and expanding specialised AI adoption.
He proposes energy-saving certificates, listings of proven solutions and funding for pilots and training.
Coordination failures and low rewards for avoided emissions remain barriers to deployment.