Artificial intelligence is consuming more electricity than ever, yet its potential to help manage climate risk may outweigh its carbon costs — provided the right governance frameworks are put in place. That is the central finding of new research from Vision, an Italian think tank, which attempted to weigh both the energy burden and the climate benefits of AI on a single ledger.
The numbers on AI’s energy demand are striking. According to the International Energy Agency, the power used to train and run large language models has risen roughly tenfold since ChatGPT launched in November 2022. A typical AI-focused data centre now consumes as much electricity as 100,000 households, and the IEA projects that future AI data centres could use as much as 20 times more power than today. Globally, AI accounts for around 155 terawatt-hours of electricity per year — roughly a third of all data centre consumption. That said, this still amounts to only about 0.5% of the world’s total electricity use, and around 0.1% of its total energy consumption when fossil fuels are included. The concentration of data centres in specific locations, such as Ireland and northern Virginia, is already straining local power grids, adding a regional dimension to what might otherwise seem like a manageable global share.
Efficiency improvements are real but offer limited comfort. Power consumption per AI task is falling, yet overall demand is growing faster — a dynamic economists call Jevons’s paradox, where technological efficiency gains are cancelled out by increased use. Google’s CEO Sundar Pichai recently noted that monthly usage of the company’s AI products has grown sevenfold in a single year, reaching 3.2 quadrillion tokens processed per month.
On the other side of the ledger, AI is already showing practical value in climate science. The European Centre for Medium-Range Weather Forecasts deployed an AI model called AIFS in 2025, which outperforms traditional physics-based forecasting systems on extreme weather prediction while using significantly less energy. Beyond weather forecasting, researchers and policymakers see AI as a potential tool for transforming how climate science reaches decision-makers. Rather than waiting years for Intergovernmental Panel on Climate Change assessment reports, a publicly accessible AI model could in principle deliver real-time, localised guidance — telling a farmer which crops to plant or helping a city council plan flood defences — while also reducing the potential for institutional bias in how scientific conclusions are framed.
The Vision research concludes that whether AI ultimately warms or cools the planet will depend less on the technology’s hardware than on the institutions built around it. Climate change has an imperfect but established governance architecture in the COP process, which has met almost every year since 1995. AI, by contrast, operates largely under voluntary codes and occasional international summits, with a UN scientific panel established only last year. Bridging that governance gap, the authors argue, is the most important variable in determining which side of the AI-climate equation wins out.