The United Nations University Institute for Water, Environment and Health released the most comprehensive accounting yet of what AI is doing to the planet, and its central message is that the conversation about carbon has been hiding the rest of the bill. Global data centres used about 448 terawatt-hours of electricity last year, on a par with France's national consumption. AI workloads accounted for roughly a fifth of that. By 2030 the same data centres are projected to need 945 terawatt-hours, nearly the entire annual electricity demand of Japan, with AI's share rising to 40 percent.
The interesting numbers, though, are the ones outside the carbon column. To produce that electricity will require an estimated 9.3 trillion litres of water and roughly 14,500 square kilometres of land, an area about twice the Jakarta metropolitan region. The report's lead author, Kaveh Madani, makes the point bluntly: "Public debate still often treats AI as software, but AI is also physical infrastructure: data centres, electricity generation, cooling systems, transmission networks, chips, minerals, land and water."
What makes the report unusual is that it punctures some of the comfortable assumptions of the green-AI conversation. Switching a data centre from coal to bioenergy can cut electricity-related carbon by 72 percent, the researchers note, but the water footprint of bioenergy is more than 30 times that of coal, and its land footprint is around 100 times bigger. Brazil's hydro-powered grid produces electricity 77 percent below the global carbon average, but its water and land footprints come in at nearly triple. "Low-carbon is not automatically low-water or low-land," the authors write. Evaluating sustainability by a single metric simply moves the burden somewhere else, usually somewhere with less political weight.
The report also shifts attention from training to use. Public alarm has fixed on the energy cost of training large models, but the UNU researchers find that day-to-day inference accounts for 80 to 90 percent of total AI energy demand. ChatGPT is estimated to process around 2.5 billion prompts a day. A conventional Google search uses about 0.3 watt-hours; an AI-generated response uses up to ten times that, across five trillion searches a year. Switching the model into a concise response mode can cut output by 30 percent, the authors note, saving 87 to 98 gigawatt-hours of electricity annually. Removing pleasantries, the small "please" and "thank you" people add to prompts, has the same effect at scale.
The local picture is starker than the global one. Data centres now account for 21 percent of Ireland's metered electricity, up from 5 percent a decade ago, exceeding all urban household consumption combined. The national grid operator has paused new approvals around Dublin until 2028. In the Netherlands, a data centre drawing on water during a drought year drew opposition from farmers. The European Commission, on Wednesday, said it would propose a law later this year to push households on to AI-powered smart meters that shift their consumption to off-peak hours, partly to leave more grid headroom for the industrial and data-centre demand to come.
Underneath all the numbers sits an uneven map. More than 90 percent of AI-specialised computing capacity is concentrated in two countries, the United States and China, while over 150 countries have no significant domestic AI infrastructure at all. E-waste from AI hardware is on track to reach 2.5 million tonnes a year by 2030, much of it likely to be processed in lower-income countries. The UN researchers do not argue against AI. They argue that the resource ledger needs to be visible before more capacity is committed. "Right now," Madani said, "the competition for growing faster than others overshadows the very basic principles of sustainable growth."