Nine hundred and forty-five terawatt-hours a year. That is the figure the United Nations University attributes to the data centres powering artificial intelligence by 2030, in a study published in early June by the Institute for Water, Environment and Health (UNU-INWEH) and reported by UN News. For scale: that is nearly triple the combined annual electricity use of Pakistan, Bangladesh and Nigeria, countries that together are home to more than 650 million people.
The number, staggering on its own, is only the start. The study also calculates a water footprint, tied to cooling systems and energy production, that could by the end of the decade equal the basic domestic water needs of 1.3 billion people in Sub-Saharan Africa. And a land footprint, linked to power generation and supply chains, that could exceed 14,500 square kilometres, roughly twice the size of the Jakarta metropolitan area.
Perhaps the most counterintuitive finding concerns where AI’s energy use actually piles up. Public debate has focused for years on the energy needed to train large models, but according to the study, day-to-day use, every prompt to a virtual assistant, every generated image, accounts for 80 to 90% of total energy demand. One widely used AI service alone processes an estimated 2.5 billion prompts a day, with consumption measured in hundreds of gigawatt-hours a year. And not every request weighs the same: generating a single image can take more than a thousand times the energy of simple text classification, while generating video demands even more.
There is also a paradox the study does not shy away from. Switching to renewable energy sources can cut carbon emissions, but in some cases significantly raises water consumption or land use, shifting the problem rather than solving it. And the environmental costs are not evenly spread: in some countries, data centres already absorb a growing share of national electricity, while elsewhere new facilities draw heavily on water supplies already under strain, sometimes amid drought.
Rounding out the picture is a striking figure on where computing power sits geographically: more than 90% of specialised AI computing capacity is concentrated in just two countries, the United States and China, while more than 150 nations lack significant domestic AI infrastructure, a gap that UNU researchers say also raises a question of environmental justice, with some countries bearing the costs without sharing in AI-driven growth.
The study, its authors are careful to note, is not an argument against artificial intelligence itself, but a call to plan its expansion within the planet’s limits, folding AI infrastructure into energy, water and land-use planning rather than letting it grow as if resources were infinite.
