Renewable-Powered AI Isn’t As Clean As It Looks, According to UN Scientists

Swap a data centre’s coal-fired electricity for bioenergy and, on average, you slash its carbon footprint by around 70 percent. Job done, you might think. Except the water footprint of that same electricity climbs more than thirtyfold, and the land it needs balloons by a factor of a hundred. The carbon goes down. Almost everything else goes up. That uncomfortable swap sits at the heart of a new report from the United Nations University, and it is forcing a rethink of what “green AI” actually means.

The report, from the UN University Institute for Water, Environment and Health, is one of the most comprehensive attempts yet to tally what artificial intelligence costs the planet. Not just in carbon, the metric everyone reaches for, but in water and land too.

“What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land,” says Dr. Miriam Aczel, the report’s lead author. It is a deceptively simple point, and a slippery one. Carbon, water and land footprints do not move in step. Push one down and you can shove another up, often somewhere else entirely. Brazil’s grid, dominated by hydropower, runs about 77 percent below the global carbon average; its water and land footprints, though, are nearly triple the global mean.

So the cheerful industry line that renewable-powered data centres are clean, green and sustainable turns out to be, at best, a partial truth. Low-carbon is not automatically low-water or low-land.

The scale of the thing is what makes those trade-offs matter. In 2025 the world’s data centres swallowed an estimated 448 terawatt-hours of electricity, which, if they were a country, would rank them eleventh on the planet, just behind France and ahead of Saudi Arabia. By 2030 that figure could roughly double to 945 TWh. The associated water footprint at that point, some 9.3 trillion litres, would cover the basic annual domestic needs of all 1.3 billion people in sub-Saharan Africa.

“If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean but that is solving one problem while creating other problems, often in places that didn’t ask for it,” says Aczel.

The tip of the iceberg is the training

Here is the part that has been hiding in plain sight. Most of the worry about AI’s energy appetite has fixated on training, the months-long, GPU-guzzling process of building a model like GPT-4. But training, it turns out, is the tip of the iceberg. Once a model goes live, it shifts into what engineers call inference: the endless business of answering everyday prompts, billions of them, day after day. That phase accounts for an estimated 80 to 90 percent of total AI energy use. ChatGPT alone fields something like 2.5 billion prompts a day, which works out to roughly 383 gigawatt-hours a year for a single product. And the per-query cost is wildly uneven. A typical chat is about 200 times more demanding than a basic spam filter; an AI-generated image, roughly 1,450 times; a single short video can match 200,000 of those spam classifications.

Which is why the report keeps circling back to a counterintuitive worry: that getting better at this might make things worse. Cheaper, faster, more efficient AI just gets used more.

“A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient,” says Professor Kaveh Madani, who directs the institute and led the investigation. “But that is only a partial picture of the overall problem.”

It is the rebound effect, sometimes called the Jevons Paradox after the Victorian economist who noticed that more efficient steam engines led Britain to burn more coal, not less. Madani puts the modern version bluntly: more efficient and affordable AI means more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains. Without hard caps, on tokens, on resolution, on default output length, the savings from each clever optimisation tend to evaporate under the sheer volume of use. Even something as small as how politely you phrase a prompt nudges the dial; the report reckons a concise response mode that trims tokens by about 30 percent could save somewhere between 87 and 98 GWh of electricity a year, roughly the residential power use of three-quarters of a million people in sub-Saharan Africa.

Local costs, distant benefits

The burdens, though, do not fall where the benefits do. That is perhaps the report’s sharpest edge.

In Ireland, data centres ate up 21 percent of all metered electricity in 2023, more than every urban household combined, and the grid operator has frozen new connections around Dublin until 2028. In Querétaro, Mexico, and in drought-stricken Uruguay, thirsty new facilities have drawn on water supplies that local people were already short of.

“And the communities living near these sites are not necessarily the ones using the AI being run there,” says Dr. Mir Matin, who runs the institute’s geospatial analytics programme. “That asymmetry is the issue. Without fixing it, we’ll just be repeating older patterns, where some places carry the costs and other places capture the benefits.”

The geography of who builds AI is just as lopsided. Only 32 countries host the specialised data centres that frontier AI runs on, and more than 90 percent of that capacity sits in just two of them, the United States and China. More than 150 countries have essentially no sovereign compute at all, yet many still supply the critical minerals dug out for AI hardware and inherit the e-waste at the other end, up to 2.5 million tonnes a year by 2030, the equivalent of binning about 250 Eiffel Towers. “AI can certainly advance prosperity and human well-being,” says Professor Tshilidzi Marwala, the university’s rector and a UN under-secretary-general. “Whether it does so equitably is now a governance question, not a technical one.”

None of this, the authors are at pains to stress, is an argument against AI. Madani calls the report a call for using the technology responsibly rather than a case against it, and points to a narrow window in which the backbone of this technological revolution might still be steered to develop within planetary limits. The fix they propose is mundane and, in a way, reassuring: measure all three footprints, not just carbon; show users the cost of their choices; treat where a data centre is built as the environmental decision it plainly is. Behind every prompt, image and video, as Madani puts it, lies a growing infrastructure of energy, water, land, minerals and waste, profoundly physical for something so often described as weightless.

The footprint, in other words, is not fixed. It depends on how much we use these systems, what we use them for, and crucially where. Get those choices right while the window is still open, and the next decade of AI need not be written in drained reservoirs and exported rubbish.

DOI: 10.53328/INR26RMA002

Frequently Asked Questions

Why isn’t a low-carbon data centre automatically a green one?

Because carbon, water and land footprints do not move together. A grid that runs on hydropower or bioenergy can cut carbon emissions sharply while using far more water and land per unit of electricity, sometimes by factors of thirty or a hundred. Judging AI sustainability on carbon alone can therefore hide where the real burden is landing, often on regions already short of water.

Is it true that running AI uses more energy than training it?

For most deployed models, yes. Training a frontier model is enormously energy-intensive, but it happens once; answering everyday prompts happens billions of times a day. That ongoing “inference” phase is estimated to account for 80 to 90 percent of AI’s total energy use, which is why the report argues attention should shift from the biggest training runs to the defaults baked into everyday products.

How can making AI more efficient end up increasing its footprint?

It is a version of the Jevons Paradox: when something becomes cheaper and faster, people use a lot more of it. Efficiency gains per query get swallowed by sheer growth in volume, so total energy, water and land use can keep climbing. The report argues that efficiency only helps if it comes paired with hard limits on things like tokens, resolution and output length.

Does the kind of prompt I type really change AI’s environmental impact?

At the scale of billions of daily queries, surprisingly, yes. The report estimates that a concise response mode trimming output by roughly 30 percent could save somewhere between 87 and 98 gigawatt-hours of electricity a year. The energy gap between tasks is even starker: a single AI image can cost over a thousand times more than a simple text classification, and a short video far more again.

Which places carry the cost of AI they may never use?

Often communities near the data centres and at the ends of the supply chain. Ireland’s grid is already straining under data-centre demand, while facilities in Mexico and Uruguay have drawn on scarce water, and many countries that host mining or inherit e-waste capture little of AI’s benefit. With more than 90 percent of specialised compute concentrated in two countries, the report frames this imbalance as a governance challenge rather than a technical one.