The Thirst of the Machine: What the UN's Water Institute Found When It Looked Past AI's Carbon Footprint

There is a particular kind of blind spot that comes from measuring what is easy to measure. For the better part of a decade, the debate about artificial intelligence and the environment has been conducted almost entirely in the currency of carbon. How many tonnes of carbon dioxide did it take to train a large language model? How many flights' worth of emissions does a month of chatbot queries represent? These are useful questions. But a new study from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) — the body known within the UN system as "the UN's think tank on water" — argues that they are also dangerously incomplete. Published on 3 June 2026 under the title Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints, the report makes a case that should unsettle anyone who has taken comfort in a data centre's renewable-energy pledge: judging the sustainability of AI by its carbon footprint alone can hide the damage it does to water and to land, and in some cases can actively encourage decisions that make that damage worse.

The central insight of the UNU-INWEH team, led by institute director Professor Kaveh Madani and lead-authored by Dr Miriam Aczel, is deceptively simple. Every kilowatt-hour of electricity that trains or runs an AI system carries not one environmental footprint but three. There is the carbon released when that power is generated. There is the water consumed — both to cool the servers and, crucially, to generate the electricity in the first place. And there is the land occupied by the energy infrastructure and the supply chains that feed it. The trouble is that these three footprints do not move together. They can, and frequently do, move in opposite directions. As the report puts it, "low-carbon" is not automatically "low-water" or "low-land." Reduce one and you may quietly magnify another.

Nowhere is that trade-off starker than in the report's headline example of switching a power source. Moving electricity generation from coal to bioenergy, the study notes, can on average cut the carbon footprint of that electricity by around 70 per cent. It is exactly the kind of move that looks like unambiguous progress on a carbon ledger. Yet the same switch increases the water footprint of the electricity more than thirty-fold and its land footprint roughly a hundred-fold. A decision that scores as a triumph in a carbon-only accounting system turns out to be, in the fuller picture, a transfer of environmental pressure from the atmosphere onto rivers, aquifers and soil. "What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land," Aczel said on the report's release. "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."

That last phrase — places that didn't ask for it — is the moral centre of the water and land story, and it is worth dwelling on the scale involved before turning to who bears it. The report's projections for 2030 are arresting. If current trajectories hold, the global data centres powering AI are on course to consume around 945 terawatt-hours of electricity by the end of the decade, nearly three per cent of projected world electricity use and roughly twice France's entire 2025 consumption. The carbon associated with that power gets the headlines, and it is large: an estimated 399 million tonnes of carbon dioxide, enough that offsetting it would demand some 6.7 billion trees grown over ten years, about twice the number of trees estimated to stand in the whole of the United Kingdom. But it is the water and land figures attached to the very same electricity that the UNU-INWEH team wants readers to sit with. The associated water footprint is projected at 9.3 trillion litres a year. To make that number legible, the researchers translate it into human terms: it is equal to the basic annual domestic water needs of all 1.3 billion people living in Sub-Saharan Africa. The land footprint, meanwhile, is projected to exceed 14,500 square kilometres — an area roughly twice the size of the Jakarta metropolitan region, home to more than 32 million people.

These are the kind of comparisons that can numb rather than inform, so the report also works at the opposite end of the scale, at the level of a single interaction with an AI system. Here the water footprint becomes almost intimate. Generating one typical AI image, the study estimates, carries an electricity-associated water footprint of roughly 29 millilitres — about two tablespoons. That sounds trivial, and for a single image it is. But the same calculation applied to a single high-complexity AI-generated video yields around 4.1 litres of water, which the report points out is almost the two-day drinking-water requirement of one person. The energy behind those numbers scales just as sharply: the electricity to generate a typical image would run a 10-watt LED bulb for about 17 minutes, while a high-complexity video would keep that same bulb lit for 42 hours. Put differently, a typical AI image demands around 1,450 times the energy of a basic text-classification task, and a single short AI video can consume as much electricity as 200,000 such classifications. The water rides along with the watts.

"AI's Hidden Thirst" visualises the water and land story behind AI's energy use, drawn entirely from UNU-INWEH's published figures: data-centre electricity roughly doubling from 448 TWh (2025) to a projected 945 TWh (2030); the two footprints usually left off the ledger — 9.3 trillion litres of water and 14,500 km² of land by 2030; the "carbon trap," where switching coal to bioenergy cuts carbon ~70% but raises water 30-fold and land 100-fold; and the per-query water cost, from ~29 mL for an image to 4.1 litres for a complex video.

What makes these per-query figures more than a curiosity is the sheer volume they are multiplied by, and the fact that the multiplication is largely invisible to the people doing it. The report is emphatic that the public conversation has fixated on the wrong stage of an AI system's life. Training the model — the one-off, headline-grabbing computation — is not where most of the environmental cost lives. Training GPT-3 was estimated to require 1.3 gigawatt-hours of electricity, and GPT-4 somewhere between 50 and 70 gigawatt-hours, but those are effectively sunk costs paid once. It is inference, the continuous running of a deployed model to answer everyday prompts, that dominates, accounting for an estimated 80 to 90 per cent of an AI system's total energy use. ChatGPT alone is estimated to handle around 2.5 billion prompts a day, which the report translates to roughly 383 gigawatt-hours of electricity a year for that single product. The water footprint of that one service is equivalent to the minimum annual domestic water needs of about half a million people in Sub-Saharan Africa. Every one of those 2.5 billion daily prompts spends a few tablespoons of water somewhere, and almost none of the users pouring them out will ever see the tap.

The report's most sobering contribution to the sustainability debate may be its treatment of efficiency, because it dismantles the most comforting assumption in the whole field. The intuitive hope is that as AI models become more efficient — as the energy and water per query fall — the overall footprint will fall with them. The UNU-INWEH team invokes the rebound effect, sometimes called the Jevons Paradox, to explain why this hope so often fails. As models become more efficient they also become cheaper to run, and as they become cheaper they get used far more. The per-query savings are swallowed whole by the growth in the number of queries. "A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient," Madani said. "But that is only a partial picture of the overall problem. More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains." For water in particular, this is a warning that technological optimism alone will not close the tap; without explicit limits on tokens, output resolution and default video length, the study argues, efficiency improvements will simply be absorbed by volume.

If the abstract projections make the water and land story feel distant, the report's site-level cases bring it uncomfortably close, and they are where the phrase "places that didn't ask for it" acquires real faces and real geography. Consider Ireland, which has become the world's cautionary tale of AI infrastructure outrunning the systems meant to support it. By 2023, data centres accounted for 21 per cent of all metered electricity in the country — more than every urban household combined. The strain grew severe enough that the national grid operator paused new data-centre approvals around Dublin until 2028. Ireland is not, in the report's framing, an outlier so much as a preview: a documented example of what happens when compute growth simply outpaces energy and, by extension, the water systems tied to power generation. The water dimension becomes explicit in the study's other cases. In Querétaro, Mexico, expanding compute infrastructure is drawing on water supplies during prolonged droughts. And in Uruguay, plans for a water-intensive data centre coincided with a 2023 drought that depleted Montevideo's freshwater reserves to the point that tap water became unsafe to drink. These are not hypotheticals or 2030 projections. They are places where the water cost of the global AI boom has already landed on local communities, most of whom are not the ones running the models that the water is helping to cool.

That asymmetry — between where AI's benefits accrue and where its water and land burdens fall — is the thread the report keeps returning to, and it is what elevates the study from an environmental audit into a question of justice. "If you map where data centres are getting built against where water stress is worst, you tend to see the same regions in some instances," said Dr Mir Matin, who manages UNU-INWEH's geospatial and infrastructure analytics programme and co-authored the report. "And the communities living near these sites are not necessarily the ones using the AI being run there. 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 land side of the ledger sharpens the point still further. The report warns that AI infrastructure could generate up to 2.5 million tonnes of electronic waste each year by 2030 — the equivalent, in the study's own image, of discarding nearly 250 Eiffel Towers annually — much of it processed in low-income economies with limited safeguards. And the critical minerals that AI hardware depends on are extracted disproportionately in jurisdictions with weaker environmental oversight, often in the Global South. The land footprint of AI, in other words, is not only the acreage under the servers; it is the mines at one end of the supply chain and the waste dumps at the other.

The geography of who benefits is astonishingly narrow. Only 32 countries in the world host AI-specialised data centres at all, and more than 90 per cent of that capacity is concentrated in just two of them, the United States and China. More than 150 countries currently have little or no access to sovereign AI compute. The report frames this not merely as an economic divide but as an environmental-justice one: excluded countries can still find themselves supplying the minerals and absorbing the e-waste, bearing pieces of the water and land footprint while the strategic and economic benefits flow elsewhere. "The concentrated development of AI infrastructure in the privileged areas of the world is creating a large digital divide," said Professor Tshilidzi Marwala, Rector of the United Nations University and a UN Under-Secretary-General. "AI can certainly advance prosperity and human well-being. Whether it does so equitably is now a governance question, not a technical one."

"Local Costs, Distant Benefits" maps where AI's water and land burden actually falls, using the report's documented cases: Ireland, where data centres took 21% of national metered electricity in 2023 and new Dublin grid approvals are paused to 2028; Uruguay, where a 2023 drought left Montevideo's tap water unsafe as a water-intensive data centre was planned; and Querétaro, Mexico, where compute expansion draws on scarce water amid drought. It closes on the digital divide — only 32 countries host AI data centres, over 90% of capacity sits in two, 150+ have almost none — and up to 2.5 million tonnes of AI e-waste a year by 2030.

For all its warnings, the report is careful not to become a case against artificial intelligence, and that restraint is worth taking seriously. "This report is not a case against artificial intelligence, a technological transformation that is improving the lives of billions of people around the world," Madani — who was named the 2026 Stockholm Water Prize Laureate — was at pains to say. "It is a call for using it responsibly and addressing its unintended impacts proactively to make it sustainable and equitable." The constructive core of the study is a plea for better measurement. Its foremost recommendation is that disclosure standards for AI's environmental impact should require carbon, water and land footprints to be reported jointly, in standardised units, across both training and inference and across jurisdictions, so that regulators, investors and the public can finally compare like with like. Because siting decisions are, in the report's phrase, environmental decisions — the same AI workload can have wildly different water and land profiles depending on where it is run and which grid powers it — the study urges that permitting, environmental-impact assessment and community consultation reflect that reality. And because efficiency alone will not contain growth, it calls for demand-side guardrails: caps on tokens, default low-resolution outputs, and resource budgets that stop the rebound effect from erasing every technical gain.

The value of a study like this is not that it tells us to stop using AI; it plainly does not. Its value is that it changes the unit of account. For years the industry has been able to point to a falling carbon number and call itself greener, while the water drawn from a drought-stricken Uruguayan reservoir and the land opened up for a critical-minerals mine went unrecorded on the same balance sheet. What UNU-INWEH has done is insist that these belong on one page. A world serious about keeping AI within planetary limits, the report concludes, is achievable — capability and stewardship can grow together — but only if we agree first to measure all three of the things AI consumes. Carbon was always only ever one of them. The machine, it turns out, is also thirsty, and it is hungry for ground.