UN University report quantifies AI’s carbon, water and land footprints
A UN University report projects AI data centres will draw 945 TWh by 2030, with water and land footprints rivalling the needs of over a billion people.
Photo by Quang Nguyen Vinh on Pexels
By 2030, the data centres that power artificial intelligence could draw 945 terawatt-hours (TWh) of electricity a year — close to triple the combined annual use of Pakistan, Bangladesh and Nigeria, home to more than 650 million people. The figure comes from a United Nations University report, released on 3 June 2026, that measures the AI environmental footprint across three resources — carbon, water and land — rather than carbon alone.
The report, “Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints”, was published by the United Nations University Institute for Water, Environment and Health (UNU-INWEH) and led by the institute’s director, Professor Kaveh Madani. It argues that judging AI sustainability by carbon emissions alone hides trade-offs that shift environmental burdens onto regions already short of water or land.
Measuring the AI environmental footprint beyond carbon
The report’s central claim is that the AI environmental footprint is being mismeasured. Most existing assessments count the carbon emitted while training large models. Every kilowatt-hour used to train or run a model also carries a water footprint, from cooling and power generation, and a land footprint, from energy infrastructure and supply chains. These three footprints do not move together. Switching electricity from coal to bioenergy can cut its carbon footprint by about 70%, the authors note, while raising its water footprint more than thirty-fold and its land footprint a hundred-fold. “Low-carbon”, the report concludes, is not automatically “low-water” or “low-land”.
The scale is already country-sized. Global data centres used an estimated 448 TWh of electricity in 2025; treated as a nation, they would have ranked as the world’s 11th-largest electricity consumer, behind France and ahead of Saudi Arabia. By 2030 that demand is projected to reach 945 TWh, with a water footprint of about 9.3 trillion litres and a land footprint above 14,500 square kilometres.
| Metric | Figure | What it represents |
|---|---|---|
| Data-centre electricity, 2025 | 448 TWh | World’s 11th-largest electricity user, if a country |
| Data-centre electricity, 2030 (projected) | 945 TWh | About 3% of projected world electricity; roughly twice France’s 2025 use |
| Water footprint, 2030 | 9.3 trillion litres | Basic annual domestic water for 1.3 billion people in Sub-Saharan Africa |
| Land footprint, 2030 | >14,500 km² | About twice the Jakarta metropolitan area |
| Inference share of AI energy | 80–90% | Running deployed models, not training them |
| ChatGPT prompts per day | ~2.5 billion | Roughly 383 GWh per year for one product |
| AI compute concentration | >90% in 2 countries | United States and China; 150+ countries lack sovereign compute |
| Projected AI e-waste, 2030 | up to 2.5 Mt/year | Equivalent to discarding nearly 250 Eiffel Towers a year |
Source: UNU-INWEH, “Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints”, 3 June 2026.
Inference, not training, drives AI energy use
Public debate has focused on the electricity needed to train large models — an estimated 1.3 GWh for GPT-3 and between 50 and 70 GWh for GPT-4. The report calls that framing outdated. Once a model is deployed, inference — the continuous running of models to answer everyday prompts — accounts for 80 to 90% of total AI energy use. ChatGPT alone is estimated to handle about 2.5 billion prompts a day, or roughly 383 GWh of electricity a year for a single product.
Per-query demand varies by orders of magnitude with the task. A typical chat query uses around 200 times the energy of basic text classification, an AI-generated image around 1,450 times, and a single short AI video as much electricity as 200,000 text classifications. Model choice, prompt length, output format and resolution all shape the footprint, yet most are set by product defaults the user never sees.
Efficiency gains alone will not contain that growth. The report invokes the rebound effect, also called the Jevons Paradox: as models become more efficient they become cheaper and are used more. “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,” said Madani, who was named the 2026 Stockholm Water Prize Laureate. The authors call for demand-side limits — caps on tokens, resolution and default output length — alongside efficiency.
Local costs, distant benefits
The report stresses that the burdens and benefits of AI’s expansion are unevenly split. In Ireland, data centres accounted for 21% of total metered electricity in 2023, more than all urban households combined; the grid operator has paused new approvals around Dublin until 2028. In Querétaro, Mexico, compute infrastructure is drawing on water supplies during prolonged drought, and in Uruguay a water-intensive data centre was planned as a 2023 drought depleted Montevideo’s reserves. These pressures land where the water footprint bites hardest — a concern Winss Solutions has examined in its look at how the world is heading towards a major water crisis.
Access is concentrated too. Only 32 countries host AI-specialised data centres, and more than 90% of that capacity sits in two — the United States and China — while over 150 countries have little or no sovereign AI compute. That divide tracks the same geography as carbon: see the broader picture in the status of carbon emissions in China, the USA and Europe. The report also warns that AI hardware could generate up to 2.5 million tonnes of electronic waste a year by 2030, much of it processed in low-income economies, while the critical minerals it depends on are mined under weaker environmental oversight.
The water and land penalties of “clean” power are part of the argument for matching AI growth to genuinely low-impact electricity — a question that connects to whether renewable energy is really unreliable due to intermittency. The report’s point is that the greenest choice on carbon can be the worst on water or land, so the trade-off has to be measured, not assumed.
What the report recommends
UNU-INWEH frames a “responsible AI ecosystem” around six principles: transparency; efficiency by design; equity and environmental justice; lifecycle responsibility; global cooperation; and sustainable use. Its core policy message is that carbon-only metrics no longer suffice: disclosure should report carbon, water and land footprints together, in standardised units, across both training and inference. The authors direct specific duties at governments, AI developers, data-centre operators, investors and international institutions, arguing that where a data centre is built — and which grid powers it — determines the footprint of the same workload.
“This report is not a case against artificial intelligence,” Madani said. “It is a call for using it responsibly and addressing its unintended impacts proactively.” Tshilidzi Marwala, Rector of the United Nations University, framed the concentration of AI infrastructure as “a governance question, not a technical one”.
About UNU-INWEH
The United Nations University Institute for Water, Environment and Health is one of 13 institutes that make up the United Nations University, the academic arm of the UN. Known as “the UN’s think tank on water”, it has been hosted and supported by the Government of Canada since 1996 and is based in Richmond Hill, Ontario. The institute marks its 30th anniversary in 2026 and works on water, environment and health challenges through research, training and policy advice for UN member states. Its latest report, carrying the DOI 10.53328/INR26RMA002 and released ahead of World Environment Day, extends that water-centred mandate to the infrastructure behind artificial intelligence, arguing that the technology can stay within planetary limits only if its carbon, water and land costs are measured together.
Sources: United Nations University (UNU-INWEH); EurekAlert! / United Nations University; Phys.org
Featured image: photo by Quang Nguyen Vinh on Pexels (free Pexels license).
For more on this topic, see our explainer on what a prompt actually costs in energy and water.
Become a Sponsor
Our website is the heart of the mission of WINSS – it’s where we share updates, publish research, highlight community impact, and connect with supporters around the world. To keep this essential platform running, updated, and accessible, we rely on the generosity of you, who believe in our work.
We offer the option to sponsor monthly, or just once choosing the amount of your choice. If you run a company, please contact us via info@winssolutions.org.
I specialize in sustainability education, curriculum co-creation, and early-stage project strategy. At WINSS, I craft articles on sustainability, transformative AI, and related topics. When I’m not writing, you’ll find me chasing the perfect sushi roll, exploring cities around the globe, or unwinding with my dog Puffy — the world’s most loyal sidekick.
