Artificial intelligence (AI) is now embedded in daily life. It drafts emails, diagnoses illnesses, predicts markets and answers questions in seconds. For most users, this convenience feels weightless, delivered instantly through a screen.
But behind every artificial intelligence query sits a vast, power-hungry physical infrastructure of data centres, cooling systems and electricity grids. New research shows this infrastructure carries an environmental cost that is large, growing fast and largely hidden from public view.
The numbers behind the AI boom
Global data centres consumed 448 terawatt per hours (TWh) of electricity globally in 2025, more than the entire annual consumption of some countries, according to a report published in June by the United Nations University Institute for Water, Environment and Health (UNU-INWEH). Artificial intelligence accounted for roughly one-fifth of that total, driving a rapid expansion in resource use that highlights the physical and environmental footprint of digital infrastructure.
By 2030, annual power consumption from data centres is projected to double to 945 terawatt-hours. This is nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria—countries collectively home to more than 650 million people, according to the UNU-INWEH report. Artificial intelligence is accounting for 40 percent of the total electricity consumption by data centres. If data centres were a country, researchers say, they would rank among the world’s largest electricity consumers.
Their associated water footprint will equal the basic annual domestic water needs of all 1.3 billion people in Sub-Saharan Africa, and their land footprint will exceed 14,500 square kilometres, roughly twice the Jakarta metropolitan area, home to more than 32 million people.
Behind these projections lies a simple driver: it is not model training but everyday use that consumes most of the power. UNU-INWEH researcher Miriam Aczel said around 90 percent of artificial intelligence-related power use comes from operational requests rather than model training. In other words, the environmental bill grows every time someone types a prompt, not just when a new model is built.
Water consumption and carbon emissions
Electricity is only part of the story. Data centres also depend on enormous volumes of water, largely for cooling servers and for the power plants that supply them.
Last year, data centres consumed 4.5 trillion litres of water, enough to meet the needs of more than 600 million people in Sub-Saharan Africa, according to the UNU-INWEH report. By 2030, that figure is expected to more than double. Water consumption is projected to reach 9.3 trillion litres, while carbon dioxide emissions rise to 399 million tonnes. Researchers note this volume would be enough to cover the basic annual domestic water needs of some 1.3 billion people.
The land required to support this expansion is also increasing. The data centre land footprint is forecast to grow from 6,900 square kilometres last year to more than 14,500 square kilometres by 2030.
Kaveh Madani, director of UNU-INWEH, argues that public debate has failed to grasp the physical scale of artificial intelligence. He said the technology should be understood as infrastructure, not merely software, encompassing everything from power plants to the minerals used in chips. The strain will not be felt everywhere equally, but in specific regions where artificial intelligence expansion collides with existing water and energy pressures.
Real-world warning signs
Communities in several countries are already grappling with the consequences of hosting the artificial intelligence boom’s physical backbone. In Uruguay, plans for a water-intensive data centre coincided with a severe 2023 drought that depleted Montevideo’s freshwater reserves, leaving tap water unsafe to drink. The developer ultimately switched from evaporative cooling to air cooling after public opposition and the country’s worst drought in seven decades.
In Chile, an environmental court partially overturned a data centre permit in 2024 amid concerns over water use. In Querétaro, Mexico, expanding computing infrastructure has drawn on water supplies amid prolonged drought conditions, according to UNU-INWEH researchers.
Elsewhere, governments have moved to slow construction altogether. The Netherlands and Ireland have imposed moratoriums on new data centre projects, citing cumulative water and energy pressures.
Increase in temperature
Meanwhile, research points to another overlooked consequence: data centres appear to be warming the land around them.
A working paper led by Andrea Marinoni at the University of Cambridge, using two decades of NASA satellite temperature data mapped against more than 6,000 data centre locations, found land surface temperatures rose by an average of 2.07 degrees Celsius after facilities began operating, with extremes reaching as high as 9.1 degrees Celsius in some locations.
The study found that more than 340 million people living within 10 kilometres, or about 6.2 miles, of a data centre could be affected by these temperature increases.
The pattern held across continents. In Mexico’s Bajío region, a growing data centre hub, temperatures rose by about 2 degrees Celsius over two decades, and a similar pattern was recorded in Aragón, Spain, where increases in nearby provinces were not mirrored.
Also Read: The Hidden Cost of AI Productivity?
Marinoni has said the planned scale-up of data centres could carry significant consequences for society, welfare and the economy.
However, the paper has not yet been peer-reviewed, and some independent researchers have argued much of the measured land warming stems from construction itself, clearing vegetation and laying concrete, rather than heat generated by computing equipment.
Deborah Andrews, an emeritus professor at London South Bank University who was not involved in the study, said it was nonetheless the first research she had seen focused specifically on the heat islands data centres may produce.
AI and the efficiency trap
One argument frequently used to ease concerns about artificial intelligence’s resource demands is that future models will become more efficient and therefore need less power. UNU-INWEH researchers caution this reasoning is flawed.
The report points to the “Jevons paradox,” an economic principle predicting that when efficiency improvements make a resource cheaper to use, total consumption rises rather than falls.
As artificial intelligence becomes cheaper and more capable, researchers expect wider adoption to outpace any savings gained through efficiency, eroding or eliminating the environmental benefit altogether. Efficiency, in this reading, is not a solution but an accelerant.
The report also flags a growing electronic waste problem. Artificial intelligence infrastructure could generate up to 2.5 million tonnes of electronic waste annually by 2030, with much of the burden likely to fall on lower-income countries with limited capacity for safe disposal.
Researchers behind the UN report are not calling for artificial intelligence to be abandoned. Instead, they are urging governments and technology companies to treat its environmental cost with the same seriousness as its economic promise.
Their recommendations include mandatory environmental disclosures from AI firms, more careful siting of new facilities away from water-stressed regions, and international cooperation on sustainable growth.
