By 2030, AI could consume nearly 3 percent of the world's electricity, produce carbon dioxide emissions comparable to the entire UK's emissions last year, use enough water to quench the thirst of every person on Earth for more than a year and a half, and produce e-waste equivalent to the annual disposal of 250 Eiffel Towers.
Those are the findings of a worrying new report released Wednesday by the United Nations University's Institute for Water, Environment and Health, which says it provides the most comprehensive assessment to date of the environmental costs of artificial intelligence.
While most calculations regarding AI's impact on climate have focused on carbon emissions, researchers say this only tells part of the story. And reducing emissions alone may not be enough to significantly reduce AI's environmental impact.
«A low-carbon economy does not automatically mean water scarcity or low-lying land,» the report says, «and assessing resilience using a single metric can obscure trade-offs and shift the burden onto places already experiencing water or land stress.».
Data centers—massive warehouses filled with servers and cooling systems that operate continuously to power AI—consumed an estimated 448 terawatt-hours of electricity in 2025, roughly equivalent to France's total national consumption. A terawatt-hour is one billion kilowatt-hours, a unit of measurement used in household electricity bills.
Artificial intelligence tasks accounted for approximately 20 percent of the total. If this share grows to the expected 40 percent by 2030, AI-related electricity consumption could reach 374 terawatt-hours. At the current rate, this figure is projected to roughly double to 945 terawatt-hours—enough to power all 1.3 billion people in sub-Saharan Africa for over five years. The area required to generate this amount of electricity would exceed 14,000 square kilometers, roughly the size of Northern Ireland.
Water consumption for cooling this infrastructure also poses additional challenges. By 2025, data centers are estimated to have used 9.3 trillion liters of water—a figure that, according to the report, would meet the drinking water needs of 8.1 billion people worldwide for approximately 1.6 years.
Even if some of this water is returned to the environment, large-scale water extraction puts pressure on aquifers and river systems, especially in regions already experiencing water shortages. In the Netherlands, a large data center consuming large amounts of water during a drought year has sparked protests from local farmers.
Protesters hold signs in front of the Utah State Capitol building to protest the construction of the Stratos data center in Box Elder County (Getty).
Training a single large AI model like ChatGPT-5 requires approximately 100 gigawatt-hours of electricity, equivalent to the annual electricity consumption of 770,000 people in sub-Saharan Africa, as well as approximately one billion liters of water and an area equal to approximately 215 football fields.
However, the report found that the environmental costs of training, significant as they may be, are already outweighed by the costs of daily use. ChatGPT alone processes approximately 2.5 billion queries per day. A typical Google search consumes approximately 0.3 watt-hours of electricity, while AI-powered generative search consumes up to 3 watt-hours—ten times more, given the estimated 5 trillion search queries per year.
The report notes that user choice has a much greater impact on these metrics than is commonly believed. Switching to short-response mode could reduce ChatGPT performance by 30 percent, saving 87 to 98 gigawatt-hours of electricity per year—equivalent to the annual electricity consumption of nearly 760,000 people in sub-Saharan Africa. By eliminating pleasantries and avoiding "please" or "thank you," users make requests more concise and reduce their overall environmental impact at scale.
According to research, the growing popularity of AI-generated videos is becoming a serious environmental concern. Creating a single high-quality AI video requires over 415 watt-hours of electricity, exceeding the energy consumption of hundreds of AI-powered images. The quality of AI-generated videos is also rapidly improving. However, as resolution and frame rate increase, energy consumption also increases exponentially.
Video creation has become an integral part of major social media platforms, and sites are encouraging users to create and publish more AI-powered videos to follow viral trends. The report warns that this is becoming an infrastructure-wide problem.
Professor Alistair Knott of the Centre for Data Science and Artificial Intelligence at Victoria University of Wellington, who was not involved in the report, said that while the study points to growing investment by AI companies, it does not show that these companies are dependent on the growth of the AI market for their own survival.
«"The only way for companies to survive is to continually grow the market for AI-based products, but that's not necessarily what the world needs," he said. "Governments, elected by citizens, are better equipped to make the right decisions about how much AI we need and to balance that need with the environmental impact.".
Pictured are Douglas County servers (UNU Institute of Water, Environment and Health).
The report found that using renewable energy to power data centers does not automatically make them environmentally sustainable. Shifting from coal to bioenergy can reduce the carbon footprint of electricity production by 72 percent, but bioenergy's water footprint is, on average, more than 30 times greater than that of coal, and its land footprint is 100 times greater. For example, Brazil's hydroelectric grid produces electricity with 77 percent fewer carbon emissions than the global average, but its water and land footprint are nearly three times greater than the global average.
In Ireland, data centers now account for 21 percent of the country's total electricity consumption, up from 5 percent ten years ago, exceeding the combined electricity consumption of all urban households. Researchers say this is a result of the growth of artificial intelligence infrastructure outpacing energy sector planning. The National Grid Operator has suspended new permits in the Dublin region until 2028.
Professor Te Taka Keegan of the University of Waikato's Institute for Artificial Intelligence, who was also not involved in the report, said infrastructure concentration raised concerns about environmental justice.
«"The environmental burden falls most heavily on communities least likely to reap the benefits," he said. "As AI becomes embedded in everyday platforms and enabled by default, whether users choose it or not, this environmental footprint accumulates on a massive scale.".
Researchers are calling on governments to begin considering AI infrastructure when planning water and energy supplies. Tech companies should also incorporate environmental considerations into the planning of new features they implement.
«"Technological progress must remain environmentally managed," the authors write. "Real progress depends on embedding sustainable development principles at every level, from equipment and model design to implementation, management, and public use.".