A report by UNU-INWEH argues that a carbon-only assessment misses the point: water, land and e-waste weigh just as much.
Published on June 3, 2026 by the United Nations University's Institute for Water, Environment and Health, the report "Environmental Cost of AI's Energy Use" places AI on three axes that are rarely discussed together. By 2030, its data centers would consume 945 terawatt hours of electricity, about three times the combined consumption of Pakistan, Bangladesh and Nigeria, countries with a total population of 650 million.
On water, the figure put forward reaches 9,300 billion liters, the equivalent of the basic domestic needs of 1.3 billion inhabitants of sub-Saharan Africa. The land area is estimated to exceed 14,500 km², almost twice the size of the Jakarta metropolitan area, and the associated electronic waste is estimated to rise to 2.5 million tons per year. The authors, led by Miriam Aczel under the direction of Kaveh Madani, also point out that inference, not training, accounts for 80 to 90% of total energy expenditure.
Their conclusion is less about volumes than about geography: the concentration of capacity in two countries places environmental costs elsewhere than where the benefits are captured, which they say calls for multi-stakeholder governance and not just a carbon footprint.
Read the full article on the original source
Go to the source →