UN report quantifies AI's environmental footprint

A new report from FN-universitetets institut UNU-INWEH (United Nations University Institute for Water, Environment and Health) puts numbers on the environmental footprint that AI leaves through its energy use. The point is that AI is not just code, but also physical infrastructure in the form of data centres, chips, cooling water and electricity. For the first time, the researchers quantify carbon, water and land together and emphasise that a green power source does not automatically mean low water or land use.

The most important figures from the report:

• The world’s data centres used about 448 TWh of electricity in 2025. If they were a country, they would be the world’s 11th-largest electricity consumer.
• AI accounted for roughly 20 per cent of that consumption and is expected to reach 40 per cent by 2030.
• Day-to-day use, not model training, is the main culprit. Inference accounts for 80 to 90 per cent of AI’s total energy consumption.
• ChatGPT alone processes around 2.5 billion queries per day.
• A single high-resolution AI-generated video can draw as much electricity as tens of thousands of simple text tasks.
• Only 32 countries host specialised AI data centres, and over 90 per cent of capacity is located in the United States and China.

The report also warns of a rebound effect: when models become cheaper and faster, usage simply rises, so the overall footprint grows despite efficiency gains. The researchers also point to a global imbalance, where communities neighbouring the facilities often bear the local costs in the form of pressure on water resources and pollution, while the benefits accrue elsewhere.

The Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints

AI’s rapid growth drives huge energy, water, and land use, raising environmental and equity challenges across its global infrastructure.

United Nations University