Does AI even fit within our planetary boundaries?

By 2030, artificial intelligence (AI) will consume enormous amounts of electricity, land, and water – equivalent to the needs of billions of people, scientists write in a recent report by the United Nations. Is AI even possible within planetary boundaries? In principle, yes, states the UN report. ‘Yes,’ says Xin Sun of the University of Groningen, ‘but not with the current geography of data centre expansion.’

FSE Science Newsroom | Charlotte Vlek

The fact that AI data centres require a lot of energy is probably well known by now, states the UN report. But its land and water footprints are often overlooked. The report presents a long list of figures about the projected consumption by 2030, which include:

• 945 terawatt-hours of electricity – roughly twice France’s 2025 consumption

• 9.3 trillion litres of water – equal to the basic annual domestic water needs of all 1.3 billion people living in sub-Saharan Africa

• 14,500 square kilometres of land – about a third of the Netherlands. Part of this land will be used for windmills and/or solar panels needed to generate electricity for these data centres.

‘The huge energy demand of AI data centres results in a large carbon footprint,’ Sun explains. ‘This issue can be solved relatively easily by transitioning to renewable energy sources.’ But this only shifts the problem: by using less fossil fuel, the carbon footprint will drop, but the land and water footprints will rise.

‘The water footprint is the biggest issue,’ Sun says. ‘Water is needed to cool these data centres. And when data centres are built in areas where water is scarce, this causes problems.’ Together with Yanqiu Tao (Nanjing University of Aeronautics and Astronautics), Sun analysed the energy consumption and geographic locations of over 11,000 data centres all over the world, and he warns that these coincide with areas of water stress, such as in Mediterranean Europe, South Africa, and Northern China.

‘My own view is that we should not consider these issues at a global level, but at a local level,’ Sun says. ‘You can talk about AI within planetary limits, but this does not help a local community that faces water restrictions, higher electricity prices, land use pressure, or weak bargaining power.’

‘Data centres in water-stressed regions should face stricter conditions,’ Sun concludes. ‘These could include transparent reporting of energy and water use, limits on freshwater withdrawal, use of reclaimed or non-potable water, or alternatives methods of cooling.’ And the local community should also benefit from the presence of the data centre, Sun remarks, as they are usually not the ones who make use of the AI that is running on these data centres.