UN warns that AI will “drink” as much water as 1.3 billion people by 2030

Brussels – Behind every image generated by software or every response provided by artificial intelligence lies an environmental cost that is growing and is often underestimated. According to a report published yesterday (3 June) by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), assessing the sustainability of artificial intelligence using a single parameter “can hide trade-offs and shift environmental burdens onto regions already facing water or land stress.” 

Projections for 2030 outlined by scientists estimate that global data centres will consume 945 terawatt-hours (TWh) of electricity, almost three times the combined annual consumption of countries such as Pakistan, Bangladesh, and Nigeria, whose combined population is around 650 million people. The water footprint, however, will be the greatest impact: by the end of the decade, AI will require 9.3 trillion litres of water, an amount equivalent to the annual domestic water needs of all 1.3 billion inhabitants of sub-Saharan Africa. Land consumption will also be massive, exceeding 14,500 square kilometres, an area twice the size of the Jakarta metropolitan area, which is home to over 32 million people.

The core of the report stresses that assessing the sustainability of AI based solely on carbon dioxide emissions is a mistake: “What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land“, said Miriam Aczel, a UNU-INWEH researcher and lead author of the study. According to the report, most expert assessments have remained almost exclusively focused on the carbon emissions generated by training large language models. However, every kilowatt-hour of electricity used to train or operate an artificial intelligence system also entails a water footprint, resulting from cooling and energy production, and a land footprint, resulting from energy infrastructure and supply chains. 

In addition, the report highlights that although public debate has largely focused on the energy required to train the models, 80–90 per cent of total energy consumption stems from day-to-day use. ChatGPT alone processes around 2.5 billion inputs per day, consuming enough electricity in a year to require the planting of trees across an area the size of Manhattan just to offset the emissions.

  Depending on their complexity, different AI products have varying energy consumption levels. For instance, generating a single image requires 1,450 times as much energy as a simple text classification task. A short AI-generated video can consume as much electricity as needed to process 200,000 spam emails. A highly complex video can consume as much as powering an LED bulb for 42 hours, or almost the equivalent of a person’s water consumption for two days (4.1 litres). The report invokes the rebound effect, or the Jevons paradox. This economic concept explains that as models become more efficient, they become cheaper and are used more frequently. But overuse creates new pressures. Incidentally, Professor Kaveh Madani, director of UNU-INWEH, warned that “more efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains.”

The economic and environmental impact is not distributed equally. The report highlights that although AI infrastructure entails environmental costs, it also offers significant economic, security, and sovereignty benefits. Only 32 countries worldwide host data centres specialising in AI, and 90 per cent of computing capacity is concentrated in just two countries, the United States and China. However, over 150 countries currently have limited or no access to sovereign computing resources for AI. The report frames this not just “as an economic divide but as an environmental justice issue” because the environmental costs often fall elsewhere and “excluded countries bear critical minerals extraction and e-waste burdens while the strategic benefits flow elsewhere.” Uruguay and Mexico are cited as examples, where the expansion of digital infrastructure is putting pressure on and putting at risk water reserves during severe drought periods. Furthermore, by 2030, 2.5 million tonnes of electronic waste are expected to be produced annually, often disposed of in low-income economies with weak environmental safeguards. 

“This report is not a case against artificial intelligence,” Madani explains, “but a call to use it responsibly”. Scientists are calling on governments to integrate data centres into national energy and water plans, and on companies to be transparent not only about carbon emissions but also about their water and land use.