AI could consume as much water as 1.3 billion people by 2030, warns UN report – Hidden cost of every prompt

A new report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) has put hard figures on what was until now only vaguely understood, the full environmental toll of the world’s rapidly growing appetite for artificial intelligence.

The report, which quantifies the carbon, water, and land footprints of AI’s global electricity use, finds that by 2030 data centres powering AI are projected to consume 945 terawatt-hours of electricity, nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria, countries that together are home to more than 650 million people which means AI could use as much water as 1.3 billion people by 2030.

The reality behind the data

Most people think of electricity when they think of AI’s environmental impact. The UN report argues that framing misses two equally important dimensions: water and land. Cooling the high-density servers inside data centres requires enormous volumes of water, and building the energy infrastructure that powers those servers consumes vast tracts of land. When we put those figures together, these three footprints including carbon, water, land tell a far more troubling story than carbon figures alone.

A standard ChatGPT-style text prompt carries an electricity-associated water footprint of roughly 29 millilitres , about two tablespoons. That sounds trivial until you consider scale. With an estimated 2.5 billion prompts sent every day, the annual water footprint of text queries alone equals around 3.8 billion litres, enough to cover the domestic water needs of half a million people in Sub-Saharan Africa for a year.

Generating a single AI image carries a similar water cost. But video is where consumption shoots up. A single complex AI-generated video can require 4.1 litres of water, almost a two-day drinking supply for one person. If just one-fifth of daily AI video requests are high complexity, the total annual water footprint of AI video generation alone could exceed 13 billion litres.

“If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean, but that is solving one problem while creating other problems, often in places that didn’t ask for it, ” says Miriam Aczel, lead author of the report, UNU-INWEH in a press release.

Training the biggest AI models takes a water supply the size of a small city

The water consumed during the inference phase, answering your questions is only part of the equation. Training large-scale AI models is itself extraordinarily water-intensive. The report estimates that training GPT-4 consumed approximately 600 million litres of water, enough to meet the minimum annual domestic water needs of 81,000 people in Sub-Saharan Africa, or to fill 237 Olympic-sized swimming pools. Next-generation models such as GPT-5 are projected to consume roughly one billion litres during training alone sufficient for the annual needs of more than 135,000 people.

Zooming out to the entire data centre ecosystem, the picture becomes even more arresting. In 2025, the electricity consumed by global data centres carried an associated water footprint of 4.5 trillion litres, enough to fill 1.8 million Olympic pools. AI workloads specifically accounted for approximately 900 billion litres of that total. By 2030, the report projects the water footprint of data centres will reach 9.3 trillion litres, an amount that would meet the basic annual domestic water needs of every person in Sub-Saharan Africa all 1.3 billion of them.

Going green on carbon can make the water problem worse

The report’s most counterintuitive finding is that the choices that look greenest from a climate perspective can magnify water stress. “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, the study’s lead author.

Switching from coal to bioenergy, for instance, cuts electricity’s carbon footprint by 70 per cent but increases its water footprint more than 30-fold and its land footprint 100-fold. The trade-off is not academic, it plays out in specific communities and landscapes,far from the data centres themselves.

Hydropower is perhaps the sharpest illustration of the paradox. Brazil’s electricity grid has a carbon footprint 77 per cent below the global average, making it look like a model of clean energy. But its water intensity, 29 litres per kilowatt-hour is nearly triple the global mean. Canada, Switzerland, and Sweden, all heavy users of hydropower, sit at 21 litres per kilowatt-hour, more than double the global average. By contrast, grids in Hong Kong, the United States, and Australia consume just three to six litres per kilowatt-hour, even though they rely more on fossil fuels.

Real communities are already feeling the pressure

In Mesa, Arizona, a Google data centre is permitted to use 5.5 million cubic metres of water per year, enough to supply 753,000 people in Sub-Saharan Africa. In the Netherlands, a Microsoft complex consumed 84 million litres in 2021 during a severe drought, far exceeding earlier estimates and triggering local opposition. In Querétaro, Mexico, fast-tracked data centre plans are drawing on limited water supplies during prolonged droughts, intensifying pressure on residents already facing shortages.

In Uruguay, the tension boiled over into street protests. In May 2023, a severe drought in Montevideo depleted freshwater reserves and made tap water unsafe to drink. When plans for a water-intensive Google data centre were announced in the middle of the crisis, residents took to the streets, furious that industrial demands appeared to be taking priority over basic human needs.

A digital divide that also divides the environmental burden

The report flags a deeper inequality layered beneath the resource question. As of 2025, only 32 countries, which is 16 per cent of the world’s nations host AI-specialised data centres, and 90 per cent of that capacity is concentrated in just two countries: the United States and China. Wealthier nations are building the infrastructure; lower-income nations are often left absorbing the environmental costs without accessing the economic benefits.

The hardware side of the equation compounds the problem further. AI infrastructure could generate up to 2.5 million metric tonnes of electronic waste every year by 2030. Much of that waste is exported to low-income countries, where frontline communities face exposure to the toxic substances inside discarded servers, cables, and cooling equipment.

What the report says needs to change

The UNU-INWEH researchers argue that the standard practice of measuring AI’s environmental impact through a carbon lens alone is dangerously incomplete. A data centre that switches to hydropower may look responsible on a carbon scorecard while draining a river system or contributing to water stress in a drought-prone region. Without transparent water and land footprint reporting built into AI governance frameworks, the report warns, technological progress will continue to shift environmental burdens onto communities that had no say in where the servers were built or how the electricity was generated.

The message is not that AI must stop expanding. It is that the full cost of expansion needs to be counted and that counting should start now, before the infrastructure of 2030 is already in the ground.