AI Could Use Water Equal To 1.3 Billion People By 2030; Each ChatGPT Prompt Has a Hidden Cost

The cost of water due to AI usage is a reason for concern across the globe. That question you typed into ChatGPT this morning? It cost roughly two tablespoons of water. Multiply that by 2.5 billion daily prompts, and the numbers start to look very different.

A landmark new report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) has put hard figures on what was previously only vaguely understoodv- the full environmental toll of the world's surging appetite for artificial intelligence. The findings go well beyond carbon emissions, and they paint a far more complicated picture of what it means to build a green AI future.

The UNU-INWEH report quantifies AI's carbon, water, and land footprints together for the first time. Its headline projection is stark - by 2030, data centres powering AI are expected to consume 945 terawatt-hours of electricity annually, nearly triple the combined electricity use of Pakistan, Bangladesh, and Nigeria, three countries with a combined population of over 650 million.

The water consequences of that energy demand are equally striking. The report projects that by 2030, the water footprint of global data centres will reach 9.3 trillion litres, an amount sufficient to meet the basic annual domestic water needs of every person in Sub-Saharan Africa, all 1.3 billion of them.

What your prompts actually cost

The UN report breaks down consumption to the level of individual interactions, and the figures are revealing.

A standard text prompt sent to a ChatGPT-style model carries a water footprint of approximately 29 millilitres, about two tablespoons. Trivial in isolation, but at an estimated 2.5 billion prompts per day, text queries alone account for roughly 3.8 billion litres of water annually. The report notes that is enough to cover the domestic water needs of half a million people in Sub-Saharan Africa for an entire year.

Image generation carries a comparable per-query cost. Video is where the numbers become harder to dismiss. A single high-complexity AI-generated video can require up to 4.1 litres of water, nearly a two-day drinking supply for one person. If just one-fifth of daily AI video requests fall into that high-complexity category, the annual water footprint of AI video generation alone could surpass 13 billion litres.

Training models is even more water-intensive

The water consumed each time a user sends a prompt is only part of the equation. Training the large-scale models that power these tools is itself enormously water-intensive, and the demands are growing with each new generation.

The report estimates that training GPT-4 consumed approximately 600 million litres of water, enough to fill 237 Olympic-sized swimming pools, or to meet the minimum annual domestic water needs of 81,000 people in Sub-Saharan Africa. Next-generation models such as GPT-5 are projected to consume around one billion litres during training alone, sufficient for the yearly needs of more than 135,000 people.

In 2025, the report found, global data centres consumed electricity with an associated water footprint of 4.5 trillion litres in total, with AI workloads specifically accounting for roughly 900 billion litres of that figure.

Going Green on carbon can make the water problem worse

Perhaps the report's most counterintuitive finding concerns the relationship between clean energy and 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 at UNU-INWEH.

Switching a data centre's power supply from coal to bioenergy, for instance, cuts the electricity's carbon footprint by 70 percent. But it simultaneously increases its water footprint more than 30-fold and its land footprint 100-fold. The trade-off does not stay theoretical, it plays out in specific communities, often far from the data centres themselves.

Hydropower offers perhaps the sharpest illustration of the paradox. Brazil's electricity grid has a carbon footprint 77 percent below the global average, making it appear a model of clean energy. But its water intensity sits at 29 litres per kilowatt-hour, nearly triple the global mean. Canada, Switzerland, and Sweden, all significant users of hydropower, consume more than 21 litres per kilowatt-hour. By contrast, grids in Hong Kong and Australia, which lean more heavily on fossil fuels, consume just three to six litres per kilowatt-hour.

What the UN says must change

The UNU-INWEH researchers argue that measuring AI's environmental impact through a carbon lens alone is dangerously incomplete. A data centre can look responsible on a carbon scorecard while simultaneously draining a river system or exacerbating water stress in a drought-prone region.

The report calls for transparent water and land footprint reporting to be built into AI governance frameworks before, not after, the infrastructure of 2030 is already in the ground. The argument is not that AI must stop growing. It is that the full cost of growth needs to be counted, and that counting needs to start now.