AI will consume as much water in 2030 as 1.3 billion people

By 2030, water consumption linked to the use of artificial intelligence will be equivalent to that of 1.3 billion people in sub-Saharan Africa, while it will require nearly three times the annual energy consumption of Pakistan, Bangladesh and Nigeria — countries with a combined population of 650 million. In terms of carbon emissions, these could reach 400 million tonnes of CO₂ equivalent, comparable to the United Kingdom’s total emissions. The operation of AI will require 14,500 square kilometres of land, including infrastructure and supply chains — twice the size of the Jakarta metropolitan area, a megacity with more than 32 million inhabitants, or 10 times that of Mexico City (21 million).

These are some of the figures cited by the authors of a report published this Wednesday by the United Nations University Institute for Water, Environment and Health (UNU-INWEH). In addition to these projections, based on conservative growth estimates, the report also contains striking data about the current situation: if the data centers powering AI were a country, their present electricity consumption (448 terawatt-hours, TWh) would be on a par with that of France.

The institution had previously published reports warning about the carbon emissions associated with the growing use of AI. On this occasion, researchers have also taken into account the energy and water consumed by the data centers that power AI (in the case of water, this includes both cooling systems and electricity generation).

“This report is not a case against artificial intelligence,” said Professor Kaveh Madani, director of UNU-INWEH, in a press release. “It is a call for using it responsibly and addressing its unintended impacts proactively to make it sustainable and equitable. We have a narrow window to ensure that the backbone of the technological revolution of our era develops within planetary limits.”

“The report is an important and timely reminder that AI is not limited to models and algorithms, but also has a real physical and environmental impact determined by data centers, power systems, water-supply systems, land use and hardware supply chains,” said Shaolei Ren, professor of computational engineering at the University of California, Riverside, and an AI sustainability specialist who did not participate in the study.

The underestimated environmental cost of AI

The authors of the report highlight several key messages. One of them is that the environmental cost of AI is being systematically underestimated. Most analyses published so far focus on the carbon emissions associated with training models — the stage before their release, in which tens or hundreds of millions of parameters are processed day and night over several weeks using massive datasets.

“Every kilowatt-hour of electricity used to train or run an AI model carries environmental footprints, including a carbon footprint from the generation mix; a water footprint from electricity production and cooling; and a land footprint from energy infrastructure, reservoirs, and fuel extraction,” the report stresses.

The carbon footprint can vary by up to 70% if, for example, coal is replaced by bioenergy as the source of electricity powering AI. However, this would in turn increase the water footprint thirtyfold and the land footprint a hundredfold. The complexity of managing AI’s environmental impact is therefore extremely high. Low emissions do not equate to low water consumption or low land use. Assessing AI’s environmental impact using a single metric can obscure its harmful effects and shift them to other regions.

“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,” explained Miriam Aczel, the study’s lead author.

Which uses are more polluting

The report also draws another interesting conclusion. Until recently, the prevailing consensus was that most of the energy consumption associated with an AI model occurred during the training phase (that is, before it is used by the public). However, Aczel’s team’s data challenge this view: inference — the computations carried out every time a user submits a query so the model can respond — accounts for the dominant share, between 80% and 90% of total consumption. The success of these tools, used by hundreds of millions of people every day, has reversed the balance.

Researchers also assessed the energy consumption linked to different uses of AI. A standard conversation with a chatbot such as ChatGPT or Gemini uses 200 times more energy than a basic AI function like classifying suspicious emails into spam. Using that as a baseline, generating a synthetic image consumes 1,400 times more, while a short video can require up to 200,000 times more energy.

“This is one of the most comprehensive technical reports on the environmental impact of current AI systems, but the conclusions focus on the impact of GPT-4, which is a model from more than three years ago. And three years in the AI sector is an eternity,” said Álex Hernández, a researcher at the Quebec AI Institute (MILA), led by Yoshua Bengio at the University of Montreal, who did not take part in the study.

That the report’s conclusions are based on data from older models, Hernández says, speaks to the sector’s lack of transparency. “The main limitation of the study is the difficulty of obtaining concrete data on the consumption of current systems,” he added.

Inequality in the distribution of externalities

Another conclusion of the study is the unequal distribution of AI’s benefits and negative externalities. In Ireland, for example — whose lax tax regime has made it the preferred EU location for many major tech companies’ headquarters — data centers already accounted for 21% of total energy consumption in 2023. This has led the country to impose moratoriums on the construction of new facilities of this kind in Dublin.

In Uruguay, plans in 2023 to build a large, water-intensive data center coincided with a drought that depleted Montevideo’s drinking water reserves, making tap water unsafe to consume.

Meanwhile, the authors estimate that by 2030, AI infrastructure will generate 2.5 million tonnes of electronic waste per year (mainly obsolete processors), much of which will accumulate in low-resource countries.

The report also highlights inequality in infrastructure. Only 16% of countries have specialized facilities to run AI, and two of them — the United States and China — account for 90% of total installed capacity. While electronic waste, carbon emissions and water consumption are distributed across many countries, the benefits — namely access to AI applications — are concentrated in a few.

Towards sustainable AI

Like most U.N.-sponsored reports, this one also includes policy recommendations. It calls on governments to require operators to produce standardized reporting on AI’s environmental footprint, and on developers to prioritize selecting appropriate models for each task (avoiding the use of the largest, most resource-intensive systems for simple problems). This idea of “efficiency by design,” as well as the call for greater transparency, are the report’s main demands on the industry.

Hernández, from MILA, says it is important that the U.N. engage in publishing reports on the environmental footprint of AI, a topic until now primarily addressed by academia and investigative journalism. “This report seems to seek the legitimacy of an academic paper while also reaching the policy realm,” he said.

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition