By 2030, the water consumption associated with the use of artificial intelligence (AI) will be equivalent to that of 1.3 billion people in sub-Saharan Africa, while it will require almost three times the energy power consumed annually by Pakistan, Bangladesh, and Nigeria, countries with a combined population of 650 million. As for carbon dioxide emissions, these could reach 400 million tons of CO₂ equivalent, comparable to the total emissions of the United Kingdom. The operation of AI will involve occupying 14,500 square kilometers, including infrastructure and supply chain, twice the metropolitan area of Jakarta, a megacity with more than 32 million inhabitants, or 10 times that of Mexico City (21 million).
These are some of the figures presented 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, they also have striking data about the current situation: if the data centers where AI is computed were a country, their current electricity consumption (448 terawatt-hours, TWh) would be at the level of France.
The institution had previously published reports warning about the carbon emissions involved in the growing use of AI. This time, the researchers have also taken into account the energy and water consumed by the data centers that power AI (in the case of water, including both that used to cool systems and that dedicated to generating electricity). “This report is not against AI,” says Professor Kaveh Madani, director of UNU-INWEH. “It is a call to use this technology responsibly and an attempt to assess its unwanted impacts to make it sustainable and equitable. We must try to ensure that this technological revolution develops within planetary boundaries.”
“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, electrical systems, water supply systems, land use, and hardware supply chains,” points out Shaolei Ren, professor of computational engineering at the University of California, Riverside, and 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 systematically underestimated. Most published analyses focus on the carbon emissions associated with training models (the process before their launch in which tens or hundreds of millions of parameters are computed over massive databases for several weeks, day and night). “However, every kilowatt-hour of electricity consumed to train or operate an AI system also entails a water footprint (associated with cooling and energy generation) and a land footprint (energy infrastructure and supply chains),” the report emphasizes.
The carbon footprint can vary by up to 70%, for example, if coal is replaced by bioenergy as the source of electricity generation powering AI. But that, in turn, would multiply the water footprint by 30 and its land impact by 100. The complexity of managing AI’s environmental footprint is very high. Low emissions do not equate to low water consumption or low land impact. Evaluating AI’s environmental impact with a single metric can hide its harmful effects and shift them to other regions. “If we only rely on carbon emissions, we might think that renewables make AI infrastructure clean, but that means solving one problem while creating others, often in places that did not request it,” says the study’s lead author, Miriam Aczel.
Which uses pollute more
The report offers another interesting conclusion. Until recently, the majority consensus was that the most significant part of the energy consumption associated with an AI model occurred during the training process (that is, before the public starts using it). However, Aczel’s team data refute this: the inference process (the calculations executed each time a request is made to the model so it can respond to commands or prompts) represents the dominant cost, between 80% and 90% of total consumption. The success of these tools, which have hundreds of millions of daily users, has changed the situation.
The researchers have also evaluated the energy consumption associated with different uses of AI. Thus, a standard conversation with a chatbot like ChatGPT or Gemini consumes 200 times more than a basic AI function, such as classifying suspicious emails into the spam folder. Continuing with that reference, 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 over three years old. And three years in the AI sector is an eternity,” notes Álex Hernández, researcher at the Quebec AI Institute (MILA) led by Yoshua Bengio, professor at the University of Montreal, who did not participate in the study.
The fact that the report’s conclusions are based on data from older models, says Hernández, 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 adds.
Inequality in the distribution of externalities
Another conclusion of the study is the inequality in the distribution of AI’s benefits and negative externalities. In Ireland, for example, whose lax tax regime makes it the preferred EU country for most large tech companies to locate their headquarters, data centers already accounted for 21% of total energy consumption in 2023. This has led the country to implement moratoriums on the construction of new such infrastructures in Dublin. In Uruguay, plans in 2023 to build a large data center with intensive water consumption coincided with a drought that depleted Montevideo’s drinking water reserves, causing tap water to become unfit for consumption.
On the other hand, the authors estimate that by 2030, AI infrastructure will generate 2.5 million tons of electronic waste annually (mainly obsolete processors), and much of that waste will accumulate in low-resource countries.
The report also talks about inequality. Only 16% of countries have specialized infrastructure to compute AI, and two of them (the US and China) concentrate 90% of all installed capacity. While electronic waste, carbon emissions, and water consumption are distributed among many countries, the benefits (that is, the use of AI applications) remain in few.
Towards sustainable AI
Like most UN-sponsored reports, this one also includes recommendations for action. For example, it asks governments to require operators to provide standardized reports on AI’s environmental footprint. Developers are urged to favor selecting models appropriate for each task (trying not to use the largest ones, which need more resources, to solve simple problems). That “efficiency by design” and increased transparency are the two main requests to the industry.
Hernández, from MILA, believes it is important for the UN to get involved in publishing reports on an issue, AI’s environmental footprint, which until now has been mainly reserved for academia and journalistic research. “This report seems to aspire to have the legitimacy of an academic article and at the same time reach the political sphere,” he opines.
Read more EL PAÍS launches ‘ConversIA: new dynamic subscription management system optimized by AI’
