A prompt seems like a light touch. You write a question, wait a few seconds and an answer comes. Everything appears immaterial, almost magical. But behind every request made to ChatGPT or other artificial intelligence systems lies a gigantic infrastructure made up of servers, cooling systems, electrical networks and large quantities of water and energy.
This is the picture that emerges from the new report by the United Nations University Institute for Water, Environment and Health (UNU-INWEH), which invites us to look beyond the surface of AI. Because artificial intelligence is not just software: it has an increasingly heavy physical footprint in terms of electricity, water consumption, land use and raw materials.
How much data centers consume
In 2025, the world’s data centers consumed approximately 448 TWh of electricity. If they were a state, they would have been the eleventh largest energy consumer on the planet. And growth does not seem destined to slow down: by 2030 demand could reach 945 TWh, almost as much as Japan’s entire annual consumption and equal to around 3% of global electricity.
Much of this growth is linked to artificial intelligence. Not only for the training of large models, but above all for their daily use: every text generated, every image created, every video produced requires computing power and therefore energy.
One of the most surprising aspects concerns water. According to the report, the water footprint associated with electricity used by data centers could reach 9.3 trillion liters by 2030. A huge amount, equivalent to the essential domestic water needs of around 1.3 billion people in sub-Saharan Africa.
Water is used both to directly cool the servers and to produce the energy that powers them. For this reason, the authors underline that simply measuring CO2 emissions is not enough: every energy source also brings with it a water and territorial footprint that often remains invisible.
In 2025, data centers consumed approximately 4.5 trillion liters of water and produced 189 million tons of CO2. By 2030, emissions could rise to 399 million tonnes, while land occupation could exceed 14,500 km².
The so-called digital “cloud”, therefore, is much less ethereal than we imagine. Behind the cloud are buildings, industrial plants, transformers, cooling systems and infrastructure that take up space and require resources. And they are often built in areas where water and energy are already under pressure.
Jevons’ paradox and the rebound effect
Then there is another element that deserves attention. Many argue that AI models will become increasingly efficient and therefore consume less. But history teaches us that it doesn’t always work that way. It is the so-called Jevons paradox: when a technology becomes more efficient and less expensive, it tends to be used much more, canceling out part of the benefits obtained.
In the case of artificial intelligence the risk is clear. Faster and cheaper models are being integrated into a growing number of applications, services and platforms. Thus the consumption per single operation decreases, but the total number of operations increases enormously.
The report also highlights that most of the energy consumed by AI does not come from training models, but from so-called inference, i.e. from daily use to respond to user requests. This phase can represent between 80% and 90% of the overall energy consumption.
The type of content required also makes a big difference. Generating an image requires much more energy than a simple text response, while creating videos with AI consumes even more. According to the study, a single generated image can require up to 1,450 times the energy required for simple text classification.
Growth concentrated between the United States and China
Another aspect concerns the global distribution of infrastructures. The study reports that only 32 countries host specialized cloud infrastructure for AI, and more than 90% of capacity is concentrated between the United States and China. Many other countries remain excluded from direct economic benefits but still suffer environmental impacts related to the extraction of critical minerals and the management of electronic waste.
By 2030, AI infrastructure could generate up to 2.5 million tons of e-waste each year. Servers, chips and components are being replaced rapidly to support the race for computing power, further increasing pressure on natural resources.
Possible solutions
For the authors of the report the challenge is not to stop innovation, but to make it transparent and sustainable. It means publishing comparable data on energy consumption, emissions, water and land use. It means designing more efficient models and using computing power in a way that is proportionate to real needs.
Europe is also trying to intervene. The European Commission is working on new efficiency standards for data centers, with criteria that include water consumption and renewable energy.
Meanwhile, according to the International Energy Agency (IEA), the global electricity consumption of data centers could double by 2030. A figure that makes it clear that the future of artificial intelligence is now closely linked to the energy and environmental challenges of the planet.
For years we have thought of digital as something clean because it is invisible. But the UN report reminds us that behind every response generated by AI there is a reality made of electricity, water, raw materials and infrastructure. The cloud, seen up close, is much less impalpable than it seems.
You might also be interested in:
