A.I. and Water

Generative artificial intelligence does not consume water because it literally “drinks”, but because it depends on a vast material infrastructure: data centres, servers, GPUs, cooling systems, electricity production and semiconductors. Behind the apparent immateriality of a prompt, a generated image or an automated response, there is a physical network made of energy, heat, metals, water and territories.

A.I.’ s water consumption occurs mainly in three ways.

The first concerns the cooling of data centres: the servers that run generative models produce large amounts of heat and, in some cases, are cooled through evaporative systems or cooling towers that use freshwater.

The second concerns electricity production: even when a data centre uses little water directly, the energy that powers it may require significant amounts of water in thermoelectric or nuclear power plants.

The third concerns hardware production: chips, GPUs, servers and electronic components require water-intensive industrial processes, including the use of ultrapure water.

A widely cited study, Making A.I. Less “Thirsty”, distinguishes between on-site water consumption, meaning water used inside the data centre, and off-site consumption, linked to the production of the electricity required for its operation. The same study estimated that training GPT-3 in Microsoft’s data centres in the United States could have consumed around 5.4 million litres of water, including approximately 700,000 litres of direct consumption for cooling.

It is important to distinguish between training and inference.

Training is the initial phase in which a model is built by processing enormous amounts of data: it consumes a great deal of resources over a concentrated period of time.

Inference, by contrast, is the everyday use of the model: every time someone generates a text, an image, a video, code or an audio track. A single request may have a relatively small impact, but the accumulation of billions of daily requests becomes an industrial phenomenon.

For this reason, “per prompt” figures must be treated with caution. The often-cited estimate that GPT-3 would consume around half a litre of water every 10 to 50 medium-length responses depends on many factors: location, time of day, type of data centre, energy source, climate, cooling method and system load. Some analysts consider these estimates too high for more recent models or for infrastructures cooled in different ways. It is therefore more accurate to read them as indicative orders of magnitude, not as universal values.

The problem, however, is not the individual prompt. The problem is systemic multiplication: chatbots integrated everywhere, A.I. search engines, automated generation of images, videos, code, advertising, music, corporate assistants, and financial, medical, military and creative tools. Generative AI is no longer a marginal experiment: it is becoming a widespread infrastructure, embedded in every sector of cultural, economic and communicative production.

The issue is not only environmental, but also political and social. Water is not an abstract resource: it is local, limited and contested. If a large data centre consumes water in a region already affected by drought or water stress, the impact falls on surrounding communities, agriculture, ecosystems and local economies. In some cases, the water consumption of a large data centre can approach that of a small city.

Here, an essential question arises: is it right to use freshwater to power the automated production of content, advertising, synthetic images, industrial texts and mass-generated entertainment, especially in places where water is already scarce?

A.I. can certainly be useful, creative and transformative. But it must be understood for what it is: not an immaterial, invisible technology with no ecological cost, but a system with a high infrastructural impact. It must be designed, used and regulated by taking into account not only its generative power, but also its energy, water, social and territorial costs.

The promise of A.I. is that we may all become artists, musicians, writers, image producers and creators of worlds. But the critical question remains open: will we all become digital creators while the landscape around us turns into desert?

And what about me: how much water did I use to write this article or to create the images ?