Artificial Intelligence Will Drink the Water of 1.3 Billion People: The Real Numbers Behind the Alarm

A headline that makes noise, a number that needs explaining

By 2030 artificial intelligence will consume as much water as is needed by 1.3 billion people, the entire population of sub-Saharan Africa. The figure comes from a United Nations report published on 3 June 2026, and it is the kind of number built to make news: enormous, concrete, immediately understandable. Before letting it frighten us, though, it is worth understanding exactly what it says and what it does not say, because it is precisely in big numbers that the most frequent misunderstandings lurk.

The report, produced by the Institute for Water, Environment and Health of the United Nations University (UNU-INWEH), estimates that by 2030 AI systems will absorb around 9,300 billion litres of water a year and 945 terawatt-hours of electricity, almost 3% of global electricity consumption. To give a sense of the scale, 9,300 billion litres would be enough to cover the drinking-water needs of the entire world population, more than 8 billion people, for about a year and a half. The comparison with 1.3 billion people is precise but narrow: it refers to the basic annual domestic needs of that population, not to the total water consumption of a region, which would also include agriculture and industry. It is a distinction that many headlines omit, but which changes the meaning of the comparison. The research director, Kaveh Madani, was keen to make it clear: the document is not an indictment of AI, but a call for its responsible use.

This article starts from that number to do something different from raising alarm: to line up the real data, distinguish the solid estimates from the shaky ones, and look at what is actually happening in Italy, where the issue stopped being abstract within the space of a few months.

Why AI needs water

Let us start with the basics, because the link between a chatbot and a glass of water is not obvious. Water comes into play because data centres, the huge warehouses full of servers where AI is trained and run, produce heat. A lot of heat. And for most facilities the cheapest way to get rid of it is evaporation: water is passed through cooling towers, part of it evaporates carrying the heat away, and that part does not come back.

To this direct consumption an indirect one is added, less visible but often larger: the water used by the power plants that produce the electricity the data centres run on. Several studies indicate that this indirect consumption can be up to four times the direct one, and almost no company declares it. It is one of the reasons why estimates vary so much from one source to another: often they are simply counting different things.

There is also a physical reason why data centres built for artificial intelligence are far thirstier than traditional ones. A classic server rack draws between 5 and 15 kilowatts of power, a load that air conditioning handles without problems. Racks designed for AI play in a completely different thermal league: the most recent configurations can reach 120-140 kilowatts per single rack, and at those levels air cooling is no longer enough. Water is needed, because it absorbs heat thousands of times more efficiently. The result is that AI-dedicated facilities can consume from ten to fifty times more water than a traditional structure of the same size.

The numbers, and why they should be taken with a pinch of salt

Alongside the UN estimate, the most discussed figure of recent months comes from a study by researcher Alex de Vries-Gao, published in the journal Patterns in December 2025. His assessment is stark in its breadth: in 2025 artificial intelligence is said to have consumed between 312 and 764 billion litres of water, a quantity comparable to the entire annual world production of bottled water. The enormous gap between the minimum and maximum value is not a flaw: it is the most honest signal we have, because measuring AI's water consumption is even harder than measuring its emissions. De Vries-Gao himself has described the data published by the big companies as "only the tip of the iceberg".

To make sense of these abstract figures, some more tangible examples help. Training GPT-3 is estimated to have evaporated around 700,000 litres of fresh water. A single request of about a hundred words to a large model consumes roughly half a litre: negligible on its own, enormous when multiplied by hundreds of millions of daily interactions. A one-megawatt data centre using evaporative towers can consume up to 25.5 million litres a year for cooling alone, the equivalent of the daily water needs of about 300,000 people.

The technical metric used to measure all this is called WUE, Water Usage Effectiveness: it indicates how many litres of water are needed for every kilowatt-hour of energy going to the servers. The lower it is, the better. But on its own it tells only half the story, because it must be read together with the PUE, the indicator of energy efficiency. Here an engineering paradox hides: evaporative cooling is efficient in electrical terms, because by consuming water it reduces the work of the compressors. In other words, energy is often saved precisely by wasting water. Improving one indicator worsens the other, and for years the industry optimised energy efficiency while leaving water efficiency in the background.

Italy and the Lombardy case

For an Italian reader the natural question is: does all this concern us? The short answer is yes, and more than was thought a year ago. According to a survey by the I-Com institute updated to October 2025, Italy has around 209 data centres, concentrated above all in Milan, with 73 facilities, followed by Rome and Turin. Lombardy is the country's digital centre of gravity: according to a study by A2A and Teha Group, the region concentrates about 65% of the electrical capacity of Italy's data centres, with dozens of active structures in the Milan area alone.

The problem is where all this happens. In recent years the Po Valley has experienced the most severe drought of the last two centuries, and the idea of concentrating water-hungry infrastructure there has sparked a lively debate. Estimates for a single medium-sized facility point to an intake that can reach up to 1.3 million litres of water a day for cooling. In a territory where water is already contested between agriculture, industry, residential use and now digital computing too, it is a pressure that mayors and local committees have begun to challenge openly.

The institutional reaction has arrived. In 2026 Lombardy approved Italy's first regional law on data centres, which mandates high-efficiency cooling technologies and introduces limits on the withdrawal of drinking water for industrial use. It is a first step, but several observers, including the trade unions, have raised an awkward point: once operational, this infrastructure guarantees few direct jobs against an enormous consumption of land, energy and water. The cost-benefit balance for the territory is anything but obvious.

The UN report also offers two concrete cases that explain why the issue is explosive. In Ireland, one of Europe's main hubs, in 2023 data centres consumed 21% of national electricity, surpassing the entire urban population. In Uruguay, plans for a large water-intensive data centre coincided with the 2023 drought that drained Montevideo's reserves, making tap water undrinkable. These are examples showing how the impact is local even when the benefit is remote: a facility can serve users thousands of kilometres away while offloading the pressure onto the water basins of the place where it stands.

What can be done: the technical solutions

The good news is that AI's water consumption is not a fatality: it is largely a design choice. The technologies to reduce it already exist and are spreading. Closed-loop cooling, for example, fills the system with water once and then keeps recirculating it, effectively eliminating evaporation. Some operators estimate a saving of more than 125 million litres a year per single facility compared with traditional evaporative towers, and several companies have announced the switch to these systems in new data centres from 2026 and 2027.

There are also direct-to-chip cooling, which brings the coolant into contact with the processors, and immersion cooling, in which the servers are literally submerged in a non-conductive fluid. Both drastically reduce both water and, in many cases, energy. Alongside these solutions is a growing trend to use recycled or non-potable water in place of drinking water: in 2025 drinking water still accounted for more than half of data centre consumption, but alternative sources are growing at double-digit rates every year.

On the regulatory front, the European Union has introduced the obligation to report the water efficiency of facilities, and several big companies have made public commitments to become "water positive", that is, to return more water to territories than they consume, by the end of the decade. These are commitments to be verified in practice, but they point in a direction. The central demand of the UN report, after all, is not to stop AI: it is to impose transparency on consumption, because without public, verifiable data every discussion remains based on estimates with enormous margins.

What it means for those who use AI every day

The question that concerns everyone remains: does it make sense to worry about the glass of water behind every request to a chatbot? The most reasonable answer is neither alarmism nor indifference. The individual use has a minimal impact, and giving up AI to save half a litre of water would make little sense, especially when the same technology is used to optimise water networks, forecast droughts or reduce industrial waste. The point is not to use AI less, but to be aware of it and to demand transparency from those who build and host these systems.

Awareness, in this case, is already a result in itself. Knowing that behind the apparent immateriality of the digital world there are warehouses, servers, electricity and water helps us make more informed choices, both as citizens facing a new facility on our territory and as companies deciding where and how to invest in technology. The real waste is not the single request, but the disorderly, criterion-free use of a resource that, behind the scenes, has a very concrete physical cost. And the number that makes the most noise, those 1.3 billion people, is worth above all as an invitation to look at what usually stays off the screen.

A note on the data

The figures cited come from recent and authoritative sources: the UNU-INWEH United Nations report picked up by ANSA on 3 June 2026, the study by Alex de Vries-Gao published in Patterns in December 2025, data from the World Economic Forum and the I-Com and A2A-Teha analyses of the Italian context. Estimates of AI's global water consumption remain characterised by very wide margins, due to the limited transparency of companies and the difficulty of measuring indirect consumption. We report them for what they are: indications of order of magnitude, not precise measurements.

At A126 we help companies use technology sensibly: choosing the right tools for each objective, without oversized projects or needless consumption. It is first and foremost a matter of efficiency and budget, but in a world where every computation has a physical cost, it is also a way not to waste resources. Contact us to find out how to integrate the digital into your business in a clear and efficient way.

A126 Corporate Advisors — Understanding the digital, beyond the headlines.