You have probably experienced the warmth of your phone in your hand. Felt it while doing something demanding on it. But that heat is just one end of the picture. On the other end, imagine that same heat, but multiplied by the millions of people continuously interacting with their chosen model, going into a single building that has to be actively and constantly cooled with water and power so it doesn’t cook itself. That warmth in your palm is the only part of the bill you’ll ever feel.

This is the strange paradox of Artificial Intelligence (AI). We have built a technology whose costs are almost perfectly invisible to the people who enjoy its benefits. And as AI has moved from novelty to essential infrastructure, that invisibility is becoming a problem we can no longer afford to overlook.

The numbers are now arriving, and they are sobering. According to the June 2026 report by the United Nations University, the data centres powering AI are projected to consume 945 terawatt-hours of electricity each year by 2030, nearly three times the combined annual electricity use of Pakistan, Bangladesh and Nigeria, countries that are together home to more than 650 million people.

The water needed to cool these facilities tells an even starker story: by the end of the decade, AI’s water footprint could match the basic annual domestic water needs of all 1.3 billion people living in Sub-Saharan Africa.

Consider what this means in practice. A data centre does not sip water politely. In many designs it evaporates fresh, drinkable water to dissipate heat, and it does so continuously, day and night. When such a facility is built in a region already short of water (as many of those regions are because land and electricity are cheap in those locations), it’s competing directly with farms and households for a resource that is already scarce to begin with.

The prompt that produced your image might have been free, but the water that cools the servers is not. There is an implicit price, and somebody, somewhere, is feeling its absence.

There is a common assumption that the real cost of AI lies in training these enormous models, a one-off, headline-grabbing burst of computation. The actual truth is a bit less convenient. The UN University study finds that ordinary, day-to-day use of AI by end users accounts for roughly 80 to 90 per cent of its total energy demand.

In other words, the bill is not confined to a one-off burst of computation in a laboratory. It accumulates every day, by all of us millions of users, one prompt at a time. A single popular frontier model is estimated to handle around 2.5 billion prompts a day. The wonders these systems produce are not occasional. They are constant, and so is their appetite for electricity and water.

But not all the wonders they produce are equal in cost. Asking a model to sort a line of text is cheap. Asking it to conjure an image is not: generating a single picture can demand more than a thousand times the energy of a simple text task, and video is hungrier still.

The playful portrait in a Renaissance dress, multiplied across millions of users indulging in the same trick, is therefore a far heavier consequential act than it seems. We are, in a sense, commissioning oil paintings by the second and assuming they are free because no canvas is involved.

There is also the afterlife of the machines themselves. AI is hungry for specialised chips, and chips age quickly. The same report projects that AI infrastructure could generate up to 2.5 million tonnes of electronic waste every year by 2030.

This waste does not vanish. It is shipped, sorted and often dismantled by hand in communities far from the boardrooms where the next model is announced, exposing people to toxic substances so that the rest of us can enjoy a more capable chatbot.

Here is the part we tend to miss. None of this is an accident or a malfunction. It is the design working exactly as intended. The genius of modern AI products is that they hide their machinery so completely that the user experiences only the magic.

We are encouraged to imagine intelligence as something that happens in “the cloud” — a word chosen, one suspects, precisely because clouds are weightless and clean. But there is no cloud. There is only someone else’s computer, in vast data centres, in someone else’s town, drinking someone else’s water.

A fair objection deserves a fair hearing. Efficiency, the optimists say, will save us: chips improve, models slim down, and each computation costs less than it did a year ago. This is true, and it is also a trap.

The UN University researchers point to a rebound effect familiar to anyone who studies energy: when something becomes cheaper and better, we simply use far more of it. Efficiency does not shrink the footprint when it triggers an explosion in demand. A more fuel-efficient engine is no comfort if it persuades the world to drive ten times as much.

So what would it mean to pay honestly for the machine to think?

Not, I think, to abandon AI. A tool of this power will not be un-invented, nor should it be. The first step is simpler and harder: to make the bill visible. Providers should be required to disclose the water, energy and carbon cost of their systems, much as a food label discloses calories.

A user told that generating a video costs many times more than generating a paragraph might well choose differently and could no longer pretend not to know.

Beyond disclosure lies responsibility. The siting of a data centre should attract the same environmental scrutiny as any other heavy industry, with water and land use weighed honestly and the affected community genuinely consulted.

We already demand this of factories and mines. There is no principled reason to exempt the factories of thought.

The model’s answers will keep arriving, instant and effortless, and our phones will keep growing warm in our hands. But that warmth is the smallest part of the bill. The rest is paid in water, power, and waste, somewhere we conveniently choose not to look.

The question is who else pays for this convenience, and whether we are willing to feel the full weight of a warmth we have so far only held in our hands.

Notes

1. United Nations University Institute for Water, Environment and Health (UNU-INWEH), Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints (Richmond Hill, Ontario, 2026). The report projects that data centres powering AI will consume 945 terawatt-hours of electricity annually by 2030 — nearly triple the combined annual electricity use of Pakistan, Bangladesh and Nigeria — and that AI’s water footprint could match the basic annual domestic needs of 1.3 billion people.
2. UN News, ‘AI’s environmental costs threaten water, land and climate’ (4 June 2026), reporting the UNU-INWEH finding that day-to-day use of AI accounts for roughly 80 to 90 per cent of its total energy demand, rather than the one-off training of models.
3. Ibid. The same study finds that generating a single AI image can require more than a thousand times the energy of a simple text-classification task, with video generation more demanding still.
4. UNU-INWEH (n 1). The report further projects that AI infrastructure could generate up to 2.5 million tonnes of electronic waste annually by 2030, with much of the disposal burden falling on lower-income countries.
5. Rose Willis & Kathryn Conrad / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/