The Cost of Better · Part Two

There Is No Cloud

We were told computing had become weightless. It hadn’t — the weight was simply moved out of sight, and the books were quietly arranged so no one would have to add it up.

Of all the things the technology industry has sold the world, the most quietly brilliant is a single word. Not a product. A word. Cloud.

Consider the craftsmanship in it. It tells you your data is up there — aloft, ambient, weightless, somewhere between the weather and heaven. It is the most reassuring noun in modern business, and it was chosen, I have always suspected, by someone who understood that you will not ask hard questions about a thing that sounds like a pleasant afternoon. Nobody worries about the carbon footprint of a cloud. Nobody asks how much water a cloud drinks. That, of course, is the point.

I spent a good part of my career running the machinery underneath that word. I have signed for the diesel generators, costed the cooling, and sat in the rooms where “let’s just move it to the cloud” was said with the airy confidence of a man ordering a second coffee. So let me say the unglamorous thing plainly. There is no cloud. There is someone else’s computer, in someone else’s town, drinking someone else’s water, drawing power from someone else’s grid — and the bill, the real one, is paid by people who never attended the meeting.

There is no cloud. There is only someone else’s computer, in someone else’s town, drinking someone else’s water.— Sumir Nagar

This is the second essay in a series about the hidden cost of technological progress — about the way each advance arrives with a price tag we have become expert at relocating. The first essay followed the cost into the future, into the debt we quietly book against tomorrow. This one follows it somewhere more immediate, and harder to deny once you have seen it: out of the abstract and into the physical world. Into the grid. Into the reservoir. Into the air.

A Word That Weighs Nothing, Attached to a Thing That Weighs a Great Deal

Begin with electricity, because it is the easiest to measure and the hardest to euphemise.

The International Energy Agency — not an activist outfit, but the sober institutional referee of global energy — reports that the world’s data centres consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5% of all the electricity generated on Earth. By itself that sounds modest. The trajectory is what should hold your attention. The IEA projects data-centre consumption will roughly double, to around 945 terawatt-hours, by 2030 — with electricity demand from these facilities growing at several times the rate of overall demand, driven overwhelmingly by artificial intelligence. If today’s data centres were a country, they would already sit among the larger electricity consumers on the planet.

Here is the comparison that lodged in my mind. By 2026, the electricity drawn by the world’s data centres is on the order of the entire annual consumption of a large industrialised nation — the kind of figure usually reserved for whole countries, not a category of building. We constructed a second national grid’s worth of demand out of server racks, and we called it a cloud so that nobody would picture the power stations.

We built a nation’s worth of power demand out of server racks, and called it a cloud so that nobody would picture the power stations.— Sumir Nagar

And this is among the few corners of the economy running directly against the tide. While most of the world is being asked — slowly, painfully, expensively — to decarbonise, the emissions tied to this sector are still climbing. The IEA’s analysis notes that natural gas and, in places, coal are still being drawn upon to cover the surge in demand. In some regions, the newest and most advanced computing on Earth is kept running by the oldest and dirtiest fuel we have. The machine we built to carry us into the future is, in places, powered by the past. There is a metaphor in there, and it is not a flattering one.

The Thirst Nobody Mentions

Electricity, at least, gets discussed. The water does not — and the water is where this stops being an abstraction and starts being your neighbour’s tap.

To stop tens of thousands of servers from cooking themselves, many data centres rely on evaporative cooling, which is a polite way of saying they boil off large quantities of fresh water into the air. Not seawater. Not greywater. Frequently clean, potable, drinking-grade water, because that is what keeps the equipment from fouling. A single large facility can consume on the order of millions of litres a day — the daily water use of a small town. One building, drinking like a community of thousands, so that a query can be answered a little faster.

The aggregate figures are difficult to read calmly, and I will be careful to mark how firm each one is. Industry analyses estimate that AI-associated data centres consumed on the order of hundreds of billions of gallons of water in 2025 — a figure equivalent to the annual household use of a mid-sized population, though the precise totals vary considerably between studies and methods, and I would not stake a claim on any single number. What is better documented is the geography, and it is the part that should trouble us most: an analysis of more than 800 planned US data centres found that around two-thirds are slated for land that has experienced drought within the past year. We are building the thirstiest machines in industrial history, by preference, in some of the places with the least water to spare.

Sit with the arithmetic of that. In one county, residents are asked to let their lawns brown and to shorten their showers. Down the road, a windowless building evaporates a small reservoir’s worth of fresh water so that the rest of us can generate, in a second, an image of a cat in a spacesuit. Nobody designed this as an insult. It emerged, decision by reasonable decision, from a thousand rooms where the cost was always somebody else’s to carry.

In one county, residents are asked to shorten their showers. Down the road, a windowless building evaporates a reservoir. Nobody designed this as an insult. It simply emerged.— Sumir Nagar

I should name something here, because I have sat on the side of the table where it happens. The same boardroom that approves the glossy sustainability report — the one with the leaf on the cover and the carbon-neutral pledge for some comfortably distant year — is frequently the boardroom that approves the data centre in the drought county. There is no contradiction felt in the room, because the two decisions are taken in different meetings, measured on different dashboards, owned by different people, and never laid side by side. Often the pledge itself is honoured not by emitting less but by purchasing the claim of neutrality — carbon offsets, many of which independent investigations have found to represent reductions that never actually occurred, a subject that deserves and will receive its own examination elsewhere. This is the quiet failure of ESG as it is actually practised: it measures the carbon it has learned to put on a slide, and not the reservoir it has not. A sustainability metric is not a conscience. It is a spotlight — and like any spotlight, its chief product is the darkness it creates everywhere it is not pointed.

A sustainability metric is not a conscience. It is a spotlight — and its chief product is the darkness it creates everywhere it is not pointed.— Sumir Nagar

The Objection I Owe You

This is the point in the essay where the well-informed reader prepares to push back, and the push-back is a good one, so I will make it for them — properly, rather than pretend it does not exist.

It goes like this: but we are getting more efficient. And this is true — emphatically, measurably true. The IEA itself notes that the energy required for a given unit of computing work has been falling rapidly, a rate of improvement with few precedents in the history of energy. Every generation of chip does more per watt. The industry has even built itself yardsticks for this — Power Usage Effectiveness and Water Usage Effectiveness, the ratios by which a data centre’s overhead is measured and, year on year, improved. By the metric of efficiency-per-task, this is one of the great optimisation stories of our age.

And it does not save us, for a reason an economist named more than a century and a half ago. When William Stanley Jevons studied coal-fired steam engines in 1865, he observed that making engines more efficient did not reduce Britain’s coal consumption. It increased it — because efficiency made the engines cheaper to run, which made them worth running everywhere, which consumed far more coal in total than the wasteful few ever had. Efficiency did not shrink demand. It unleashed it. The paradox has carried his name ever since.

That is precisely what is happening here. We made each unit of computation cheaper, so we now run an astronomically larger number of them. The per-task cost fell and the total bill exploded, because the cheaper a thing becomes, the more reasons we invent to do it. Efficiency, in a system addicted to growth, is not a brake. It is an accelerator wearing the costume of a brake.

Efficiency, in a system addicted to growth, is not a brake. It is an accelerator wearing the costume of a brake.— Sumir Nagar

The engineers are not idle in the face of this, and it would be unfair to suggest otherwise. They are pursuing genuinely ingenious remedies — direct-to-chip liquid cooling, the full immersion of servers in non-conductive fluid, closed-loop systems that evaporate no water at all. Microsoft went so far as to sink a sealed data centre to the floor of the sea off the Orkney Islands, cooled by the ocean itself; over two years, the submerged servers proved roughly eight times more reliable than their equivalents on land. And then the company set the idea aside — not because it had failed, but because the appetite it was built to serve had moved on. The newest workloads, driven by AI, demand hardware so dense, so power-hungry, and so constantly upgraded that no sealed vessel resting on a seabed could keep pace with the servicing they require. The remedy worked. The hunger outgrew it.

That single episode contains the whole difficulty. We are extraordinarily clever at making each unit of computation cheaper and cooler, and not clever at all about wanting less of it. Every gain on the supply side is met, and exceeded, by an expansion of demand that no efficiency has yet managed to outrun — which is why the cooling towers keep rising and the reservoirs keep falling even as each individual machine grows more frugal than the last.

We are extraordinarily clever at making each unit of computation cheaper and cooler, and not clever at all about wanting less of it.— Sumir Nagar

And the demand does not only grow; it churns. Every few years the industry decides that some new framework, language, or architecture is the future, and the physical plant beneath it must be rebuilt to suit — new hardware for the new paradigm, then new hardware again when that paradigm turns out to have been a detour. I have written separately about that graveyard of sure things, the long parade of technologies we mistook for destinations; what matters here is that each turn of that wheel lands, eventually, as concrete and steel and water in a field somewhere. The software fashions are weightless. The data centres they require are not.

The Cost Was Never Removed. It Was Relocated.

Which returns me to the spine that runs through this entire series.

The promise of the cloud — of all of this — was dematerialisation. The seductive idea that we had finally escaped the physical: no more discs, no more paper, no more weight, only clean weightless information humming in the ether. It was one of the most successful stories the industry ever told, and like the best stories it was not quite a lie. Your laptop genuinely got lighter. Your shelves genuinely emptied. The friction genuinely vanished — for you.

But matter is not so easily abolished. The weight did not disappear; it moved. It moved from your desk to a server hall in some out-of-town business park. From your hands to a substation you will never see. From your awareness to a reservoir being drawn down, year on year, in a county whose name you do not know. Every frictionless thing you now enjoy — the instant search, the streaming film, the AI that drafts your email before you have finished thinking it — is frictionless for you precisely because the friction was exported to a place you will never have to look at. Smooth here means rough somewhere else. The convenience is real. It was simply never free, and you were never the one invoiced.

I want to be careful here, because the lazy version of this essay ends in finger-wagging, and I have no interest in telling you to feel guilty about sending an email. Guilt is a poor substitute for accounting, and individual restraint is not the lever that matters. The lever that matters is visibility. The reason this cost grows unchecked is the same reason the technical debt in the first essay grows unchecked: nobody has to look at it. It does not appear on your screen. It does not appear on your conscience. It appears, if it appears anywhere, as a line in a utility report in a county you have never heard of, read by an official whose objection will be quietly outweighed by the jobs and the tax base.

That is the deeper pattern, and it is why “the cloud” was such an effective piece of language. A cost you cannot see is a cost you will never vote against. The genius was never only in the engineering. It was in the naming — in choosing a word soft enough that no one would think to weigh it.

There is one more relocation to follow, and it is the hardest of the three, because at least the grid and the reservoir can be measured by anyone willing to look. The final essay in this series is about a cost we have moved somewhere even measurement cannot easily reach — out of the future, out of the physical world, and out of the one place that should have stopped all of this long ago: the question of whether we should have done it at all.