I have known that the cloud was never clean for about as long as I have been measuring websites for a living. It is close to the first lesson of this work, that every page drawn on a screen is powered by a machine humming away in a building somewhere, and putting honest numbers to that fact, in grams of carbon a client can actually read, has been part of my job for years. So the evening I want to describe here did not begin with a revelation.
The evening in question found me at my desk, reading through the latest figures on how much electricity, and water, the new wave of AI actually draws, and I had let numbers like these slide past me before, the way you learn to when your field produces them weekly. One of them caught this time and would not let go, the projection that the world’s data centres are on course to roughly double their electricity use inside five years, climbing towards something close to the entire consumption of a country the size of Japan. I sat with it for longer than I meant to. The dirt itself was old news to me, and what held me there was the slow understanding that, for the first time in my working life, I could not find the lever. Every environmental cost the web has handed me until now arrived with a handle attached, where you measured the site, trimmed what could be trimmed, and watched the cost come down where a client could see it. This was a second problem laid on top of that old familiar one, far larger and built along different lines, and for the life of me I could not see how we optimise our way clear of it the way we have been doing with websites for years.
It helps to say plainly what the older problem looked like, because knowing the cloud was dirty never felt like grounds for despair, and for a long stretch of years the dirt answered to careful effort. We streamed films on a Friday night, we moved our clients’ sites and systems onto servers we would never see, and we told ourselves the internet was marching steadily towards net zero. The footprint was real, and those of us doing this work knew it was real, but it sat within reach of the tools we already had. The villain of that era, if it had one, was video, and the useful thing about video was that you could optimise it. You could reach for a better codec, defer the scripts, lazy-load the images, put a CDN in front of the heavy assets and watch the numbers fall. Craft visibly moved the needle, and because it did, it was easy to believe that the craft was the whole answer.
Looking back at that evening now, what strikes me is that nothing new had actually been revealed to me. The knowledge itself was old and settled. What gave way was the belief that had grown up beside it, the belief that careful work could always reach the problem, and it gave way without my expertise failing me in the slightest. I was doing the work I have always done, measuring, trimming, handing a client a smaller number than the one they walked in with, and the work was sound. But a single product is fielding at least two and a half billion prompts a day, a figure that was already a floor when it was published last year, and each one of those reaches back into a data centre to be answered. Set against something like that, my careful savings were a teaspoon held against a rising tide. The discipline I had spent years building held up under that scale, working well enough to let me see the size of the thing clearly and without flinching, which is not at all the same as despairing of it.
The figures behind that doubling are worth setting out properly, since a cost is hard to weigh until it has been written down. The International Energy Agency gathered the energy side of this into one place earlier this year, and the shape of it is plain. Electricity drawn by the world’s data centres grew by around seventeen per cent in 2025, which is already steep, but the electricity drawn specifically by AI-focused centres jumped by roughly half in that single year. The climb from somewhere near 485 terawatt hours now to around 950 by the end of the decade is the doubling I had been staring at that evening, and it is not the kind of problem you optimise your way out of one website at a time.
What fewer people have noticed, though, has nothing to do with electricity at all, and it is the part I found genuinely surprising. AI has started to eat the world’s capacity to make memory. In a single quarter at the start of this year, the contract price of ordinary computer memory rose by something close to eighty or ninety per cent, the steepest jump the industry has on record, and the cause, when you trace it back, turns out to be a decision about where to point the factories rather than any shortage of raw material in the ground. The three companies that make roughly nine in ten of the world’s memory chips have been turning their wafer capacity over to the high-bandwidth memory that sits beside AI accelerators, and that memory is hungry in a way that compounds the squeeze, because every bit of it swallows about three times the manufacturing capacity of the ordinary kind. Make more of the scarce thing and you make the common thing scarcer still. The forecast is that AI data centres will absorb something like seventy per cent of the high-end memory produced this year, and the result lands on a bill that has nothing to do with AI at all, in the rising price of the RAM and the storage inside an ordinary laptop or phone. The cost of nearly every computer that is not an AI server has gone up, in order to feed the ones that are.
I want to be fair to the other side of this, because there is a reasonable version of it. The same energy report is careful to point out that data centres, even after all this growth, will still account for only around three per cent of global electricity by 2030, which is a long way from the thing that ends the world. The companies building these centres are signing more renewable power agreements than almost anyone, pouring money into closed-loop cooling and water-positive pledges, and there is a genuine case that AI will pay some of its debt back by making other systems leaner, through smarter grids, faster materials discovery, fewer wasted journeys. I do not wave any of that away. But none of it touches the thing I actually saw, which is that demand at the source is growing faster than efficiency can answer, and we have been here before. We were sold the same comfort about the internet itself, that it was heading towards net zero on the strength of better hardware and cleaner grids, and demand simply grew to swallow every gain we made. When a query gets cheaper to run, the world responds by running a great many more of them, and the savings evaporate into the growth. A pattern that has caught us once already deserves more suspicion than hope.
This is where I have had to be honest with myself about what my own field is actually for. For ten years a great deal of sustainable web design was practised as optimisation, and I practised it that way too. Shave the bytes, compress the video, defer the scripts, and the footprint of the thing you built came down where you could see it. That work was good and it remains good, and I am not about to tell anyone to stop doing it. But I have come to think that optimisation was only ever the practice. The purpose, underneath all the technique, was learning to see the true cost of what we build, the whole of it, including the parts that never reach the invoice and never appear on a chart in a client meeting. The byte-shaving was how we trained the eye. And the strange gift of this moment is that the eye is exactly what the moment asks for, only now it has to look at something far larger than a single codebase.
So I have stopped telling people that optimising their own site is the same as doing their part, because I think that was a half-truth even in the era of video, and it has hardened into a more dangerous one now. A well-built and lightly-loaded site is still worth making, and I will go on arguing for it for as long as anyone will listen. It has simply become the floor of this work rather than the finish line. The larger leverage has moved somewhere harder to reach, into the choices most of us are never asked about directly, what a company buys and from whom, what its products do by default, whether a feature reaches for a generative model because the task genuinely needs one or only because the option happened to be sitting there. It lives in the willingness to leave some things unbuilt, to ask for less rather than more, and in the procurement and policy decisions that sit above any single developer’s craft. That is a less satisfying answer than a smaller number in a report, and I know it. But pretending that a tidy site settles the account is only the next version of the net-zero story we already told ourselves once. The belief I am really asking us to retire is the oldest comfort the cloud ever offered, that uploading a thing is the same as making it disappear.
I stopped believing it years ago, and yet the industry, and most of the people it serves, still runs on it, and AI has stretched the gap between that comfort and the truth too wide to paper over. The proof sits in a warm and humming building somewhere, in a grid and a water table and a wafer of silicon that someone you will never meet is paying for. Learning to see that plainly is the whole of the craft now, and in a sense it always was. The work was always larger than the bytes. It was, underneath everything, a refusal to look away from the true cost of what we build, and that refusal is needed at a scale none of us trained for, which is exactly why it matters more now, and not less.