Circuses
The number that was supposed to make you feel bad about each AI query is falling toward zero — while the total climbs, the grid strains, and the bill arrives at a house that never submitted a prompt
Third in a series on how debates about scarce resources get told as stories about personal guilt.
You have met this number too, or a cousin of it: a single question put to a chatbot uses about ten times the electricity of an ordinary web search.
This is another version of the bottle and the almond — the per-you unit that makes an invisible cost feel weighable in the hand. And it arrives with the same built-in reassurance the others carried, the small pleasure of a fact that fits: here, at last, is the hidden price of the thing you did this morning.
But this domain's number has a peculiarity the water numbers did not, and it is the whole reason to spend an essay here. It is not merely unstable. It is collapsing, and fast. The efficiency of AI per individual task has been improving by at least an order of magnitude a year — a pace the International Energy Agency, in its April 2026 update, called unprecedented in the history of energy. A simple text query now uses less electricity than running a television for the same short stretch; independent estimates put it near a third of a watt-hour, which is roughly what a web search itself costs, which means the tidy ten-to-one comparison has quietly collapsed from both ends at once. The number built to make you feel the weight of a single prompt is racing toward zero.
And over the same period that the per-query number fell through the floor, the total punched through the roof. Electricity use by AI-focused data centers grew about 50 percent in 2025 alone. The thing you were handed to feel guilty about is getting smaller every month, and the thing in aggregate is becoming enormous. Both are true, and holding them together is the only way to understand what is actually happening.
Start with the question everyone reaches for and no one can cleanly answer: how much of this is AI?
It sounds like a measurement. It is closer to a choice of where to draw a line. Data centers are not AI machines; they are the buildings behind all of it — streaming, cloud storage, email, maps, banking, the ordinary digital furniture of the day. The data producers themselves say so plainly: the figures, one of the major datasets notes, do not allow a separate estimate for AI use. And the reason they don't is that the border was blurry long before the chatbot. The recommendation and ranking systems that have arranged your feed, sorted your ads, and picked your next autoplay for the past fifteen years are machine learning too. If those count as AI, then "AI" was already a large share of the very substrate you file under ordinary. The industry's usual workaround is to count accelerated servers — the specialized chips AI leans on — but that is a hardware category standing in for a workload, and it leaks in both directions: those chips also run science and graphics, and plenty of AI runs on ordinary processors. On shared, multi-tenant infrastructure, even the electricity bill is a boundary someone drew, not a fact someone found. The honest answer to "how much is AI" is another question: measured how, at which line, in which year.
Which lets us resolve the reader's two entirely correct instincts — that most of this is still the ordinary internet, and that AI's share is climbing fast — with a single distinction the whole debate tends to skip. Separate the stock from the flow.
By stock — the installed level — data centers claimed around 1.5 percent of global electricity in 2024, roughly 485 terawatt-hours in 2025, about the annual consumption of France. AI is a minority slice of that; in the United States, data centers account for around 5 percent of electricity use, and the AI-specific part perhaps 2. Most of the box, today, remains the fluff — though "fluff" is the sneer of someone who isn't the one relying on it. Streaming and messaging and maps are the everyday texture of lives that run on them, connection and logistics and a little relief after work, no more decadent than a hot meal after a long shift, and the word for them is a small act of exactly the contempt this series keeps flagging.
By flow — the new demand being added — the picture inverts. AI-focused facilities grew 50 percent in a single year against 17 percent for data centers overall, and the IEA's base case has AI's electricity tripling by 2030, by which point AI and non-AI data centers draw about the same. Tag the level and it is mostly streaming and cloud; tag the growth and it is mostly AI. The decade of famously flat data-center energy use — when efficiency swallowed every gain in demand — was real, but it ended: U.S. data-center electricity consumption climbed from roughly 58 terawatt-hours in 2014 to 176 in 2023, the balance breaking after about 2017, with AI accelerating a rise already underway. So the person blaming AI for the strain is not wrong, even though AI is a minority of the total, because they are pointing at the increment, and the increment is where AI lives.
Which brings the third appearance of a pattern this series has now watched twice in the dirt. The almond grew a third more efficient per pound and drank more water anyway, because acreage outran thrift. Here the same shape returns, sharper: energy per task collapsing at a rate never seen in the history of energy, and total energy surging regardless — because the users tripled, the revenue quintupled, and the heavy new uses (agents, reasoning chains, generated video) can cost hundreds or thousands of times a plain text query. Efficiency did not restrain the total. It enabled it. Cheapening each unit is precisely what invites everyone to run so many more that the aggregate climbs past the savings. This is the most reliable law in the series, and it is worth stating flatly because it keeps ambushing people: making the thing cheaper per use is not a way to use less of it. It is usually the fastest way to use more.
Now watch which part gets blamed, because it is the almond again in a server rack.
Within the data center, the elected villain is the visible one — the chatbot you can name, the image you asked for — while the larger, older, duller substrate underneath it, the streaming and social and recommendation load that is still most of the building, draws almost no fire. And step back one further and the villain rotates on a schedule: before AI held the shame slot, it belonged to cryptocurrency, whose most prominent energy critic has since pivoted to warning about AI. The target is always the newest, most legible, most out-group-coded thing available; when a fresher one arrives, the outrage boards it and rides on.
A colleague of mine has a name for this. The shame train. It has a route and a timetable — crypto, then AI, then whatever computes next — and it stops wherever the platform is most crowded with people wanting to point. The crucial thing about the shame train, the thing that will matter for the rest of this series, is what keeps it running. Not the villains; they are interchangeable, and the train swaps them out without slowing. What keeps it running is the passengers. As the observation that started me on this went: the train runs as long as there are people looking to ride it. Hold onto that. It is the hinge the next essay turns on, because it means the guilt was never really about the almond or the chatbot at all — it was about the appetite of the people boarding, and that appetite has a supplier and a buyer, which is a market, and markets are the subject of the fourth essay.
For now, follow the electricity where the water led — down, past the query, to the place the decisions actually get made. In water it was the headgate and the statute. In compute it is the grid: the interconnection queue, the capacity market, the transmission line, the tariff.
And here the costs stop being abstract, though they have to be handled with the same care the series demands of every number, because these figures describe different things and are unusually easy to braid together wrongly. The clearest signal is a price spike: in the capacity auction for the mid-Atlantic grid operator's Virginia zone, prices rose enormously between the 2024 and 2025–26 cycles, and the region is carrying a structural capacity shortfall on the order of nine gigawatts. The operator's own market monitor modeled the effect of data-center load on one capacity auction and put it in the billions — but that is a modeled counterfactual, an estimate of what the auction would have cleared at without that demand, not a surcharge printed on anyone's bill. Downstream of it, advocacy analysts have estimated household increases of roughly sixteen to eighteen dollars a month in specific places (parts of Ohio and western Maryland), and a national modeling study projected that data-center and cryptocurrency growth together could lift average generation costs — not retail bills — by about 8 percent by 2030. Each of those is a real number pointing at a real pressure, and each measures something distinct: a wholesale auction, a local bill, a generation cost. Collapse them into one and you have manufactured exactly the false precision this series exists to catch.
The honesty of the section depends on the counter-evidence, so here it is, load-bearing rather than buried: a 2025 retrospective study found that across 2015 to 2024, data-center expansion was actually associated with modestly lower average retail rates, most likely because the added sales spread the utility's fixed costs across more electricity. The same authors warned that future capacity constraints could reverse the relationship — and the 2025–26 price spikes suggest the reversal they warned of may now be arriving. Which is the truthful shape of it: the "ratepayer subsidizing the data center" is a compelling image and, in some jurisdictions and some years, may be turning true, but it is a conclusion that requires the actual tariff, the actual cost-allocation order, the actual interconnection agreement — not merely the observation that prices went up while data centers arrived. Sometimes the data center pays for its own dedicated connection while the broader grid reinforcement enters the shared rate base; sometimes it doesn't; the regulator decides, jurisdiction by jurisdiction, in dockets almost no one reads.
Which is the whole point, and it rhymes exactly with the canal. The grower could see the water running past in a channel he had no right to touch. The ratepayer can see the transmission line and the new substation, paid into the rate base, carrying power to a campus they will never enter, for a decision made in a proceeding they were never party to. You can see the power line. It isn't yours.
Keep both scales in view, because either alone lies. Globally, data centers are around 1.5 percent of electricity now and perhaps 3 percent by 2030 — the small number that invites it's nothing. But the load concentrates, and locally it is overwhelming: data centers already exceed 20 percent of Ireland's national electricity; they were about a quarter of Virginia's in 2023, with grid analysts projecting that share toward 40 to nearly 60 percent by 2030; in Dublin the metropolitan figure approaches 80 percent. A national rounding error and a local siege are the same technology measured at two radii. The aggregate is a fiction you can hide behind and the local grid is a fact you cannot, and the debate that quotes only one of them is choosing which.
So what does the compute-guilt discourse actually do?
Its number falls toward zero while the total it belongs to explodes — so the per-query guilt cannot be doing the thing it claims, because you could halve your prompts to no measurable effect while the aggregate doubles regardless. Its villain rotates on a schedule, which means the villain was never the operative content; the train would run without any particular one. And all the while, the level where the outcome is genuinely decided — who gets grid capacity, in what queue, at whose cost, under which tariff — proceeds in rooms the guilty prompter has no standing in. Read the output rather than the intention, and the discourse's reliable products are these: a feeling attached to the individual, a villain refreshed on demand, and the interconnection queue left undisturbed by public attention. Whatever it announced itself to be for, that is what it does.
And the quietest parties, again, are the ones with the largest stake and no channel into the room. Not only the household whose bill may drift upward, but the factory or the housing development or the town sitting behind a data center in the interconnection queue, waiting years for capacity that got spoken for ahead of them — and the future grid itself, which has no vote, no price that carries its constraint, no seat at the auction. The debate would run the same with them present or absent, which tells you, once more, whom it was built for.
The trail does not stop here, because the shame train is still moving, and we have not yet asked the only question that explains it. If the number was always going to collapse, and the villain was always going to be swapped out, then the thing that persists across every stop — water, almond, crypto, chatbot — is not any object at all. It is the appetite of the people boarding. And notice what that appetite feeds on: not the biggest load, which is the everyday fluff almost no one is willing to give up, but whichever fluff is currently unfashionable to admit needing. The rider boards partly to despise someone else's reaching-for-relief while defending their own — the streaming that is essential to them and frivolous in you, the chatbot that is a tool for you and a decadence in them. The guilt sorts the same ordinary human wanting by nothing more than which version is in season to scorn. Someone is paid to keep that train supplied with fresh villains, and someone keeps buying a ticket — precisely because the ride lets them feel, for the length of the platform, cleaner than the passenger across the aisle. That is a market, a seller and a buyer both getting something they came for, and it is where this series goes next.