No. 17AI Technology

If chat is a tenth of a kettle, where is the power going?

When I put the kettle piece on Hacker News, one commenter asked exactly the right question. If everyday AI costs a tenth of a kettle a day, why is the grid being asked to connect data centres on a scale that rivals the whole country’s peak demand? They’d done their homework, quoting Ofgem’s own figures, and they were right that my sum says very little about the total.

So I went and looked. The kettle sum holds up. It’s also beside the point, and the people who plan the world’s electricity have noticed the same gap.

The IEA did the same sum

In April the International Energy Agency ran almost exactly my arithmetic. It took the published figures for a text prompt, rounded them up to a generous 1 Wh to cover heavier work, and imagined 10 billion queries a day, about as many as every web search on Earth. That comes to about 3.6 TWh a year, less than 1% of the 485 TWh that data centres consume today. Its conclusion is the question I’m asking here: the bulk of the capacity being planned must be destined for other workloads.

It also noted that none of the companies publishing per-prompt figures has said how its planned capacity splits between training, video, agents and everything else. That was the commenter’s other point, about why anyone should trust a company’s own white paper, and it’s a fair one. The per-prompt numbers I used are self-reported, but they aren’t the only ones. The IEA also cites a bottom-up estimate from Microsoft Research of 0.34 Wh for the biggest models, and Hugging Face’s independent measurements of open models land in the same range. The breakdown that would let anyone check the total isn’t published at all.

How big the total is

Data centres used about 485 TWh in 2025, slightly more than 1.5 per cent of the world’s electricity and 17 per cent more than the year before. The ones built for AI grew faster, by half, to something like 155 TWh. The IEA expects the total to roughly double to 950 TWh by 2030.

That’s big, but it isn’t most of the story of rising demand. The world’s electricity use is climbing faster than it has for a decade anyway, on industry, air conditioning and electric cars, and data centres are slightly under a tenth of the growth to 2030. The strain is concentrated in America, where data centres are about half of all the growth, and the US and China between them take nearly 80 per cent of the data centre increase.

Fig. 1 · Where the new demand comes from

Data centres matter, but they aren’t most of it. Industry, buildings, transport and cooling add far more to the world’s demand by 2030.

A queue is not a load

The UK numbers the commenter quoted are real. Ofgem’s July consultation counts about 73 GW of data centres waiting to connect, against a national peak demand of 45 GW. If all of that ran flat out, it would use more electricity than the rest of the country put together.

It won’t, and Ofgem says so. A connection queue is a list of requests for capacity, not a forecast of use. In February Ofgem said it expected a significant number of the projects to be non-viable. NESO has since heard from the industry that close to half the pipeline may be speculative or duplicated, and Ofgem found at least 9 GW that had joined the queue as batteries and then relabelled themselves data centres. The IEA’s reading of American utility data is that only about a fifth of data centre requests get built in the short to medium term.

What’s actually connected in Great Britain is about 2.4 GW, by NESO’s estimate. DESNZ’s figure for 2024 was 4.5 TWh, 2 per cent of the electricity drawn from the grid, which is an average of about half a gigawatt.

None of that makes the build-out small. The government forecasts it will need at least 6 GW of AI-capable capacity by 2030, and NESO’s scenarios have data centres using 13 to 25 TWh by then, roughly three to five times the 2024 figure and 4 to 7 per cent of British demand. That’s a real planning problem. But “the grid needs half as much again” is a reading of the queue, not of the meter.

Fig. 2 · A queue is not a load

The queue is requests, the meter is use. Britain’s data centres drew about half a gigawatt on average in 2024, against 73 GW asking to connect.

Where it goes instead

So if chat is a sliver, what isn’t?

A plain text answer is the cheapest thing these systems do. Hugging Face measured the same models with reasoning switched off and on. One 70B model went from 0.05 Wh to 7.6 Wh, about 150 times as much, and smaller ones by up to 700 times, because the reasoning version writes thousands of words of working before it answers. An agent that plans and runs a multi-step task with reasoning comes to around 50 Wh by the IEA’s rough figures, half a kettle. Video is the far end: one five-second clip from one open model measured close to a kilowatt-hour.

Run the IEA’s sum again with that agent figure. Ten billion requests a day at 50 Wh rather than 1 Wh comes to about 180 TWh a year, more than a third of what every data centre on Earth used in 2025, from one kind of work. A single frontier training run, a few hundred gigawatt-hours, is small beside that. The catch, which comes up below, is that the runs that ship are only a fraction of what the labs spend on research.

Fig. 3 · A tenth of a kettle, and the rest

Chat is the bottom rung. Reasoning and agents climb by orders of magnitude, and a video clip tops the ladder at nearly ten kettles.

Volume is the next multiplier. Google processed 9.7 trillion tokens a month across its products in May 2024. In May 2026 it was over 3.2 quadrillion, more than three hundred times as many in two years, some of it the AI summaries that now sit on top of search results. And for every word a person reads, a reasoning model or an agent may have written many more that nobody reads.

The oldest big AI workload isn’t chat at all. Before anyone had heard of ChatGPT, recommendation systems, deciding which post, advert or video you see next, were over 79 per cent of the AI inference in what was then Facebook’s data centres, in 2019. That work didn’t stop when chat arrived.

Town-sized training runs

Then there’s building the models. GPT-3, in 2020, took about 1,300 MWh to train. Grok 4, five years later, took around 310 GWh by Epoch AI’s estimate, roughly a year’s electricity for 124,000 British homes, and the next generation is bigger again.

My hunch going in was that a lot of this is waste: dozens of companies racing to build the same thing in secret, each burning energy so that one of them can win. The evidence half agrees. In 2025, Epoch’s database has more than thirty organisations training models with at least 10²⁴ operations, about a twentieth of GPT-4. Four American labs shipped their flagship models within a fortnight of each other in November 2025. Mark Zuckerberg said last year that a bubble was quite possible, but that misspending a couple of hundred billion dollars would be better than being late. Economists have been predicting this since the 1970s: when firms race for a single prize, each puts in more than is good for everyone, and much of it is duplication.

The hunch had the place wrong. The headline runs are a minority of the spend: Epoch’s analysis of OpenAI, MiniMax and Z.ai puts the final runs of released models at about 10 to 23 per cent of their research compute. The duplication is mostly in the experiments, the many trial runs that never ship. It’s also a dozen or so rich players, not a hundred: only two labs have trained anything Epoch puts above 10²⁶.

And secrets don’t keep. The best open models trail the best closed ones by about four months, and in February Anthropic reported more than 16 million exchanges with its model taken through fraudulent accounts to train rivals. Some of the losers do stop: in 2024 Inflection, Adept and Character.AI each handed their founders, most of their staff and a licence to their models to Microsoft, Amazon or Google, rather than keep training.

Fig. 4 · Town-sized training runs

Frontier runs are town-sized; science is a street. Grok 4 took about three thousand times what AlphaFold 3 did.

Science is the cheap bit

The uses I’d most like to defend are the smallest. AlphaFold 3, which predicts how proteins and other molecules fit together, took roughly 0.1 GWh to train, working from Epoch’s figures: about a three-thousandth of Grok 4. ECMWF’s AI weather forecast uses about 1,000 times less energy than the physics model it runs alongside, though at a coarser resolution and by its own account. Science is a rounding error on the bill.

Efficiency doesn’t save us

Google says it cut the energy of its median text prompt 33-fold in a year, a figure it notes hasn’t been independently verified. Its total electricity use rose 27 per cent in 2024 and 37 per cent in 2025, and its data centres alone used 42 TWh last year, more than twice what they used in 2022. Making each answer cheaper made more answers worth giving. William Jevons noticed the same thing in 1865, when Britain burnt more coal, not less, as its steam engines got more efficient.

Energy is power

Who gets to run this race is decided by who can pay for the power. The four biggest American tech firms are guiding to something like $700 billion of capital spending this year between them. Big tech has signed nuclear deals for about 10 GW, from Microsoft’s restart of Three Mile Island to Meta’s 6.6 GW in January, with more in targets. And price decides where it goes: in April OpenAI paused its UK Stargate project, citing energy costs and regulation.

So the commenter was right, and the kettle sum still stands. Your own chat costs next to nothing. What costs a lot is everything around it: reasoning and agents doing more work per request, a few hundred times more requests, video, the recommendation engines that were here first, and a dozen rich companies each paying for their own route to the same frontier, mostly in experiments nobody sees. The IEA is asking the companies for the breakdown, and so should we. If a company can publish what one prompt costs, it can publish what its gigawatts are for.