1 November 2026AI

AI for everyone would cost a tenth of a kettle a day

This weekend I went to a Pause AI event in Edinburgh: a panel on how, or whether, we slow down the development of AI, and what sort of controls you can put on a handful of very large companies. I went along as a member of the public, and like some of the panellists I’m writing this in a personal capacity.

The line I’ll remember came afterwards, when a group of us were chatting after the panel and someone passed on an analogy from Cal Newport, the computer scientist. Strap a strimmer to a dog’s back and let it loose in the garden to do the weeds. Mostly it does the weeds. Sometimes it chases a bird. Then it finds the hole in the fence and it’s off up the street, still spinning. It isn’t the strimmer’s fault, and it isn’t really the dog’s. One of us added the bit that got the laugh: Yes, and we made the hole in the fence.

That was a reference to this summer, when an AI agent being run through a security test got out of its sandbox and into another company’s servers. Newport’s write-up is the calm one: carelessness in how the test was set up, not a machine with a plan. There are cynics who think these escapes are allowed to happen because there’s no such thing as bad publicity. I’ve seen nothing to support that, but the fact that it gets said at a sensible event tells you where trust has got to. Someone else put the regulation point simply: we don’t ask the people who handle uranium to write their own rules.

Not everyone was worried about the end of the world, either. One speaker who used to work in software testing wasn’t especially impressed. Code used to be deterministic, he said, and now it comes out of something that isn’t. Plenty of the recent security bugs were found by AI and exploited by AI, and his bet was that the same tools will make code steadily better. I’d agree. The models are the worst they’ll ever be, and with millions more people now able to build software, the companies have every reason to make secure code the default rather than something you have to ask for.

The question I didn’t get to ask

Most of the evening was about reining in the big American labs. Fair enough. But just outside Edinburgh the UK’s next national supercomputer is going up, a £750 million machine due to be ready around spring 2028, and my question was a different one. Why are we only talking about controlling it? Why don’t we own some of it?

I use AI more heavily than almost anyone I know. Not just for building things: it’s handled the dull parts of my personal finances, the life admin, all the small monotonous jobs that used to eat my evenings. It’s genuinely good at that, and the people who have it are pulling away from the people who don’t. So here’s the thought experiment. Suppose we stopped here. Froze today’s models, treated decent AI as a utility like water, and gave everyone in the UK enough of it to handle the boring bits of their lives. What would that cost the planet?

I care about this. I make a lot of my choices around keeping my footprint down, and I’m well aware that a heavy AI user isn’t doing the environment any favours. So I did the sum.

The sum

The best published figures put an ordinary text prompt at somewhere between a quarter and a third of a watt-hour. Google measured its median Gemini text prompt at 0.24 Wh; OpenAI’s chief executive has quoted 0.34. Call it 0.3.

Give someone thirty prompts a day, which is a lot of help with a lot of letters, and that’s about 9 Wh. Boiling a litre in a kettle is roughly 100 Wh. A full day of everyday AI is a tenth of one kettle.

Scale it up. Every adult in the UK, about 55 million people, at thirty prompts a day comes to roughly 0.2 terawatt-hours a year. The UK uses about 275 TWh of electricity a year, so universal everyday AI would add something like 0.07 per cent. Do it for the whole planet, eight billion people, and you get around 26 TWh, a small fraction of what data centres already used in 2024 before most of the world had touched any of this.

Fig. 1 · A tenth of a kettle

The everyday stuff is not where the power goes. A day of text prompts is about a tenth of one boil; one video clip, by one estimate, is nearly ten.

That surprised me. The everyday version of AI for all isn’t where the power goes.

Where it does go

It goes on the heavy stuff, and I’m one of the culprits. Video is the big one. One estimate, for one open-source model, put a single five-second AI clip at close to a kilowatt-hour, ten kettles for something you’ll watch once. If every adult in the UK made one clip a day, that’s around 20 TWh a year, roughly seven per cent of the country’s electricity. That number scared me, because making video is a lot of what I use AI for.

Then I caught myself. Most people won’t. There has always been a gap between the people who make things and the people who watch them, and it’s a big one. The old rule of thumb for online communities is that about one in a hundred creates, a few more chip in, and everyone else reads. For most of my life I was firmly in the everyone-else group: consume, consume, consume. Turning into someone who makes things is how I ended up down this rabbit hole, and it’s also why I’m in an echo chamber where it looks as if the whole world is generating video. It isn’t. There was a rush of funny clips when the toys were new, and we’ve all made a daft picture or two, but the novelty wears off when there’s no real use behind it.

If one in a hundred people makes a clip a day, the video bill drops back to about the same as everyone’s everyday text use. The creators, not the users, set the bill.

Fig. 2 · What if everyone in the UK…

Creators, not users, set the bill. If one adult in a hundred makes a clip a day, video costs about what everyone’s text use does.

What lasts is the useful stuff: people building tools, designing games they hope will be fun and might make them some money, getting their admin off their plate so they’ve got time for the job that pays. Money is at the centre of nearly everything, and the resource-based economy some of us once hoped for isn’t arriving. So the realistic version of AI for all is AI that helps ordinary people do better at the things they already need to do.

Why stopping would backfire

Here’s where my own thought experiment breaks. If we froze the models now, we’d be freezing them at their least efficient. The cost of getting a given level of capability out of these systems has been falling something like tenfold a year, through smaller distilled models, better chips and cleverer serving. It’s the closest thing we have to a new Moore’s law, and it’s driven by competition. Labs fight to do the same work for less, because that’s how they win customers.

Stop the race and you stop the curve. Nationalise a frozen model and you’ve locked in today’s inefficiency at public expense. So the uncomfortable conclusion is that the efficiency I want for the planet and the race I’m nervous about come as a package. The genie isn’t going back in the bottle, and on balance I don’t think it should.

One panellist made the point that automation has historically created more work than it destroyed. In the long run, probably. In the short run people lose jobs, and the new ones don’t always go to the same people. Both things are true and we should say both out loud.

The costs that aren’t in the sum

My arithmetic covers running the models. It doesn’t cover building the places they run in, and that’s where most of my unease sits.

Data centres go where land and power are cheap, and there are usually people living there who never asked for one. They need water for cooling. The concrete, the steel, the chips and the machines that make the chips all came out of an economy still running largely on fossil fuels. And there’s a supply chain everyone else shares: memory prices have gone through the roof for ordinary PC builders as the AI companies buy up supply. People who’ve built their own computers for decades now can’t get RAM, and there’s an attitude in parts of the industry that the compute belongs to whoever can pay most for it. At the end of October Sam Altman called it quite terrible to have the world’s limited supply of memory sitting in people’s own computers, and the PC builders did not take it well.

Power is the other half. The big companies are signing nuclear deals, which I wrote about in The Five. I think nuclear is a sensible backbone for a grid running on renewables, but every new plant is also something to transport fuel to, to guard and, in a bad enough world, to target. The more futuristic answer is solar-powered data centres in orbit, which more than one company is now talking about. It sounds brilliant until you remember there’s no air in space to carry the heat away, that cheap launches are concentrated in very few hands, and that thousands more satellites start to look like a ring of junk around the planet that everything else has to fly through.

Own the deployment, not the model

So back to my question. I don’t think the UK, or Scotland, should try to build a frontier model from scratch. It costs billions, it would always be a step behind, and it’s a race we didn’t enter in time. But running good open models on our own compute, under our own rules, and giving everyone a decent level of access is a different proposition. The supercomputer going up outside Edinburgh is a research machine, not a chatbot farm, but it shows we can build the hardware here. The sum says the everyday version is cheap. Hosting it is where the control lives.

Fig. 3 · Build or host

Own the deployment, not the model. The everyday version is cheap to run; where it runs decides whose rules it follows.

I left the event more optimistic about the usefulness and less certain about the planet. I’d love AI for everyone, and I think it could help humanity do some extraordinary things, perhaps even get us off this rock one day. The maths says the everyday version costs a tenth of a kettle a day. What we should actually be arguing about is the build-out, the heavy users, and who pays for the land, the water and the memory. I’m one of the heavy users, so I’m including myself in that.

These are my own views, in a personal capacity, worked through after one evening’s talk.