29 October 2026AI Self-hosting

The Five: your personal AI will be rented, and that is the point

In Alien: Earth, the year is 2120 and the planet is run by five corporations. No governments worth mentioning, just Prodigy, Weyland-Yutani and three others whose names you only learn from a map. And each of them runs on synthetic staff: androids crewing the ships, androids in the labs, and in Prodigy’s case something stranger, human minds moved into manufactured bodies. These aren’t companies that use AI. They’re made of it. I watched it this year and the part that stayed with me wasn’t the alien. It was how plausible the map looked.

I don’t think it takes a century. Here is the prediction, and I’ll show my working: within a few years, nearly everyone will have a personal AI that is connected to everything they own and do, they will talk to it rather than type, and they will rent it from one of a handful of companies. Not build it, not run it. Rent it, the way you rent Netflix or the gym.

Most of the argument about AI at the moment is about what the models can do. I think the more interesting question is who you’ll be paying for yours, and why the answer is going to be such a short list.

What the thing is

Start with the shape of it, because it’s already visible if you use these tools heavily. I do, mostly out walking, which is how my last book got written, and a normal week now looks like this. I tell it to find the file from a fortnight ago and send it on. I ask whether there’s anything in my personal inbox I haven’t answered. The dentist wants to move an appointment, so I say “find a slot next week that doesn’t clash” and the reply gets written. None of that involves opening a document. It is life admin, and life admin is most of what a person’s inbox actually contains.

The pattern has three parts. The AI is connected: mail, calendar, files, photos, messages, the smart home, the watch. You talk to it in ordinary language, and it understands what you mean well enough that most of the time you don’t check the output. And it knows you. Not in a creepy way, in the way a good assistant of ten years knows you: your tone, who matters, what you’d normally say.

Put those together and you have something that is not a chatbot. It’s closer to a second person running your affairs. And that is the thing nobody is going to host in a cupboard.

Why you’ll rent it

The hardware

The hopeful version of the future is that models keep getting smaller and cheaper until a good one runs on your phone and the big companies lose their grip. Half of that is true. The models on phones today are real and useful, and they’ll get better. But look at what they are. The on-device models that ship with flagship phones run to a few billion parameters and need the top-end memory to do even that. The frontier models are hundreds of times larger and run on racks of chips that each draw more power than your whole flat.

Both of those improve every year. That’s the point people miss. The gap between the phone and the warehouse doesn’t close, because both ends are moving. The phone of 2030 will run something that would amaze you today, and the warehouse of 2030 will run something that makes it look like a pocket calculator. Whatever the best model is, it will live where there’s the most compute, and the most compute is never in your hand.

The moon-landing comparison gets made a lot: the Apollo guidance computer had a few kilobytes of memory and your phone has millions of times more. True, and irrelevant, because the question was never whether your phone beats 1969. It’s whether your phone beats a million of itself wired together in a shed in Louisiana. It doesn’t, and no amount of progress changes that, because the shed gets the progress too.

The energy

Compute is two costs: the hardware and the electricity to run it. Everything else is detail, and the second cost is the one that’s about to matter most.

The International Energy Agency’s Energy and AI report, published in April 2025, projected that electricity demand from data centres worldwide would more than double by 2030 to about 945 terawatt-hours, which is more than Japan uses in a year. Its April 2026 follow-up found data centre demand had jumped 17 per cent in 2025, more than five times faster than electricity demand overall. The IEA’s director put it in one line: there is no AI without energy.

Meanwhile the four biggest cloud companies have 2026 capital spending plans of roughly 700 billion dollars between them, most of it on data centres and the chips inside them. Several have signed deals for nuclear power, in one case restarting a plant that had closed.

Nobody’s homebrew AI gets its own reactor.
The energy

You and I are not doing that. Nobody’s homebrew AI gets its own reactor. Cheap energy means cheap compute, cheap compute means a better model at the same price, and a company that can sign a twenty-year power contract will always be able to sell you more intelligence per pound than you can generate yourself. Energy has decided who’s ahead for most of industrial history. It’s about to again.

The security job

The third reason is the one I’d push hardest, because I’m a developer and I know what running your own server actually involves.

Think about what a personal AI holds. Your mail, back years. Your calendar, which says where you’ll be. Your files, your photos, your messages. If you’ve connected your watch, your heart rate while you sleep. That is the most sensitive database that will ever exist about you, and the question is who you trust to keep it.

We already answered this for everything else. Almost nobody hosts their own email. Some people do run a home server for photos, and I don’t want to annoy them, but they’re a small technical minority who are also prepared to do the off-site backups, which is the hard part. Everyone else went to iCloud or Google Photos the moment that was an option, because the backup comes with the service and the alternative was a drive that would fail. Websites went to the cloud. We outsource hosting to people who do it full time, with security teams and audits, and we were right to.

Running your own AI with your whole life attached is the same job with higher stakes and a higher skill bar. The people who should do it are the people who can patch a box at two in the morning and know what a certificate expiry looks like. That’s a hobby, not a plan for a population. Everyone else will pay someone, and they’ll pay whoever has the best model, which brings us back to the first two reasons.

Fig. 1 · The stack

You talk to the top; the money and the electricity are at the bottom.

So that’s the shape. You talk to something that feels like yours. Underneath it is a model you don’t own, memory you don’t hold, connections into every account you have, and at the bottom a warehouse full of chips and a power station. The handful of companies that own the bottom of that stack are the Five. Today I’d say Anthropic, OpenAI and Google hold three of the seats, with Microsoft, Meta, Apple and the Chinese labs arguing over the rest. The names may change. The number won’t get much bigger.

The reversal

Here is the part I find funniest, as someone who spent years telling people to care about their data.

A few years ago the whole culture was refusal. Cookie banners everywhere after 2018, and the smart thing was to reject all. Ad blockers, tracker blockers, the browser that promised not to follow you. I said this at a tech event last year and the room laughed because everyone recognised it: the same people who wouldn’t accept a cookie are now sitting in front of a screen that says “Connect your calendar?” and clicking yes before it finishes asking. Mail, yes. Files, yes. Photos, yes. Connect everything.

Fig. 2 · From Reject All to Connect Everything

Nobody changed their mind about privacy.

Nobody changed their mind about privacy. What changed is that the utility became visible. The research on this is old and solid. Susan Barnes named the “privacy paradox” in 2006: people say they care about privacy and then behave as though they don’t. Norberg and colleagues measured it directly the year after, and the gap between stated concern and actual disclosure was wide. Pew found in 2019 that four in five Americans felt they had little or no control over the data companies collected on them, and they kept using the services anyway. Cookies gave us nothing we could feel; the ad would have been an ad either way. The AI gives us back an hour a day, and for that we’ll hand over the lot.

I’m not saying that’s wrong. I’ve done it. The more I connect, the more useful it gets, and the trade is better than any the ad industry ever offered. I’m saying it’s a bigger bet than it feels like when you click, because the profile it builds is total, and it lives with one company.

What it costs

Two things.

The first is money, and the shape of the pricing is already set. The consumer tiers go free, then a monthly fee, then a top tier at two hundred dollars or more a month that buys the best model and the most of it. That’s the gym and Netflix model exactly, and it means what it means at the gym: those who pay more get more. The difference is that what you’re buying here is capability. The person on the top tier can produce in an afternoon what the person on the free tier can’t produce at all, and the gap compounds, because the output of one week is the input to the next. I’d expect that to become one of the sharper divides of the next decade, and it won’t be between people who have AI and people who don’t. It’ll be between tiers.

The second is risk, and it’s the mirror of the reversal above. The profile that makes the AI useful is the profile that makes a breach catastrophic. Someone who gets into your personal AI doesn’t get a password; they get you. Your voice can be cloned from a short clip, your face from a photo, and now your history, your tone and your relationships from the thing that was helping you answer the dentist. That deserves its own piece and I’m writing it. For now: the better the assistant, the worse the breach.

Big beats small

I want to end on the structural point, because I think it’s the one that holds whatever the model names turn out to be.

Fig. 3 · Both ends move

The phone gets better. So does the warehouse.

We have got very good at running things in parallel, and modern AI is the clearest case: the big models are big because they’re spread across thousands of chips working together, and the way you make a better one is mostly to spread it across more. So the best model is always the one with the most hardware and the most electricity behind it, and the most hardware and electricity are always in a warehouse owned by someone with a balance sheet the size of a country. Your phone improves. So does the warehouse. The warehouse improves faster, because that’s where the seven hundred billion went.

So: a personal AI each, connected to everything, spoken to rather than typed at, and rented from one of a short list of companies who own the bottom of the stack. I can’t see a version of the next ten years that doesn’t look like that, and I’ve tried.

What I would watch

The last two seats. Three are fairly settled. Who takes the others, and whether a Chinese lab is one of them, decides what the map looks like outside America.
Whether anyone regulates the memory. The profile is the product. Rules about portability, taking your AI’s memory of you to another provider, would change the whole picture, the way number portability changed phones.
The top tier’s price. If it goes up, the divide I’m describing is real. If it comes down, I’m partly wrong and glad.
A hybrid that actually works. Local model for the private stuff, rented model for the hard stuff, with a clean boundary between them. If someone ships that well, it’s the best outcome and the least likely.