The AI You're Paying For Is Cheaper Than It Should Be. That Won't Last.

Let me start with a confession. I pay for a lot of AI.

I’ve got a Claude Team plan. Gemini through my Google Workspace. ChatGPT on the premium tier, plus its coding tool. Add it up and it’s a respectable monthly bill, the kind of thing you’d expect a product person who lives in these tools to spend.

Here’s what bothers me. I’m almost certainly not paying what it actually costs.

The all-you-can-eat problem

Think about the last time a new restaurant opened in your neighbourhood with a wild launch deal. Forty percent off. Free dessert. A second main on the house. You went, obviously. The food was great and somehow cheaper than cooking at home.

You also knew, somewhere in the back of your mind, that this could not go on forever. No kitchen survives by selling meals below what the ingredients cost. The deal is a hook. It gets you in the door, gets you used to the place, gets you telling friends. Then, quietly, the prices drift back up to what they were always going to be.

That is roughly where we are with frontier AI right now. The big providers are in a land-grab. They are spending enormous sums on the computer chips, electricity, and engineers needed to answer your questions, and they are charging you a fraction of that to get you hooked while they fight for market share.

I am not going to throw made-up numbers at you, because the truth is none of us outside these companies knows the exact unit economics. But the direction is not really in dispute. The infrastructure is staggeringly expensive, the subscriptions are priced to win customers rather than to turn a profit, and that combination has only one ending.

Why this matters to you, even if you’re not technical

You might be thinking: fine, so prices go up a bit. I’ll deal with it then.

The trouble is what happens in the meantime. These tools are quietly becoming load-bearing. People draft their emails in them, plan their weeks, summarise their documents, talk through decisions. The more useful they get, the more we lean on them, and the harder it is to walk away when the launch deal ends and the real menu arrives.

I have watched this movie before in other corners of technology. The cloud storage that was free until it wasn’t. The ride-hailing that was absurdly cheap until the subsidies dried up. The pattern is always the same. Get people dependent while it’s cheap, then reprice once leaving is painful.

I’m not saying the frontier tools are a trap to avoid. They’re genuinely brilliant and I’ll keep using them. What I’m saying is that depending on something you don’t understand and don’t control is a weak position, and it’s worth doing something about while the doing is easy.

The thing almost nobody mentions

Here’s the part that surprised me. You don’t actually need the frontier models for everything.

There is a whole world of capable AI you can run on a computer you might already own. No subscription. No monthly bill. No usage meter ticking in the background. The model lives on your machine, answers to you alone, and costs you nothing but the electricity to run it.

These are called open models, and a year ago I’d have told you they were a hobbyist’s toy. That’s no longer true. The good ones are now genuinely useful for a large slice of everyday work, and they’re getting better at a pace that’s frankly hard to keep up with.

I spent a few weekends getting one running on an ordinary gaming PC. Not a data centre. Not a five-figure rig. A regular desktop with a couple of graphics cards in it. And it works.

What this series is about

Over the next four posts, I’m going to walk you through what I learned, written for someone who is smart and curious but not technical. No jargon without a plain-English translation. Plenty of analogies, because that’s how I actually think about this stuff.

Here’s the plan:

• Next up: what “local AI” even means, and why running a model on your own computer is different from using one in the cloud.
• Then: the why. How giving your model a job turns it from a private chatbot into a genuine assistant that remembers you and works for you.
• After that: how to figure out which model fits your particular computer (including the Mac-vs-PC question), and the surprising trade-off I found between a faster model and a smarter one.
• Then: the actual weekend build. How to get one running without a computer science degree.
• Finally: what these local models are genuinely good for, where they fall short, and why I think learning them now is the smart move.

The whole thing is a weekend-project diary, not a manifesto. I’m doing this because it’s fun and because I think the people who understand the alternatives will be in a much stronger position when the launch deals end.

The restaurant is still running its opening special. That’s exactly why this is the right moment to learn to cook.