There’s a lot of excited arithmetic going around about artificial intelligence. A trillion-dollar valuation here, a hundred-billion-dollar funding round there, the price of a model quoted like the budget of a moon mission. I’ve been writing this column long enough — since the Reagan administration, if you want to make me feel old about it — to have learned one durable thing about computing: the number everybody is staring at is almost never where the money ends up.

Let me tell you about two companies that figured that out early.

The first is ARM. If you’re reading this on a phone, there’s an ARM design inside it. There’s one in the tablet on your nightstand, the car in your driveway, probably the watch on your wrist. ARM is in nearly every smartphone on the planet. And here’s the part that ought to stop you cold: ARM has never manufactured a single chip. Not one. They don’t own a factory. They design an instruction set — the basic grammar a processor uses to think — and they license it. Everybody else does the expensive, dangerous, capital-soaked work of actually building the silicon. ARM collects a royalty measured in pennies on each chip, and those pennies, multiplied across the entire industry, have made a company that builds nothing worth more than many of the companies that actually run the factories.

The second is Dolby. Ray Dolby could have spent his career building the world’s best noise-reduction box and selling it to recording studios, and he’d have done fine. Instead he did something smarter. He turned his method for cleaner sound into a standard, and then he made that standard the thing every other manufacturer wanted on the box. You never bought a Dolby. You bought a tape deck, or a receiver, or a movie ticket, and somewhere on it was a little double-D logo that meant somebody had paid Ray Dolby for the privilege of sounding good. He didn’t win the audio market. He put a toll booth on it.

I bring up these two old stories because I think we’re about to watch the same trick get played in AI, and almost nobody is positioned for it.

Here’s the uncomfortable truth underneath all that trillion-dollar arithmetic: intelligence is on its way to becoming a commodity. The models are converging. The thing that felt like magic two years ago is now something you can rent by the token from half a dozen vendors, and the price only goes one direction. When you can swap one model for another the way you swap one brand of gasoline for another, you’re no longer looking at a miracle. You’re looking at a utility. And utilities don’t command trillion-dollar margins for long.

So, if intelligence gets cheap, where does the value go? It goes, as it always does, to whatever stays scarce.

And the scarce thing in AI isn’t intelligence. It’s trust.

We’ve built machines that are fluent and confident and, a meaningful fraction of the time, wrong — wrong in the specific, expensive way that sounds exactly as authoritative as being right. I’ve spent a good part of this series on that problem, because it is the problem. A model that is brilliant and occasionally invents things is not something you can put in front of an insurance adjuster, a radiologist, a loan officer, or a judge. The intelligence is already good enough. What’s missing is the guarantee.

Now here’s the interesting thing about a guarantee like that. Once somebody works out how to verify an AI’s answer — to know, before it ever reaches a human, whether the system is standing on solid ground or making it up — that capability doesn’t behave like a product. It behaves like a standard.

Think about why. You don’t want trustworthy AI in one app. You want it everywhere the AI is, the way you want the brakes to work in every car and not just the one in the showroom. A verification layer is only worth something if it sits underneath everything — and the fastest way to get underneath everything is not to build a rival to every AI company on earth. It’s to license the one thing they all need and let them keep competing on everything else.

That’s ARM. That’s Dolby. You don’t sell the trustworthy machine. You license the part that makes the machine trustworthy, and you collect a sliver every time it runs.

It’s an odd business to explain at a cocktail party, because the honest one-line description is “we sell nothing.” We don’t want to build the model. We don’t want to win the application market. We don’t want to run the data center. We want to be the instruction set for verified answers — the double-D logo on the box that tells a regulated buyer the output can be trusted — and to be paid a small, dull, recurring amount every single time, across a base that is growing faster than almost anything in the history of this industry.

Small and dull and recurring, multiplied by enormous, is exactly how ARM quietly became worth more than the giants whose chips it lives inside. You don’t have to win the market. You have to be the toll on it.

I find this clarifying, and not only for business reasons. For a decade the working assumption in AI has been that the prize goes to whoever builds the biggest brain. I think that’s the wrong race, or at least the wrong finish line. The biggest brain is going to get cheaper every year, the way the fastest chip and the cleanest sound did before it. The durable money — the ARM money, the Dolby money — is going to belong to whoever owns the standard that everyone else has to license in order to be believed.

Trust is going to be the most valuable thing in artificial intelligence. And trust, it turns out, is something you license, not something you sell.

Disclosure: I’m a co-founder of 2Brains, Inc., which is built on exactly the idea in this column — so season my enthusiasm with the appropriate grain of salt. I’ve tried to make the argument stand on its own two feet, the way ARM and Dolby do, whether or not you ever pay me a nickel