In five years, the companies everyone dismissed as "just a wrapper" will be worth more than the frontier models they run on. I know how that sounds. Two years ago, "you're just a wrapper" was the fastest way to end a pitch meeting. I think it's about to become a compliment.

Last week I made the case that intelligence probably has a ceiling, and that we may be closer to it than the first-inning crowd believes. This week is the follow-up question. If that's right, where does the money go?

Start with something most people in the industry would agree on. The open models trail the frontier models by roughly six to eight months. Whatever Anthropic or OpenAI ships today, expect something comparable with open weights a couple of quarters later. That lag is what the labs are actually selling. You pay frontier prices for the window when nobody else can do what their model does.

Now add the ceiling. If each new model adds less appreciable value than the one before it, the frontier eventually stops moving in any way a customer can feel. Fable 5 is better than the model before it. At some point the next release is better on a benchmark and indistinguishable inside your workflow. And the moment the frontier stops moving, the shadow closes. Six to eight months later, the open models are standing in the same spot, and you have a Fable 5-class model you can run wherever you want.

I've started calling that the parity point: the moment an open model is as good as the frontier on everything you would actually pay for.

Here's the irony. The labs are in a race to AGI, and I think reaching it is what undoes them. Once you build the model that no next model meaningfully improves on, you've handed the market a fixed target. Everyone catches up to a target that doesn't move. Each model after that one is worth less than the one before, and the importance of the model layer as a whole starts shrinking from that day forward.

At the parity point, every CFO is going to ask the same question: why are we still paying Anthropic or OpenAI for tokens?

The honest answer would be lock-in, and the labs don't have much of it. Switching model providers today is closer to a config change than a migration. Compare that to a database vendor or an ERP, where leaving takes years and a consulting army. The frontier labs have an API and a monthly bill. They could build real lock-in between now and then, through memory, agents, and integrations that are painful to unwind. It's possible. But they're starting from close to zero, and their customers can see them trying.

So the price of tokens heads toward the cost of compute. Electricity, hardware depreciation, a thin margin. Not free, but priced like a utility rather than like software. The labs will still be enormous businesses. They just won't be the ones charging a premium.

Go back to the original insult. If you build on top of Anthropic, Anthropic is going to clean your clock the day they decide to ship your feature themselves.

That's true. They can. But only if they build a better version than you did.

Look at Harvey. They built the whole experience around how attorneys actually work, and they've gone deep enough that they now build their own models. That's what makes them a useful example. The model was never the point. The point is who uses the models best, and behind the scenes the model doesn't matter a whole lot.

Could Claude clean their clock by incorporating all the legal stuff? Anthropic will certainly have the option to go after that market. But having the option isn't the same as building a better workflow than a company that has focused on nothing but the legal workflow for years. Legal is one of dozens of things Anthropic is working on. It's the only thing Harvey is working on.

The counterargument I hear is that soon you'll be able to tell Fable or Mythos, "build me the best attorney workflow product," and it will. I don't buy it, at least not on any timeline that matters for a business decision. Getting a workflow right takes nuance and experience: knowing which step people skip, which output they'll never trust without a human signature, where the review has to happen for compliance reasons nobody wrote down. That knowledge comes from sitting with the people doing the work. The model doesn't have it, and the lab that owns the model doesn't have time to go get it for every profession at once.

Run it forward. Model capability tops out. Open models reach parity. Token prices fall to compute. Who is left to charge a premium?

The software that built the best experience for a given workflow. The wrappers.

Two years ago the consensus was that wrappers had no moat and the models had all the power. Within five years I think it's the opposite. The models have very little pricing power, and the moat is the accumulated understanding of one specific job, encoded in a product that a lab with dozens of priorities will never replicate well enough to matter.

Harry Stebbings covered some of this ground on 20VC with Eno Reyes, the CTO of Factory, including whether American enterprises should run open-source Chinese models and how few of the new labs are likely to survive: https://www.youtube.com/watch?v=h9VNB9TA2Hk

I could be wrong. Maybe there is no ceiling, the shadow never closes, and the labs stay six months ahead forever with a product nobody can match. In that world, the frontier keeps its premium. But notice that this is the world every lab valuation is pricing in, and it depends on a claim about intelligence that nobody has proven.

For the operators reading this, the model layer is becoming the least important decision in your AI stack. The clients I work with almost never ask which model is under a tool. They ask whether it fits how their people work, and whether their team will actually use it. That instinct is correct. Judge AI software the way you judge any software: does it fit the workflow, and how hard is it to leave. The model underneath is going to be swappable, and if a vendor can't swap it, that's their problem, not a reason for you to pay more.

The same logic applies to anything you build internally. When we build something for a client, I want the model to be a line in a config file, because in a couple of years it will be, and the thing we actually built, the workflow, is the part that keeps its value.

Homework: Run the swap test. List every AI tool your company pays for. For each one, ask a single question: if the model underneath were replaced tomorrow with an open model six months behind the frontier, would anyone here notice? If the answer is no, you're paying for the workflow, so evaluate it and negotiate it as software. If the answer is yes, write down exactly what would break. That short list is the only part of your AI spend that is actually buying intelligence, and it's probably shorter than you think.

Two things I'd genuinely like back from you. First, hit Reply and tell me which part of this matches what you're seeing. Second, if there's someone in your world who should be on this list, forward it to them.

- Robbie