Let’s talk about the company that sells you the engine — and keeps a window into every customer you win with it.
Picture a small team. Four or five people. They pick a boring, narrow problem — the admin that swallows an independent insurance broker’s whole afternoon. They build a tool on a frontier model. It works. Brokers pay for it. Eighteen months in it’s real revenue, and every session is tokens burned on somebody else’s API.
They feel safe because they’re narrow. Nobody big cares about insurance brokers.
Then the lab ships a first-party app that does the same thing. Not a smarter one. The same one. Priced inside a subscription people already have.
That’s not a conspiracy theory. It’s the oldest move in platform economics, and the AI labs just picked it up.
Here’s the mechanism, and it’s almost embarrassingly simple. When you build on a model API, the provider sees the shape of your traffic. Which use cases burn the most tokens. Which verticals are growing. Which features people ask for, and how often. They don’t need to read your code. They read your usage. The investor note from Foundation Capital puts it plainly — the major labs have a panoramic view of what’s working across the whole AI ecosystem. Then they point a team at the part that’s working.
We’ve watched this exact movie before, and this time we have the receipts. Amazon ran a marketplace for third-party sellers, watched which products sold, then launched its own private-label versions underneath them. Reuters, the Wall Street Journal and CNBC all documented it — employees pulling third-party sales data while developing competing products, the opposite of what the company told Congress. The seller carries all the risk of finding out what the market wants. The platform collects the answer and sells it back against them.
The AI version is faster, because the telemetry is live. Amazon had to wait for a product to sell. A model provider knows which prompt is popular before your app has finished launching.
Now add the second half of the squeeze. OpenAI’s Apps SDK puts third-party apps inside ChatGPT, in front of hundreds of millions of users. It’s built on an open standard — MCP — and the company says it’s open source so your app can run anywhere that adopts it. All true. And that word open is exactly what gets builders to walk through the door.
But the standard being open is not the part that matters. What matters is everything sitting around it: the directory that decides who gets found, the featured placement that decides who gets seen, the review that decides who gets published, and the checkout that has to run through the platform’s own protocol. Open port. Private gate.
And the platform gets to watch. Every app that connects, every tool that gets called, every vertical quietly catching fire — that’s a map handed to the company that can act on it first, with better distribution and a cheaper cost of capital than you will ever have.
So what do you actually do with this?
First, stop treating the model as your stack. The model is the weather. It changes on someone else’s schedule and you don’t get a vote.
Second, name the thing the provider cannot see through an API call. If your advantage is a clever prompt or a clean interface, you don’t have a business. You have a prototype the lab can ship in a quarter.
Third, and this is the one people skip: build the closed loop. The asset isn’t the data you scraped or the docs you bought. It’s the loop where real users do real work, outcomes get measured, and the system gets better from it. Sees the deal close, not just the email. Sees the claim settle, not just the form. The labs can’t see what happens between your customers’ tools, in their actual workflow, on the days nothing got logged anywhere. That’s still nobody’s. You can own it.
Fourth, own a relationship, not a listing. Distribution the platform can hand you is distribution the platform can take back. A customer who chose you because you solve one specific problem one specific way — and would follow you out the door — isn’t a user. That’s a moat, and it’s the only one that doesn’t ask permission.
None of this means don’t build on the models. Of course you do. Build there the way you’d build on land you don’t own. Know where the fence is. Know who holds the deed. Make sure what you’ve built can walk. And keep your own copy of the way it works.
Because here’s the part that should bother you more than the competition. When the platform can watch you win and then take the win, the deciding stops being yours. You’re not choosing which products exist anymore. You’re auditioning ideas for a company that keeps the ones that work — and pays you nothing for the audition.
So before you build the next thing on top of a model, hold it up to one question. If the lab shipped this exact feature next quarter, would a single customer stay?
If the answer is no, you already know what you’re building. You’re building their roadmap. Do it on purpose or don’t do it at all.