Rent the Model, Own the Exit

In August, investor Gavin Baker repeated a second-hand claim that Dario Amodei had said Anthropic might one day become the only private company left in the world. There was no recording or transcript.

An Anthropic researcher called the claim false. Amodei’s own public forecast, made months earlier, was less cinematic: AI would probably resemble cloud computing, with three or four large players rather than one.

The rumor was better than the transcript.

It also captured a fear that now shapes how companies buy AI.

Subscribe to a frontier model and you expose sensitive work, fund a powerful supplier, and risk becoming dependent on it. Refuse, and a competitor may automate first. That sounds like a prisoner’s dilemma.

The arms-race case is respectable. If two law firms can cut the time needed to review a contract, neither can comfortably agree to wait. The first adopter may win clients for a while.

Once both adopt, faster work becomes the new minimum. Prices adjust, and some of the gain passes to clients. The firms paid to run faster but did not necessarily become more profitable. The model provider collected a toll from both.

There is a second concern. A company may believe its advantage sits in documents and databases. Then it discovers that much of it actually lived in the way experienced people asked questions, checked answers, and handled exceptions.

Put those interactions into a vendor’s product and the fear is that the supplier learns the trade while the customer teaches it.

That version is too broad. Major enterprise and API contracts now say business inputs and outputs are not used for model training by default. Eligible customers can also get retention and regional-processing controls.

These promises do not abolish security risk. A contract cannot retrieve a secret after a breach. But the choice is not consumer ChatGPT or a cave with no electricity.

I think the procurement case starts here. Companies are not choosing once between adoption and abstinence.

They can buy several models. They can reserve the best hosted system for tasks where its extra quality matters, then run cheaper or open-weight models for stable, high-volume, or sensitive work.

Amazon Bedrock already exposes a wide catalog through several common APIs. Meta distributes downloadable Llama weights. Mistral says most of its customers run models inside their own data centers or cloud environments.

None of this makes switching free.

The API call is often the cheapest part.

Lock-in accumulates in prompt libraries, retrieval pipelines, proprietary embeddings, fine-tunes, tool schemas, evaluation habits, and minimum-spend contracts. A team that claims to be model-agnostic because it renamed one endpoint may be enjoying a particularly modern form of self-esteem.

The real exit is an operating capability. Keep the business logic outside the model. Own the data, the tool permissions, and the test cases that define an acceptable answer.

Run those tests against a second provider before the first provider gives you a reason. Measure how long a switch takes, what quality falls, and which workflows cannot move.

Portability is not a diagram.

It is a fire drill.

Open weights strengthen that threat, but they do not grant free independence. Self-hosting replaces a variable API bill with GPUs, engineers, capacity planning, patches, security work, and the privilege of being paged at 3 a.m.

For some regulated or high-volume workloads, that trade is excellent.

For a midsize company chasing the frontier on every task, it may be an expensive way to own yesterday’s model.

So the optimal strategy is a barbell. Rent frontier intelligence where the quality gap earns its cost. Own the evaluation and workflow that tell you whether the gap still exists.

Move repeatable or sensitive work toward models you can host or replace. The point is not to avoid suppliers. It is to stop any supplier from becoming the only place where your company knows how to work.

Who wins under that arrangement? Model labs can build large businesses, especially if frontier performance stays scarce. Cloud companies and chip suppliers can win across several models because almost every path consumes their infrastructure.

Enterprises keep more of the value when AI is buried inside proprietary distribution, data, and decisions. Thin wrappers that merely add a logo to somebody else’s tokens should keep the champagne inexpensive.

I may be wrong about the market structure.

Training frontier models demands enough capital and expertise to support an oligopoly with real margins.

Yet this market lacks the clean network effect that made one social network more valuable merely because everyone else was already there. Customers can use more than one model. Old capabilities spread into open weights, and today’s leader can be tomorrow’s fallback.

This is why “one company left” is a poor base case even before we note that the famous line is not a verified Amodei quote.

Three or four frontier labs could still hold serious power.

The durable contest, however, is not only between those labs. It is between suppliers trying to make their intelligence indispensable and customers trying to make it replaceable.

The customer does not escape by cancelling every subscription.

It escapes by owning a door and proving, regularly, that the door opens.