Stripe Is Buying OpenRouter: A Payments Company Becomes the Gatekeeper of Models

What happened

Stripe's announcement says it has signed an agreement to acquire OpenRouter; it calls the router "a leading AI model gateway and routing platform". OpenRouter's post the same day says the deal has not closed yet: "The transaction is subject to customary closing conditions." and "We expect to close in the coming weeks." This is not an acqui-hire: the product, the brand and the team stay, and the team is growing.

No official statement gives a purchase price. The figures circulating in the press all come from anonymous sources, and they do not even agree with each other: above 7 billion dollars (Bloomberg), 7.5 billion (New York Times), more than 8 billion in cash and stock (Axios), 6.8 billion euros (Forbes España). If you see a single hard number somewhere labelled "the value of the acquisition", that is not official data.

How big is the thing they bought

OpenRouter's own figures, published on the day of the announcement:

• more than 10 trillion tokens a day pass through it,
• more than 400 models, from more than 80 providers according to Stripe's announcement,
• more than 10 million developers and companies use it,
• founded in early 2023, the team is 90 people.

The growth curve shows in the earlier official announcements too: at the 113 million dollar Series B in May they were still writing about 25 trillion tokens a week, and by 10 August it was 55 trillion a week. These are all the company's own claims; none of them has independent confirmation.

What they promised — and what they did not

OpenRouter's post states four things:

• Independence: "same mission, same name, same product, same roadmap."
• Continuity: "If you build on OpenRouter today, nothing about your integration changes."
• Neutrality: the level playing field for models "doesn't bend to any model, any provider, or any parent company".
• Routing: "Routing decisions will remain driven by one thing: what's best for you, the user."

What no official statement covers: pricing, the fee structure and data handling. They neither announce a change nor promise that things stay as they are. The most that is said is the "current commitments are unchanged" formula. At an acquisition that is standard — but if your process runs through this layer, those three are exactly the things you care about.

Where a model router makes its money

Worth knowing, because it is the logic of the acquisition. OpenRouter's documentation says it puts no markup on the token price: "We pass through the pricing of the underlying providers without any markup on inference pricing" — you always pay the provider's list price. The revenue is in the credit top-up: 5.5% on a card top-up (minimum 0.80 dollars), 5% with crypto. With your own key (BYOK), 5% above a monthly allowance.

you top up 100 USD of credit
  -5.50 USD  the router's fee
  = 94.50 USD goes to tokens
the token list price stays the same

So the router does not earn on the model, it earns on the movement of money. From that angle it makes complete sense that a payments company is the buyer: Stripe's own announcement points out that it has been working since last year on helping companies optimise their token costs, and that it launched a separate product for token-based billing. The token as a unit of account and the payment infrastructure are the same business.

Router or aggregator — not the same thing

The two words get mixed up, and yet their risk is different.

• Aggregator (gateway): one API, with many models and many providers behind it. Unified authentication, billing and logging, plus failover when something goes down. You name the model in the call.
• Router: per request, the system decides which model serves it. Stripe's announcement describes what it bought in exactly those terms: it will "dynamically evaluate each request, routing it to the optimal model based on task complexity, price, speed, and reliability".

OpenRouter is both: an aggregator by default, but it also has an automatic routing mode. The new version announced on 10 August uses the market's token spend over the past seven days as its decision signal — that is, what everyone else is spending on right now. It is convenient, and it is exactly the kind of decision worth switching on deliberately rather than inheriting by accident.

The field is not a one-player one, by the way: Vercel AI Gateway also passes provider prices through "without markup", Cloudflare AI Gateway offers caching and failover, Microsoft Foundry uses a routing model it trained itself, AWS Bedrock predicts the expected response quality per request, and LiteLLM and Portkey start from an open-source gateway. So the router is not one company's product but a layer — and layers are replaceable.

What this means for you

If your company's AI calls go through a router, then you have an intermediary between you and the model, and it has just turned out that this intermediary is changing owner. That alone is no reason to move out. Four things are worth checking, though, and each of them fits inside an hour:

1. What does the intermediary cost you, in real money? Not in percent: 5.5% of the past three months' top-ups is a concrete number. That is what you weigh against what you get for it — one key for many models, one invoice, failover when something goes down.
2. Is there a way out, and how long does it take? If you switched tomorrow, in how many places in your code would you have to rewrite the endpoint and the model identifier? If it is one configuration line, you have no exposure. If it is scattered, you do.
3. What data goes through it? The intermediary sees the prompt. If text carrying personal or business-confidential material passes through as well, then the data-handling terms are not a small-print question — and that is precisely what neither party said anything about in the announcement.
4. Which model are you actually getting? If automatic routing is switched on, the quality of the output depends on what the system picked that day. It is worth logging the model identifier returned in the response — without it you cannot say afterwards what wrote that text.

There is a separate article about model routers and when they are worth using: model routers and aggregators — worth reading before you build on a gateway.

Common misconceptions

• "The acquisition has already happened." On the day of the announcement it had not closed: it is subject to customary closing conditions, and the closing usually gets a statement of its own.
• "The router adds a markup to the token price." This router, by its own documentation, does not; the fee sits on the credit top-up. At another provider it can work differently — go and look.
• "After the acquisition it will get more expensive." No official source claims that. But nobody promised it will not, either. The difference matters.
• "If everything goes through one gateway, I see the cost in one place." Yes — and your exposure is in one place too. The two are the same decision seen from two sides.
• "Neutrality is guaranteed, they wrote it down." A blog post is not a contract. If this matters commercially, it has to be in the contract.

What to take away

On its own the news is a corporate acquisition. What sits behind it is a structural change: the token has become a unit of account, and whoever handles the movement of money has an interest in handling the route of the token too. For the user that can be convenient — one key, one invoice, many models. The practical job is not to move out, it is to know: what the intermediary costs you, how long it would take to replace, what it sees of your data, and which model actually served your request. Those four answers are useful even if you never switch.