Bloomberg reported on August 16 that Stripe finalized a deal to buy OpenRouter for more than $7 billion. OpenRouter sits between developers and the model providers, routing each request to whichever of 400-plus models fits the job on quality, price and latency. It claims 8 million users. It has fewer than 100 employees. In May it raised $113 million at a $1.3 billion valuation, so Stripe paid roughly five times what the company was worth in the spring.
The revenue math depends on whose figure you trust, which is the normal problem with private companies. Reported annualized revenue is in the neighborhood of $140 million, up from something closer to $19 million when 2025 ended. Cost to serve runs under 30% of that. So: a business with maybe $100 million of gross profit, sold for $7 billion, at somewhere between 50 and 70 times revenue. Most of the commentary has been arguing about that multiple, which is a fine argument for people who trade private-market secondaries and a useless one for anyone who sells AI for a living.
Here is what a gateway does to a deal. It turns switching from a conversation into a configuration.
The switch stops being a threat and becomes a setting
In July I wrote about Kimi K3 showing up in renewals before it shows up in production, and that argument was about negotiating position. A buyer picks up a credible-sounding zero-dollar alternative, waves it across the table, and takes a discount whether or not anyone was ever going to run it. I sold through years of that at Oracle. The migration stayed six months away forever. The discounts were immediate.
A router changes the shape of that. Nobody waves anything. The rule in the config says route this class of request to whichever model clears the quality bar at the lowest cost per token, and traffic moves the day a competitor cuts price. No renewal cycle, no bake-off, no meeting. You find out from the usage line.
I lived the infrastructure version at Fastly. The customers who mattered most ran multi-CDN, with a traffic manager in front splitting requests across us and a competitor by percentage. Those percentages were a dial somebody adjusted on a Tuesday. The commercial negotiation had effectively concluded the day that architecture went in; everything after was arguing about the size of the commit and pretending we had more room than we did. The reps who did well in that book stopped selling delivery and started selling the things the traffic manager couldn’t arbitrate: edge logic the customer had written against our platform, observability their team had built dashboards on, support engineers who answered at 2am.
That is the same job now, one layer up. A router can shop a token. It cannot shop your evals, your guardrails, the fine-tune sitting on the customer’s data, the workflow their operations team built around your outputs, or who signs the indemnity when a model says something expensive. Price your product on cost per finished task rather than per million tokens, and make sure the parts a rules file can’t compare are the parts your renewal rests on.
Stripe is collecting both meters on the same company
The reason this deal is worth more than a shrug about multiples is who bought it.
Stripe co-wrote the Agentic Commerce Protocol with OpenAI and released it under Apache 2.0, along with the Shared Payment Token that lets ChatGPT trigger a purchase without handing the merchant’s payment credentials around. It powers Instant Checkout inside ChatGPT. In July, Stripe and Advent bid $60.50 a share for PayPal, about $53 billion. Now it owns the gateway that meters what a company spends on models.
So one vendor is in position to measure the money an enterprise pays out to AI suppliers and the money it collects from customers whose agents did the buying. Those used to be two unrelated systems owned by two departments that meet twice a year. I wrote in July about Salesforce buying m3ter and what happens when your vendor reads your usage curve before your own finance team does. This is that, on both sides of the ledger, at a company most CFOs already have an account with.
There’s also a plainer commercial logic. Stripe’s take on a payment is a fraction of a percent. OpenRouter’s take on a routed request is several percent. Same job, sitting in the middle of somebody else’s transaction, at ten-plus times the rate, in a category where volume is growing faster than payments ever did.
Two things can be true about the toll booth
I don’t think the router position is as safe as $7 billion implies, and it’s worth saying so.
Databricks closed $5 billion on August 13 at a $190 billion valuation, with proceeds earmarked in part for its own Unity AI Gateway. Every major cloud ships routing. LiteLLM is open source and free. And the frontier labs have every incentive in the world to make their models hard to substitute, because a router is a machine for turning their product into a commodity. Toll booths only pay when the road has no shoulder.
At the same time, aggregation tends to stick where the customer relationship sits, and OpenRouter has 8 million developers who already send it their traffic and already get billed by Stripe. Both of those can hold. The category may compress to a few points of take rate while the position itself stays exactly where it is.
What to do about it this quarter
If you sell an AI product, ask your five largest accounts a question you probably haven’t: are we called directly, or through a gateway? If the answer is a gateway, ask which one, and get your win rate inside it. A router publishes your quality and price against every substitute in the same table, refreshed constantly, and the developer who edits that table has more influence on your revenue than the VP who signed your contract. Nobody has that person in the CRM.
If you buy AI, the gateway is turning into a governance surface whether you planned it or not. Somebody in your company can change which vendor gets the traffic by editing a rules file, and that decision moves real spend, crosses a data-residency line, and answers a question your security review thought it had settled. Write down who is allowed to touch it and what happens when a new model shows up cheaper.
Then go find the file. In most companies it already exists, somebody wrote it in an afternoon, and it is functioning as a procurement policy that nobody signed.