A lot of the time, when you ask an AI a question, there’s a “chooser” standing between you and the model. Sometimes it picks the model. Sometimes you picked the model and it picks who serves it. Sometimes it only picks the capacity tier, or maybe the fallback when something breaks. What looks like one API call can hold a small market’s worth of decisions. This is what kids today are calling a router. And until this week, routers were the kind of thing only infrastructure engineers argued about.
Then the money showed up.
Last week, Bloomberg reported that the financial services company Stripe was buying OpenRouter for more than $7B. By Wednesday, Stripe confirmed it had agreed to acquire the company — price undisclosed, though reports claimed it had crept toward $8B by then. OpenRouter, which launched in early 2023, was valued at $1.3B in May. Three months later, it’s worth roughly six times that. Next act: That same Wednesday, Ramp — yup, the corporate card company — launched Router.com. (They bought the domain for the occasion.) Routing is now free through the end of 2026!
In one week, two of the best payments companies in the world moved to own the same piece of plumbing: the thing that decides which AI model answers your question. Stripe spent a decade making money from moving money. Ramp built a business on knowing where corporate money goes. Now both are pointing those skills at intelligence.
The money people are buying the meter.
Will the Layer Stay Neutral?
Look, I’m one of the many reasons the meter is worth buying. I’m a developer who uses OpenRouter. One API key and I can call almost any model on the market. I don’t have to write a separate integration for every provider, and I don’t have to keep track of half a dozen vendor accounts. It’s super convenient.
When I ask an AI something, I’m almost never talking to one model (there are hundreds now). The router sits between the question and all of them. It picks the backup when a model fails or is unavailable, retries when calls hang, caches what it has already seen, and chooses which provider actually serves the tokens — at what speed, what price, and sometimes what quantization. (I’ve got a whole post on providers and quantization.) You can pin a model and it still sometimes load-balances providers underneath you. Today, most of that policy is still yours to override. The real question is whether override stays a first-class feature — or whether choosing is something you’ll have quietly handed over. (Even the labs run this way: GPT-5 in ChatGPT was, by OpenAI’s own description, a fast model, a deeper reasoning model, and a router deciding on every request which one you get.)
So follow the money. Routers got popular for good reasons. Mine is convenience. The deeper one is churn. Models keep changing underneath you: deprecated, repriced, and once in a while switched off entirely. A router was insurance: one API, and you never get stranded. Then the insurance provider got bought. OpenRouter’s own “about” page says it exists to eliminate vendor lock-in. But…the anti-lock-in layer got locked in. Stripe and OpenRouter both say the layer stays neutral, and I take them at their word. The word isn’t the problem. Words don’t survive ownership changes. Structure does.
The Traffic Report Has Started Steering the Traffic
Why would payments companies want this layer? Because it’s the intelligence play. The router sees the flow: which models actually get used, through which providers, at what cost, and how that shifts week to week. Not the words — OpenRouter says it doesn’t store prompts or responses by default — but the shape of the whole market. Its research team has analyzed over 100 trillion tokens of real-world traffic. When Mozilla wrote the State of Open Source AI report, that public data was one of our best instruments for measuring what developers actually run. It’s a traffic camera at one of the busiest corners in AI. And its new Auto Router already uses aggregate spend to decide where future requests go: The traffic report has started steering the traffic.
Stripe says it’s buying routing and token optimization, and Stripe CEO Patrick Collison put it plainly: “Tokens are the central currency for companies building with AI.” In markets, the intermediary who sees the whole flow learns things no single participant can. Stripe made its fortune metering money. Metering intelligence is the same business at a bigger scale.
The chatbot you ask about the mole on your arm, your kid’s tutoring app, or the tool your lawyer runs contracts through: behind each one a router decides — on cost grounds you’ll never see — which model is good enough for you. Ramp’s launch copy promises “the lowest-cost model that meets the required performance level.” Required by whom?
Finance already ran this play. Robinhood made trading free; the money came from market makers who paid to have your orders routed their way. (Basically, payment for “order flow.”) The SEC eventually fined Robinhood $65M over how it disclosed the arrangement. That’s the warning from history. But when routing goes free, it’s fair to ask who the flow is for.
On top of visibility, the moment a broker gets discretion over where your order goes, price stops being the only question. It’s the 10-blue-links problem, one layer back: first the chatbot hides the 10 blue links behind a single answer; now the router picks the brain behind the chatbot. Securities markets eventually built disclosure rules around exactly this: Rule 606 makes brokers show their routing. AI has no Rule 606. It should, and for every request: which model, which provider, which tier, which fallback, what it cost, what the router was optimizing for, and whether anyone paid for preference.
The buyers’ own industry is already regulated for exactly this. Since 2010, the Durbin Amendment has required every U.S. debit card to work on at least two unaffiliated networks, so the merchant, not the network, keeps the choice of route. Congress looked at payment routing and decided owning the route was owning too much. It’s fighting right now over whether to extend the same rule to credit cards. But AI routing is a layer with no Durbin, no Rule 606, no rules at all. Nobody prices a captured route better than the people the law forces to give one up. So the question worth pricing: Does durable value sit in owning the switch, or in defining the interface that makes switches replaceable?
Setting the Bar for Any Router
All of it would matter less if routing were neutral plumbing — same weights, same behavior, whoever serves them. It isn’t. I measured. This summer, I ran the same coding tasks, in the same harness, against the same model name served by different providers, and the solve rate ranged from 26% to 44%. Eighteen points between the best and worst serving of one model, bigger than the gap between the flagship harnesses themselves. “Same model” is not enough information to reproduce a result.
So do we self-host everything and churn our own butter? Of course not. Ownership never meant owning every dependency. It means having a choice. And the layer that decides how your application uses intelligence is exactly the choice you can’t afford to hand over. Here’s how we should set the bar, for any router — closed or open:
You can read the meter. A routing receipt on every request: model, provider, tier, fallbacks, price, latency, data policy, and the router’s own economics wherever they touch the choice. “Why did I get this answer?” should have an answer.
Your policy lives in your repo. Routing is application logic — code you can diff, test, and export. A dashboard can be a view of the policy, never its only copy.
Exit costs a weekend, not a quarter. Lock-in gets measured on your worst day. A change of owners upstream should be a config change downstream, and an exit that’s never been tested is a promise, not a property.
The chooser never profits from your worst outcome. Disclosed economics, inspectable objective. If the router gets paid differently depending on where you land, the receipt says so. You can’t audit a promise.
Open weights need open pipes. The license lets you possess the model. The serving path decides how much control you keep when you actually run it. I have the benchmarks.
The old objection was practical: Fine, the source is open, but who has time to operate someone else’s router codebase? But look at the alternative. Point a coding agent at an open router and ask it to show you exactly how fallbacks get chosen, or to add an adapter for your internal endpoint — done by the afternoon. Open source was always about leverage more than charity. Well, the leverage just multiplied.
Stripe and Ramp haven’t proved that the router will own the AI economy. They’ve placed an unusually expensive bet that the execution layer matters. For now, that layer is still yours.
Believe the bet. Then go read your own meter.