OpenRouter raised money a few months ago at a $1.3 billion valuation. The Financial Times reports Stripe agreed to buy it for around $8 billion.

Stripe left the number out of its own announcement, which is its right and also a small tell about how much the figure would have dominated the story.

Whatever the exact price, nobody pays that multiple for the ability to pick a cheap model.

What actually changed hands is the measurement point

A company running four model providers has four invoices, four reporting formats, and no clean answer to what it costs to process one supplier invoice or resolve one support ticket. The spend exists. It is aggregated by vendor instead of by job.

OpenRouter sits precisely in that gap. Every request passes through it, and every request leaves behind which model served it, how long it took and what it cost.

Patrick Collison, Stripe cofounder and CEO, said it plainly in the announcement. He called tokens the central currency for companies building with AI, and tied the real-world economic potential to making good use of scarce compute.

Read that with Stripe’s own history in mind. This is the company that got large by taking a fraction of every transaction because it stood where the money moved. It just bought where the tokens move.

We have been making a version of this argument from the operations side in price per token lies, so measure per job. Stripe wrote a very large check on the same premise.

The product, in plain terms

Per Stripe’s release, OpenRouter routes across more than 400 models from over 80 providers, evaluating each request and sending it to whichever model fits on complexity, price, speed and reliability. NVIDIA, Zoom and Lovable are named as customers.

Alex Atallah, OpenRouter cofounder and CEO, framed the deal around Stripe’s neutrality, a decade of building infrastructure that does not compete with the businesses on top of it. Standard acquisition language, and still pointing at the thing that makes the product work. A router with a house model to push stops being a router.

That is a useful test for anyone evaluating any middle layer in an AI stack, not only this one. Ask who profits when the router guesses wrong in a particular direction.

Routing gets cheap on its own

Routing between models is genuinely useful and also straightforward to copy. Open source libraries do a passable version of it in an afternoon. Once three vendors sell the same capability, its price walks toward zero, because nobody pays a premium for the service of choosing the cheap option.

Semafor reads the deal that way, and I find it persuasive. Stripe treats routing as low-margin commodity plumbing and monetizes the ring around it. Billing. Tax. Fraud. Stablecoin settlement. Treasury. Financing, eventually.

Every one of those depends on a single thing: a trustworthy per-customer record of consumption. The same data that lets Stripe tell a merchant what it sold today lets it tell a software company what it cost to serve each user tomorrow.

Which produces a number very few software companies can currently calculate: gross margin per customer when the variable cost is inference. We got into this when cost stopped being the guardrail on your AI. While models were expensive, the budget did the governing. Now that they are cheap, somebody has to govern on purpose.

A short test if you already pay three model vendors

Pick one process that already runs on AI. The highest-volume one. Now answer what the last run of it cost.

Not the OpenAI invoice or the Anthropic invoice. That process, that execution.

If the answer takes more than ten minutes to produce, the gap is not tooling. Your spend is organized by vendor rather than by job, which is the exact hole Stripe just paid billions to stand in.

The reflex will be to go shopping for a platform that fixes it. Better to first decide which unit of work you want to be able to price, because no tool makes that call for you and everything downstream depends on it. One processed invoice. One closed ticket. One reviewed contract. Pick one, instrument that, and compare platforms afterward.

Get the number for a single process and the rest of the case writes itself in front of your board.

Let us instrument one process so it has a cost per job