So Why Did Stripe Buy OpenRouter?

Stripe's founders told their investors that the singularity began on January 1. That line comes from a leaked letter about the deal that TechCrunch verified: they wrote that "we decided that January 1 marked the beginning of the singularity" and had been running the company on that basis ever since. Patrick Collison has been using the same bit onstage since April, and he admits it's half a joke.

It also isn't the reason for the purchase.

The reason is simpler. Stripe makes its money sitting between businesses and the cash coming in. OpenRouter sits between businesses and the cash going out, specifically the cash they spend on AI, which is on its way to becoming one of the biggest line items in software. Stripe bought the meter on the other side of the ledger.

First, in plain English: what does OpenRouter do?

It's a universal adapter.

A company that wants to use several AI models normally has to open an account with each provider, keep track of a separate API key for each one, learn the quirks of each API, and rewrite code every time it wants to switch. Now do that across the dozens of models a real product might touch.

OpenRouter collapses all of it into one connection. One account, one key, one integration, and behind it 400-plus models from more than 80 providers (its investors put the model count above 500). Andrej Karpathy has described it as the transfer switch of AI, and Menlo Ventures, an investor, borrowed the phrase for its own writeup of the deal: OpenRouter is the "transfer switch," the way Stripe became the switch for card payments.

Four words you'll hit in every other article about this:

• Token. The unit AI companies bill by, roughly a fragment of a word. Think kilowatt-hour, but for intelligence. Every prompt going in and every answer coming out gets counted and charged.

• Gateway. The single connection point described above. One door to every model instead of eighty separate doors.

• Routing. Choosing which model handles which task. A cheap, fast model can pull names out of a document perfectly well. Paying frontier prices for work like that is how an AI bill triples without anyone noticing.

• Inference. What it costs to run a model, as opposed to what it cost to build one. Training is a capital expense the labs carry. Inference is an operating expense you carry, every day, indefinitely.

Scale is what explains the price. OpenRouter says it now moves more than 10 trillion tokens a day for over 10 million developers and companies, with volume growing at least tenfold every year since launch. Menlo's figures are more granular: about 33% growth month over month for three years, doubling every eleven weeks, a 30,000x increase since the company started. NVIDIA, Zoom and Lovable are named customers. Headcount is around ninety.

What Stripe actually paid, and why the number keeps moving

Stripe didn't disclose terms, which is why three different figures are circulating:

Per the New York Times, the founders take roughly $1.5 billion and investors split the remaining $6 billion. The founders' share on its own is larger than the entire company's valuation twelve weeks earlier.

In May, OpenRouter raised $113 million at a $1.3 billion valuation, bringing its total raised to $164 million. Call it a five to sixfold re-rating in a single quarter.

The direction of travel matters too. When the Wall Street Journal first reported the talks in July, the figure being discussed was closer to $10 billion. It came down. In the intervening weeks, more than ten companies shipped routing products of their own, Ramp and Cursor among them, and Stripe still had to outbid Databricks. Both halves of that are true at once: the category is worth fighting over, and it's commoditizing fast enough to take $2.5 billion off an asking price inside a month.

The five real reasons

1. Stripe bought the expense side of the ledger

Almost every large acquisition Stripe has made has been about money coming in: collecting it, managing it, moving it around. This one is about money going out.

That's a real change in direction, and it's the thing to understand about the deal. Stripe is working toward owning both sides of a software company's P&L, starting with the fastest-growing expense category any of them have.

2. Whoever holds the meter sees everything

Stripe already sells a token billing product and has spent the past year helping companies track and control what they spend on AI. OpenRouter is the meter itself, sitting at the exact point where tokens get requested, priced and consumed.

Collison's framing in the announcement was that "Tokens are the central currency for companies building with AI." Currencies need clearing infrastructure. Stripe built that for dollars, and it's building the same thing for tokens.

3. The data is the real asset

Every request through OpenRouter leaves a record: what the task was, which model got it, what it cost, how fast it came back, whether the output held up. Across 10 trillion tokens a day, that adds up to an unusually complete picture of what AI is being used for and which models are good at what.

Menlo's analysis is blunt about why this matters. Routing on the prompt alone doesn't work, because in a multi-step agent run a single bad routing call compounds through everything downstream. Making it work takes context and history, which is to say it takes the dataset. Nobody else has one this size.

4. Leverage over the labs

PitchBook analyst Franco Granda told TechCrunch the deal is Stripe deliberately putting itself in the middle of capital flows in the AI economy, and that it hands Stripe some degree of power over its own suppliers: the frontier labs, the hyperscalers, the neoclouds.

OpenRouter's position backs that up. New models often launch there first. OpenAI has offered exclusive discounts through it. Zuckerberg and Musk have each personally announced model launches on the platform. Distribution to ten million developers is leverage, and Stripe owns it now.

5. Agents multiply everything above

An agent working through a long task makes hundreds of model calls, and it makes routing decisions constantly without a human in the loop. Agentic workloads are the fastest-growing source of token consumption anywhere.

Stripe has been building toward agentic commerce, meaning agents that transact, for a while now. Owning the layer where agents buy intelligence as well as the layer where they buy everything else follows from that.

The part the announcements skip

Two risks stand out.

Neutrality. OpenRouter's whole pitch is that it doesn't own a model and has no reason to favor one. It's now owned by a company that processes payments for the frontier labs. Both sides got out ahead of the question: OpenRouter promised "same mission, same name, same product, same roadmap" and said routing decisions would keep being driven by what's best for the user. When two companies spend paragraphs answering a criticism nobody has made yet, that tells you where they expect the pressure.

Commoditization. More than ten competitors shipped routers in a matter of weeks. Databricks has its own gateway. Ramp and Rippling launched AI spend tools. The bull case says OpenRouter's scale and data can't be matched. The bear case says basic routing turns into a checkbox feature inside tools companies already pay for, and Stripe paid a strategic premium for a commodity with a strong brand.

What this means for your stack

If you already build on OpenRouter: both companies say nothing changes in the near term, and the deal should close within weeks. Keep your integration layer abstracted anyway. Portability is why you adopted a gateway in the first place, so don't hand it back to one.

If you're evaluating AI gateways: the category got validated at $7.5 billion and consolidated in the same week. Expect hard competitive pushes from Ramp, Databricks and everyone else who shipped a router last month. Pricing should move your way.

If you own the AI budget: treat token spend the way your company learned to treat cloud spend around 2015, as an operating expense with tooling, an owner and a monthly review. The tooling exists now, and the vendors have decided this is a category.

If you're a model provider: your distribution is being intermediated, and the intermediary now belongs to the company that also runs your payments.

Stripe has always described its mission as increasing the GDP of the internet. That now includes intelligence.