Its CEO Called It "Stripe for AI." Three Months Later, Stripe Paid $7 Billion. | Edition 316

In May, OpenRouter’s chief executive Alex Atallah described his company as the equivalent of Stripe for AI. On August 16, Bloomberg reported that Stripe had finalized an agreement to buy it for more than $7 billion.

Three months earlier, investors had valued OpenRouter at $1.3 billion. The company builds what the industry calls an AI gateway: a single API that reaches more than 400 models from OpenAI, Anthropic, Google, Meta, DeepSeek and others, and decides which one handles each request based on price, speed and reliability. If a provider raises prices or has a bad week, traffic moves elsewhere and nothing in your code changes.

That was the entire pitch. One integration, no lock-in, no marrying a single lab. It worked well enough that a payments company paid frontier-lab money for it — and in doing so, bought the one layer whose value proposition was belonging to nobody in particular.

What Stripe is actually buying

OpenRouter was founded in 2023 by Atallah — previously a co-founder of the NFT marketplace OpenSea — and Louis Vichy. Its growth curve is the part that explains the price. Weekly throughput reached about 25 trillion tokens by May, five times the figure six months earlier. Revenue is a far smaller number: roughly $50 million annualized as of March, against about $19 million when 2025 closed.

So the reported price is many multiples of revenue — more than a hundred times the March figure, and independent estimates of more recent revenue put the live multiple materially lower. Pick whichever you like: neither is a number anyone pays for cash flow. It is a number you pay for position.

One caveat worth holding onto: the 8 million user figure is OpenRouter’s own claim, not an audited one. Neither is the 400-plus model count. They are almost certainly directionally right, and they are still marketing numbers.

Stripe already owned the meter

This is the part the deal makes sense through, and it started well before this weekend.

In January, Stripe completed its acquisition of Metronome, a platform built to handle billing for complex usage-based pricing. Announcing it, Patrick Collison put the reasoning bluntly: metered pricing, he said, is the native business model for the AI era. Stripe’s own blog names both OpenAI and Anthropic among the companies whose billing it handles. Stripe also co-authored the Agentic Commerce Protocol with OpenAI, and had been OpenRouter’s own payments provider since around January, through a token-billing integration that prices model usage automatically.

Line those up and the gap becomes obvious. Stripe could already measure how much AI got consumed, and bill for it, across much of the industry. What it could not see was which model got picked, and why. The router is that decision. It is the one piece of the transaction Stripe was handling everything around and nothing inside.

The neutrality question, carefully

The tempting version of this story is that a neutral layer just got captured. That is worth slowing down on, because the honest picture is more interesting than the alarming one.

The case that this is fine: Stripe does not make models. It has no reason to prefer Anthropic over OpenAI over a cheap open-weight alternative, because it does not sell any of them. Its incentive is volume — more requests, more metered transactions, more billing relationship. Of all the companies that could plausibly have bought the industry’s most-used router, a payments processor is a genuinely less conflicted owner than any model lab would have been.

The case for paying attention anyway: a routing layer sits on data about which models developers choose, at what price, across both closed and open-weight providers. That is an unusually precise map of what AI actually costs and what people actually pick — and it now belongs to a company that sells financial infrastructure to the labs on one side and the developers on the other. Nothing about that is improper. It is simply a different set of hands than the ones the lock-in pitch was made with.

Worth stating plainly: there is no evidence of anything untoward here, no announced change to how routing works, and no reason to expect your requests to start going somewhere strange on Monday. What changed is who you are trusting, and whether you ever consciously decided to trust them.

What nobody has said yet

The gaps in this story are as notable as its contents.

• Neither company has confirmed the deal on the record. Stripe told TechCrunch it does not comment on rumors or speculation.
• The final price could still change. Reporting in July put the talks nearer $10 billion; the number that landed is lower.
• No cash-versus-stock split has been disclosed.
• Nothing has been said about Atallah’s role after close — which, for a company whose value is partly its perceived independence, is not a small omission.

If you build on OpenRouter

Four things, in order of how soon they matter.

Nothing breaks this week. Do not rip out working infrastructure over an unconfirmed acquisition. Panic-migrating is a larger risk than the acquisition.

Find out what your fallback actually is. LiteLLM is open source and does substantially the same job. The major clouds have folded routing into their own model services. Direct provider APIs never went away. The useful question is not whether you would switch, it is how many days it would take — and most teams have never measured that.

Write down which decisions you have delegated. If you cannot state the criteria currently choosing your model, you have no way to notice when they change. That is true of any intermediary and it was true last week too.

Treat your usage record as disclosed, not private. Which models you call, how often, and at what price is information your gateway holds. Assume it is commercially meaningful to whoever owns it.

The line, read twice

Atallah’s description of his own company is going to be quoted for a while, and it deserves its second reading.

The equivalent of Stripe for AI was meant to describe a function: one door, many providers, no obligation to any of them. It was a promise about what OpenRouter would refuse to become.

Stripe appears to have read it as a description of an asset — and then bought the asset. The router that existed so developers would never have to depend on one company now depends on one company. Whether that turns out to matter depends entirely on what Stripe does next, which is the least satisfying and most accurate thing anyone can say about it today.

The reason to care is not that something bad happened. It is that the layer you chose specifically so you would not have to think about this is now a thing you have to think about.