The Real Reason Why Stripe Bought OpenRouter

0:00 On August 19th, a letter reached Stripe's investors. One of its section headings is two words long, the singularity. Not a frontier lab, not a chip maker, the company behind your

0:10 checkout page, and it gave the thing a start date, January 1st, 2026. Not a forecast, a date they had already passed 7 months earlier. Same letter, same week, their largest acquisition ever.

0:24 Billions of dollars for a company whose name you have likely not seen. So, the question is narrow. What does a payments company see in its own books that makes it write that word? Stripe does not

0:34 train models, it keeps a ledger. A ledger knows one thing, who paid whom and when. So, whatever convinced them is sitting in that ledger with a number attached, and it is not a benchmark

0:44 score. Here is the sentence, hedge included. It's a fuzzy and perhaps already overworked term, but we decided. We decided that January 1st marked the beginning of the singularity, and we

0:55 have since been operating on that basis. Operating, not predicting. And it is signed by three people, Patrick, John, and Will. Two founders and the president who runs technology and business, and

1:07 the evidence they give for it is one line long. It is the least science fiction sentence in the letter. A huge increase in the rate of new firm creation, people starting companies.

1:17 That is their whole case for the word, and it deserves taking seriously because of what Stripe can actually see. More than 5 million businesses run on Stripe, and the letter says the flows crossing

1:27 it represent almost 2% of global GDP. Whatever else that is, it is a very large sample of the world economy, and this is not a hype document. First half net revenue up 41% year-over-year. Free

1:40 cash flow up 43. 88% of the Forbes AI 50 build on Stripe, including OpenAI and Anthropic. The letter notes most of the missing 12% simply have not started charging for anything yet, and the share

1:52 of Stripe's revenue coming from AI companies and from crypto is more than doubling year over year. So, the instrument is pointed at the right part of the economy. Now, the reading. The

2:02 tool that matters is Stripe Atlas, which founders use to incorporate a company, and Stripe watches every single one get created. In the first quarter of 2026, Atlas incorporations were up 130% year

2:14 over year. Patrick Collison posted the chart himself. Not up 30%, not up 50, more than doubled. The letter adds the detail that makes it land. Atlas has now incorporated its 100,000th business, and

2:27 it accounts for over a quarter of all Delaware incorporations. One startup tool, one in four new Delaware companies. In late July, at Y Combinator's Startup School, Patrick

2:37 Collison gave the broader figure. New businesses launching on Stripe were up roughly two times year over year, the largest relative jump the company has recorded. The previous record was 2020,

2:48 during the pandemic, at about 50%. So, the pandemic small business boom, the one economists wrote papers about, is now the second biggest thing to happen to firm creation on Stripe. There is a

2:58 speed number underneath the count, and I think it matters more. 20% of Atlas startups now charge their first customer within 30 days. In 2020, that figure was 8%. And Patrick Collison

3:10 had already said the word out loud months earlier. April 29th, opening Stripe sessions, he called it day 119 of the singularity. Count it. January 1st to April 29th is 119 days. His exact

3:24 framing was, "We're being a bit tongue-in-cheek with this, of course, but only a bit." The hedge that reached investors four months later is a much milder one. Now, I have to slow down

3:33 because this next part decides whether the word is earned. Every number in that case is measured on Stripe rails by Stripe, about Stripe customers. Atlas up 130% tells you about Atlas. It does not

3:45 tell you about the economy. So, check it against something Stripe does not own, Delaware, where all of those companies get filed. Stripe's own post gives a comparison. Same quarter, Delaware

3:55 incorporations overall were up 38% year-over-year. 38 against 130. Both real, both growing, and Stripe's tool grew more than three times faster than the registry it files into. That gap is

4:09 not the singularity. That gap is Stripe winning share. Some of the parabola in their chart is founders switching to a better incorporation tool. The underlying boom is still there and

4:19 unusual by any measure. Census Bureau business applications have been running well above pre-pandemic levels all year. So, the read is an inflection, not an explosion. Which raises the question I

4:30 want you sitting with. If the biggest firm creation number in Stripe history is measured by Stripe on Stripe, what would convince you the economy changed and not just the market share? Hold that

4:40 because the letter does something more interesting than argue. It spends money. And what they bought tells you what they actually believe. The company is Open Router founded by Alex Atallah, who also

4:51 co-founded OpenSea. It does one job. You send it a prompt, it picks which AI model answers, and it bills you for the tokens. One API key. More than 400 models across 80 plus providers. OpenAI,

5:05 Anthropic, Google, and the Open Weight Labs all behind a single endpoint that speaks the OpenAI format. Which sounds like convenience until you look at what it is deciding. Every request gets

5:15 routed on price, latency, uptime, and throughput. A provider that goes down fails over to a live one mid-flight. Their measured overhead for making that decision is about 25 milliseconds.

5:26 Less time than a single frame of this video to choose between 400 models and hand your prompt to one of them. Scale is what made it worth billions. When Open Router raised its Series B in May,

5:36 it was routing 25 trillion tokens a week. Six months before that, the number was 5 trillion. The letter puts a rate on it, and the rate is a strange thing to read in an industry document. Token

5:47 consumption compounding at 9% per week. 9% a week is not a payments number. Compounded across a year and you land near 90 times where you started. So, that flow is what Stripe just bought a

5:58 position in. Not the models, the meter sitting in front of them. Their explanation for it is the clearest thing in the letter. Capital and intelligence, they write, are becoming the two digital

6:09 flows undergirding every business. The argument runs like this. Every developer already needs a reliable way to manage their revenue pipeline and serving that need built Stripe. Going forward, every

6:20 developer will also need a reliable way to manage their intelligence pipeline. Same shape of problem, different pipe. And they have receipts for the parallel from inside their own product line.

6:29 Radar was built to catch payment fraud. It is now guarding against token fraud at some of the largest AI companies. Metronome, which they bought in January for a reported billion dollars, meters

6:40 usage for Anthropic and Nvidia. So, Stripe already owns the layer that counts what an AI company sells. A darker line sits underneath that. At their conference in April, Stripe put a

6:49 number on the new kinds of fraud A AI businesses are seeing. Across eight fast-growing AI businesses, Radar blocked more than 3.3 million risky sign-ups in a single month, roughly one

7:00 attempt in six. Free trial abuse had more than doubled in six months. Now, the numbers that make this a real bet instead of a slogan. Last year, Stripe moved 1.9 trillion dollars of payments

7:10 and kept about 36 basis points. That is about a third of 1%. And on 1.9 trillion dollars, it is the 6.8 billion of revenue Stripe reported for the year. Open router charges 5 and 1/2% when you

7:22 top up your inference credits. 15 times Stripe blended rate on a flow more than 700 times smaller. About 2 and 1/2 billion dollars of annual AI spending runs through it today, which works out

7:34 to somewhere around 140 million dollars a year of revenue. Reported prices for the deal ran from over 7 billion to as high as 10, and Stripe never disclosed the figure.

7:44 At those numbers, they paid somewhere between 50 and 60 times revenue for a company valued at 1.3 billion in May. That price only makes sense as a bet on which of the two pipes gets bigger and

7:55 how fast. And the letter is blunt about where it thinks that bet ends up. Agents, they write, are on the cusp of becoming economic actors in their own right, not tools that help you buy

8:04 things, buyers. Stripe has been building for that reading for a year. Projects so an agent can register for a third-party service, a directory so agents can find products, Bridge for stable coins bought

8:16 for 1.1 billion, Privy for wallets, Tempo, a blockchain for machine payments. You can call that a crypto detour. I would call it a company that has

8:26 concluded its next customer might not have a credit card or a bank or a country. Which brings the problem. And developers raised it within hours of the news breaking. Open routers entire pitch

8:36 was neutrality. No stake in which model wins. You pick on price and quality, and the router does not care who you pick. Stripe has a stake in nearly everything adjacent. It builds the labs through

8:46 Metronome. It sells token billing inside its own product. Now it owns the gateway choosing between them. One analyst put it better than I can. A neutral referee cannot also be the house

8:56 accountant. And there is a second asset in this deal that nobody put a price on, the scoreboard. Open router sees what developers actually pay for, not what they post about. Forbes reported that

9:07 models of Chinese origin have held above 30% of US token volume there since February, peaking near 46. That is the closest thing this industry has to a real market share number, and it now

9:18 sits inside a company holding commercial relationships with almost every lab on the list. Now the part that made the week strange, because two things happened on the same day in opposite

9:27 directions. Stripe letter says, being private is a growing advantage as we venture into the vicissitudes of the singularity. The reasoning is unsentimental.

9:37 The profitability of the core payments engine lets them buy companies without diluting shareholders, and their share count is lower today than it was 3 years ago. And staying private is not passive

9:46 here. In July, Stripe and Advent bid more than $53 billion for PayPal at $60.50 a share. A private company bidding on a public one. Same day as the letter, CNBC

9:59 reported that OpenAI's Chief Financial Officer, Sarah Friar, told an all-hands the company will be a public company in 2027 or sooner if the business keeps inflecting. Two companies at the center

10:10 of the same boom. One says the uncertainty is a reason to stay private. The other says it is a reason to go public. They are not reading the same signal. So, here is my verdict, and it

10:20 is not the one the headline invites. Believe Stripe's spending. Discount Stripe's vocabulary. The billions of dollars are the honest signal in that letter. The word singularity is not.

10:30 Singularity means a capability explosion, a machine improving itself past our ability to follow. What the letter measures is an incorporation's chart. The villain here is not Stripe.

10:40 It is the move, borrowing a word's weight for a claim that does not carry the word's content. And I will argue their side because the data underneath is not nothing. Stripe has a live view

10:50 of small business formation that arrives faster than the official statistics do, and that view is climbing harder than it did in 2020. If I had to choose between that data set and most economic

11:00 forecasts, I would take the data set. But the receipts do not reach the word. Atlas grew 130% while Delaware grew 38. The letter only stated "Evidence is a payments metric." And the purchase

11:12 itself argues against the drama. You do not buy a metering layer because machines are about to leave us behind. You buy it because the bills are about to get complicated, which is a useful

11:21 claim inside all of this, and the boring one. If you build software now, you run two bills instead of one, revenue and tokens. The company that owns the first

11:30 one just paid billions to own the second. That is a forecast and far more specific than the word they led with, which leaves the question I cannot answer for you.

11:39 When the meter, the router, and the bill all belong to one company, who do you complain to when the routing gets worse?