OpenRouter keeps about eleven cents for every million tokens it routes. Stripe kept thirty-six basis points on $1.9 trillion of payments last year. The price of this deal is a bet on which of those two lines grows faster.
Stripe has finalized an agreement to buy OpenRouter for more than $7 billion, Bloomberg reported on August 16, with the people describing the deal cautioning that the final price could still move. OpenRouter charges 5.5% when a customer tops up inference credits, then passes the model provider’s token price through untouched. That fee currently sits on roughly $2.5 billion of annual AI spending. Stripe processed $1.9 trillion in 2025 and kept 0.36% of it. So Stripe is buying a take rate more than fifteen times its own blended rate, on a flow about 750 times smaller, and paying roughly twice its own revenue multiple for the privilege.
Until June 2025, OpenRouter’s credit-purchase fee read as a percentage plus a fixed $0.35 charge, and the $0.35 was Stripe’s card fee passed straight through to the developer. A pricing archive reconstructed from Wayback Machine snapshots of OpenRouter’s own pricing page dates the change to June 9 of that year, when the company folded the pass-through into a flat 5.5% with a $0.80 floor and moved crypto payments to 5.0%. Stripe had been sitting inside OpenRouter’s cost line as a line item. Fourteen months later it agreed to buy the line.
What Happened
Bloomberg reported the finalized agreement on Sunday, August 16, at a price above $7 billion. The Wall Street Journal had reported the talks in late July, floating a figure closer to $10 billion. A Stripe spokesperson told TechCrunch the company does not comment on rumors or speculation.
The step-up is the part that made the rounds. OpenRouter closed a $113 million Series B on May 26, led by Alphabet’s CapitalG, at a reported $1.3 billion valuation, with Andreessen Horowitz, Menlo Ventures and Nvidia’s NVentures alongside. Eighty-two days later, a buyer agreed to pay 5.4 times that mark.
OpenRouter reported 8 million users, more than 400 models from 70-plus providers, and 25 trillion tokens routed per week as of May, five times the weekly figure six months earlier. Revenue is a far smaller number. Sacra put it at roughly $50 million annualized in March, up from about $19 million as 2025 closed. The Information later reported it near $140 million by mid-year.
Both revenue figures point back to the same underlying flow once you divide by the fee. At 5.5%, $50 million of fees implies about $900 million of inference spend passing through the platform; $140 million implies about $2.5 billion. Third-party snapshots of OpenRouter’s public rankings endpoints late in May put weekly spend around $32.4 million, which annualizes to roughly $1.7 billion and would generate about $93 million of fees. Two independent paths land in the same band, between the March and mid-year revenue marks.
The Backstory
Alex Atallah co-founded OpenSea and served as its CTO before starting OpenRouter in 2023 with Louis Vichy. The pitch has been consistent since: one API key, one balance, no markup on tokens, and the ability to move a workload between providers without rewriting the integration. Atallah has described the company as Stripe for AI, which reads differently now.
The funding ladder was modest until this year. A $12.5 million seed from Andreessen Horowitz and a $28 million Series A from Menlo were announced together in 2025 as a $40 million round at roughly $500 million. Annualized inference spend running through the platform went from about $10 million in October 2024 to above $100 million by May 2025.
Stripe’s side of the story is a shopping list. It closed the $1.1 billion Bridge acquisition for stablecoin infrastructure in February 2025 and bought wallet developer Privy in mid-2025, a platform reported to have enabled more than 100 million programmable wallets by early 2026. On January 13, 2026 it completed the Metronome acquisition, reported at roughly $1 billion, for usage-based billing; Metronome’s customer list already included OpenAI, Anthropic and Nvidia. Stripe partnered with Paradigm on Tempo, a payments blockchain whose mainnet launched in March with a protocol for machine-initiated micropayments. It co-wrote the Agentic Commerce Protocol with OpenAI in September 2025 and shipped an AI token billing preview on March 2. At Sessions in April it announced 288 products under the banner of building economic infrastructure for AI. In July it bid, with Advent International, more than $53 billion for PayPal, which PayPal’s board rejected as inadequate.
The Plan
Read the acquisitions as one stack and the logic stops being mysterious. Metronome counts the consumption. OpenRouter decides where the consumption goes and at what price. Privy holds the wallet. Bridge and Tempo settle. ACP handles the checkout when an agent is the buyer. Stripe has spent eighteen months assembling a complete billing and settlement pipeline for a unit of economic activity that had no standard invoice: the token.
Stripe already ran a meaningful share of OpenRouter’s financial back office before any of this. OpenRouter bills its own customers through Stripe, including regional methods like Alipay and Google Pay, and the two companies shipped a token-billing integration in January that meters model usage and prices it automatically. Buying the customer is the cheapest way to stop being the vendor.
The asset that justifies a nine-figure revenue business at a ten-figure price is the ledger. OpenRouter’s rankings page has become the industry’s de facto price sheet and demand signal, cited by investors and researchers because it reports settled spending rather than survey responses. Stripe has built products on payment exhaust for a decade, from fraud scoring to lending. Real-time visibility into which models 8 million developers actually pay for, at what price, is the same primitive pointed at a different market.
The Business Model Angle
Two toll booths, wildly different tolls.
Stripe’s list price on a card charge is 2.9% plus $0.30, but after interchange, scheme fees and enterprise pricing, the company booked $6.8 billion of revenue on $1.9 trillion of volume in 2025. That is a blended net take rate of 35.8 basis points. OpenRouter charges 550 basis points on credit purchases and 500 on bring-your-own-key usage above a $25,000 monthly allowance, with no interchange underneath to surrender. On the same arithmetic that governs any marketplace take rate, OpenRouter keeps more than fifteen times as much of every dollar that crosses it.
The flow is the problem. Stripe’s payment volume is roughly 750 times OpenRouter’s implied inference spend. Run the tokens through it and the unit economics get small fast: 25 trillion tokens a week annualizes to about 1.3 quadrillion, and $2.5 billion of spend across that volume works out to a blended $1.96 per million tokens, of which OpenRouter keeps about 10.8 cents.
At $7 billion, Stripe is paying about 50 times mid-year annualized revenue, or 2.75 times the annual spend crossing the platform. Its own February tender valued Stripe at $159 billion, or 23.4 times 2025 revenue and 0.084 times payment volume. So Stripe agreed to pay 4.4% of its own valuation for a business generating about 2% of its revenue, at roughly twice its own multiple.
That gap is not absurd, which is what makes the deal interesting rather than silly. To justify $7 billion at Stripe’s own 23.4x, OpenRouter needs about $299 million of fee revenue, which at a 5.5% take means roughly $5.4 billion of annual inference spend crossing the platform. The flow has to a bit more than double. A platform that quintupled weekly token volume in six months can clear that bar in a year without anything unusual happening.
Every AI toll business faces the same question about which layer keeps the money, and BMA has argued before that margin migration down the stack is the actual story under the bubble debate. Stripe’s answer is that the margin settles where the invoice gets written. It has been right about that once already, in cards.
The Risk
The revenue and the volume belong to two different companies stapled together.
Third-party analyses of OpenRouter’s public rankings put Anthropic at roughly 12% of token traffic and about 46% of platform spending. Chinese-origin open-weight models run the opposite way: a CNBC investigation found they held at least 30% of weekly enterprise token volume on the platform from February onward, peaking near 46% by mid-year, against a twelve-month average of 11% and about 4.5% in the first half of 2025. DeepSeek alone accounted for roughly 17.6%. US labs went from about 70% of token share in June 2025 to about 30% a year later. Cheap models move enormous volume and generate almost no fees; the fees concentrate in a handful of premium suppliers who have the strongest incentive and the easiest path to sell direct.
The sophisticated buyers are already leaving. Lovable published an engineering post in August arguing that the model picker is a dead end, said the control plane is the product, and disclosed that its own post-trained routing models now handle a meaningful share of production work. That is one of the fastest-growing inference buyers in Europe explaining why it will not pay a router. BMA covered the cost-of-goods logic behind Lovable’s Series C when the round closed. LiteLLM does the same job as open source, and the hyperscalers have folded routing into their own model services for free.
Owning the gateway also means owning its politics. A US regulated payments company now operates the largest neutral on-ramp to Chinese model inference for Western developers, at a moment when Washington is tightening rules on both AI exports and foreign model access. Neutrality was OpenRouter’s product. It is a harder promise to keep once the owner has money-transmitter licences in fifty states.
One counterweight deserves stating plainly, because it cuts against the easy bear case. Token prices are collapsing, so a percentage of dollars should be a shrinking business. It has not been. Revenue went from about $19 million to roughly $140 million over seven months while weekly volume rose fivefold, which means the mix shifted toward paid, enterprise and BYOK usage faster than per-token prices fell. If that holds, the deflation argument is wrong and Stripe bought early.
There is also the awkward precedent that customers of payment infrastructure do eventually shop around. OpenAI, Stripe’s most visible AI reference account and its ACP co-author, has been diversifying its payment vendors.
Quick Questions
Why would a payments company buy an AI model router? Because the two businesses are the same business. Stripe does not make the goods it charges for; it sits at the moment money moves and keeps a percentage. OpenRouter does not make models; it sits at the moment tokens get billed and keeps 5.5%. Stripe already supplied OpenRouter’s billing, tax and fraud tooling.
How does OpenRouter make money if it does not mark up tokens? It charges 5.5% when customers buy credits, with a $0.80 minimum and a 5.0% rate for crypto, and 5% on bring-your-own-key usage above $25,000 per month of list-price inference for pay-as-you-go accounts, or $200,000 for enterprise. The token price itself passes through at the provider’s published rate.
Is $7 billion for $140 million of revenue defensible? At about 50 times revenue it is expensive against Stripe’s own 23.4 times, and cheap against what Stripe would spend building the same position. The flow needs to roughly double for the price to look ordinary. Given the growth rate, that is a live possibility rather than a stretch.
What does Stripe get that it could not build? Distribution and data. Eight million developers, 400-plus models under one billing relationship, and daily settled-spend data on which models the market actually pays for. Stripe could write routing code in a quarter. It could not conjure the ledger.
Does this make Stripe a neutral party in the AI market? Not obviously. OpenRouter’s value rested on choosing models without a stake in the answer. Its new owner has commercial relationships across the labs and a payments business that benefits from certain flows more than others.
The Business Model Analyst Take
Stripe has done this exact trade before. Card networks and acquirers spent decades earning a small percentage of an enormous, slow-growing base. Stripe’s insight in 2010 was that the developer, not the bank, chose the rail. The company then spent fifteen years widening what counted as a payment.
This deal applies the same move to inference, and the trade is legible on one line. Stripe is swapping 36 basis points on $1.9 trillion for 550 basis points on $2.5 billion. One of those numbers grew 34% last year. The other one quintupled in six months. Stripe is paying 4.4% of its own market value to find out whether the second number keeps compounding long enough to matter, and it has structured itself so the answer arrives through Metronome’s meters rather than a research report.
The honest worry is not the multiple. It is that OpenRouter’s fee has never been defended by anything except convenience, and the customers who generate the fees are precisely the ones with engineers capable of removing it. Lovable already did. If Stripe’s answer is to bolt routing onto billing, settlement, wallets and agent checkout so that leaving costs more than 5.5% is worth, the price works. If it treats OpenRouter as a standalone product line, which billing engineers have argued is what happened to Bridge, it will have paid $7 billion for a rankings page. The rest of the Stripe business model has always depended on the pieces talking to each other.
Watch two numbers. Whether OpenRouter’s published fee schedule survives contact with Stripe’s enterprise sales motion, and whether Anthropic’s share of platform spending holds. The first tells you what Stripe thinks it bought. The second tells you whether there is anything left to buy.
