Stripe Just Bought Its Own Merchant · Issue #221

Opening

Reader, on August 16th, Bloomberg reported that Stripe was acquiring OpenRouter. The price tag: over $7 billion — more than ₩9 trillion in Korean won. Most of the coverage read it as the payments company finally making its move into AI.

But open up the acquired company’s price sheet, and the story changes. OpenRouter doesn’t mark up token prices. What it actually earns is a 5.5% cut every time a user tops up credits. And the payment processor handling those top-ups? None other than Stripe.

Let me give you the conclusion up front. Stripe didn’t buy an AI company — it bought its own merchant. More precisely, it bought, whole, a toll booth that was already standing on top of AI spending.

The Price Tag That Grew 5.4x in 82 Days

Let’s start with the bare bones of the story.

OpenRouter is a company founded in 2023 by Alex Atallah, co-founder and former CTO of OpenSea. What it does is simple: with a single API key, it gives you access to over 400 models across more than 70 providers. Switch from Claude today to DeepSeek tomorrow, and you don’t have to rewrite a single line of code. It has 8 million users.

Look at the price trajectory, and you can feel the temperature of this deal. In its Series B this past May, OpenRouter raised $113 million at a valuation of roughly $1.3 billion. Alphabet’s growth fund, CapitalG, led the round, with Sequoia, Andreessen Horowitz, and Menlo Ventures following. Then, 82 days later, a price tag of $7 billion appeared. That’s 5.4x.

Set that against revenue, and the numbers get even stranger. Market data firm Sacra estimates OpenRouter’s annualized revenue¹ at around $50 million as of March 2026. That’s a sharp jump from $19 million at the end of 2025, but still only in the range of ₩70 billion in Korean won terms. Divide the acquisition price by that revenue, and you get a multiple of roughly 140x. Given that software companies typically trade at multiples of 10x to 20x, this clearly isn’t a price tag pinned to performance.

And this is where today’s real question begins. So what, exactly, was Stripe paying for?

5.5% Isn’t an AI Fee

Let’s take apart OpenRouter’s revenue structure. This company doesn’t mark up token prices. Whether you’re calling Claude or Gemini, it bills you exactly at the rate the provider publishes. “No markup” is even the company’s core sales pitch.

So where does it make money? When you top up credits. Load $100 onto your card, and the balance you can actually spend is $94.50. It takes 5.5% off the top. Fund with crypto and it’s 5.0%, but the minimum fee is $0.80, so if you only top up $5, the effective rate climbs to 16%. Even in BYOK² mode, where you bring your own API key, a 5% fee kicks in once you cross 1 million requests a month.

Look at what kind of structure this is. It never touches the price of the goods — it only skims a percentage at the gate where money comes in. And it takes that money upfront, holding it as a balance. This isn’t an AI company’s revenue model. It’s a payments company’s revenue model — more precisely, a prepaid-wallet operator’s.

There’s one more piece of evidence, and it comes from OpenRouter itself. On June 9, 2025, the company’s blog posted a notice about simplifying its fee structure. It contained this line:

We’re removing the flat $0.35 Stripe fee.

Which means the old fee formula was, literally, “percentage plus a $0.35 Stripe fee.” OpenRouter had been running its business with its acquirer’s name printed right on its own price sheet. What’s now a single 5.5% line used to be those two lines combined.

To sum up: what OpenRouter built over three years is model-routing technology, sure, but from a business standpoint it’s the payment gateway that AI spending has to pass through. 8 million people swipe their cards every month to top up balances, and agents draw down those balances token by token. And up until now, Stripe has been handling settlement at that gateway. It collected processing fees, but the 5.5% toll-booth take itself belonged to someone else.

Alex Atallah has long called his own company “the Stripe of AI.” That metaphor just reached its ending in the most direct way possible: an acquisition.

But Is This Toll Booth Sturdy

Before I push my argument further, let me set up the opposing case. This toll booth has a few structural cracks.

First, the fee layered on top of pass-through³ shrinks along with the model price. The fee base is total spend. AI model prices have kept falling for the past two years, and they’re likely to keep falling. Run the same job, and if spend drops, the absolute amount of that 5.5% drops too. The pricing analysis outlet UsagePricing reads OpenRouter’s announced plan to shift BYOK fees from usage-based to a flat subscription as a signal that the company itself knows this weakness.

Second, competitors are already calling this fee zero. Vercel’s AI Gateway advertises no token markup and no platform fee at all. Requesty charges 5%. In a market where rates are public, defending 5.5% means continuously proving the value of routing, fallback, and analytics.

Third, there’s the chronic churn problem gateways face. Once a customer grows large enough, it tends to contract directly with the provider and cut out the middle layer. OpenRouter attaching a fee even to BYOK mode is a device meant to convert that churn into revenue—but flip it around, and it’s also an admission that the churn is real.

Fourth, neutrality is this product’s entire value proposition, and now it has an owner. Not being locked to any one model is exactly why 8 million people use this service. Whether that neutrality survives the acquisition, and whether users will still believe it does, is a separate question.

Finally, this deal isn’t confirmed yet. Stripe’s position is that it “doesn’t comment on rumor or speculation,” and Bloomberg itself flagged that the final price could change. Demand for agent payments is still thin. The $20.57 billion that market research firm eMarketer projects for 2026 U.S. agentic commerce revenue is, after all, just a projection.

So today’s story isn’t “Stripe was right.” It’s closer to reading precisely what this company is betting on. And a caveat: this piece is commentary on industry structure, not grounds for an investment decision about any specific company or asset.

When You Line It Up With PayPal

The contours of this bet become clear when you overlay it with other news.

Right now, Stripe is pursuing an acquisition of PayPal alongside private equity firm Advent. In July, they offered $60.50 a share — about $53 billion — and got turned down; now they’re back at the table with a higher price. According to Reuters, the two companies would co-own PayPal on equal footing, with no plan to split it up. If it goes through, annual processing volume would hit $3.7 trillion.

What Reuters flagged as the real motive behind this deal is fascinating. Cutting reliance on Visa and Mastercard. PayPal has its own balance system and direct account-to-account rails that don’t run through card networks. As long as you’re on a card network, interchange fees⁴ flow into someone else’s pocket.

Now stack that against everything Stripe has bought up over the past two years. Bridge, the stablecoin settlement company. Privy, which runs 110 million programmable wallets. Tempo, a payments-only blockchain built with Paradigm. With OpenAI, Stripe rolled out a payment protocol for agents; at the Sessions event in April, it unveiled an agentic commerce product line — since expanded to Meta and Google — along with Link’s agent wallet. It also paired usage-based billing company Metronome with Tempo to demonstrate settlement that happens the instant value is transferred.

stripe

Do you see the picture now? Stripe is buying two doors at once. The door where humans open their wallets (PayPal), and the door where machines deduct from balances (OpenRouter). And at both doors, it’s laying down routes in advance to bypass the card networks.

For reference, Stripe processed $1.9 trillion in 2025 — 1.6% of global GDP. Where this company finds its next 1% — I think these two acquisitions are the answer.

Oswarld's Lens

I spend a lot of time designing pricing structures when building GTM strategy. And there’s a picture I habitually sketch first. I draw a line tracing where the money flows, then mark a dot on that line showing exactly where our fee attaches. Even for the same product, where that dot sits completely changes what kind of company you are.

OpenRouter’s dot doesn’t sit on tokens. It sits on deposits. Which is why I think this company should be classified from the start not as AI infrastructure, but as a payments business. And Stripe would have known exactly where that dot sat better than anyone — because it was the one processing the settlement for those very transactions. It had spent three years watching the traffic through someone else’s tollbooth, every single day.

Seen that way, the 140x multiple reads differently too. This wasn’t a purchase of a company with $50 million in revenue — it was the price of buying a single chokepoint through which AI spending passes. A chokepoint isn’t priced on the traffic passing through it now, but on the traffic that will pass through it later. So the success or failure of this deal doesn’t hinge on routing technology — it hinges on whether agents actually end up spending more money than humans do.

Nobody knows the answer to that yet. What I think Stripe is buying isn’t conviction — it’s optionality. If it doesn’t pan out, they lose ₩9 trillion (~$6.5 billion). If it does pan out and they failed to seize it, they lose a seat at the table for the next 20 years.

Bring this down to the practical level, and the implication shifts slightly. If you’re building an AI product, take a look at whether your cost sheet only lists the token price. Once you add gateway fees and foreign card conversion costs, the real burden sits 5-8% above what the sheet shows. The spot where Stripe paid ₩9 trillion is exactly that 5%.

Closing

Three takeaways.

First, OpenRouter’s revenue isn’t a token margin — it’s 5.5% of credit top-ups. The company itself announced it was dropping the “flat $0.35 Stripe fee,” which tells you this was a payment-processing structure from day one.

Second, $7 billion is roughly 140x revenue. That’s not a valuation on performance — it’s a price tag on the chokepoint.

Third, this deal has to be read together with the PayPal acquisition attempt. That said, nothing is confirmed yet, and pass-through fees shrink in lockstep with falling model prices.

If you try just one thing this week, pull up the AI service bill you’re currently paying and add up every line item that isn’t the token cost itself. That total is the real size of this deal.

💬 Have you ever switched from a gateway to a direct contract? Tell me in the comments at what spending level the math flipped in your favor. I’ll gather these cases and pick this up again in a piece on AI cost structures.

📨 If you have a colleague who digs into AI costs or payment infrastructure, please share this piece with them.

References & Further Reading

Primary sources

OpenRouter, “Simplifying Our Platform Fee”, June 9, 2025. ··· This is the decisive evidence behind today’s piece. The line “we’re eliminating the flat $0.35 Stripe fee” is right there. One pricing announcement, three months before the acquisition, explains the entire acquisition logic.
Bloomberg, “Stripe Finalizes Deal to Acquire AI Startup OpenRouter for Over $7 Billion”, August 16, 2026. ··· This is the original acquisition report. I’d recommend reading it alongside the caveat that the final price could still change.
TechCrunch, “Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+”, August 16, 2026. ··· This lays out the $113 million Series B, the 8 million users, and Atalla’s “the Stripe of AI” line.
Sacra, “OpenRouter revenue, valuation & funding” ··· This is the source for the estimated $50 million annualized recurring revenue. Worth reading with the caveat that it’s an estimate.
UsagePricing, “OpenRouter Pricing Blueprint”, June 10, 2026. ··· This traces the evolution of the fee structure through Wayback Machine snapshots. It’s where the sharpest version of the counterargument — “fees on top of pass-through cost shrink together with the underlying model cost” — is laid out.
Engadget, “Stripe is reportedly in talks to buy PayPal”, August 15, 2026. ··· This is where you’ll find the $60.50 per-share figure, the $3.7 trillion in annual processing volume, and the passage on reducing dependence on Visa and Mastercard.

Background

Stripe, “Stripe’s 2025 Annual Letter”, February 2026. ··· This contains the $1.9 trillion in processing volume and the five-stage breakdown of agentic commerce. If you want to see how far this company itself is mapping this out, the original letter is the fastest way in.
Fortune, “Visa to expand card partnership with Stripe’s Bridge to over 100 countries”, March 3, 2026. ··· This covers Bridge’s card expansion. Particularly worth reading is the passage where Bridge’s CEO says “agent transactions have a different texture than card-network transactions.”

Related past issues worth reading alongside this one

The Cloudflare Issue That Gave Bots a Name and a Wallet ··· This is the earlier chapter of the same story. That issue was about giving agents a name; today’s is about who takes a cut when that name is used to pay.
Agents Think in GPUs but Work in CPUs ··· This looks at where agent spending flows from the infrastructure side. Today’s 5.5% is a number attached to that total spend.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

Annualized Recurring Revenue (ARR): A figure derived by multiplying the most recent month’s or quarter’s revenue by 12 or 4 to project a full year’s total. It shows a fast-growing company’s current pace, but it isn’t the actual amount earned over the past year. ↩
BYOK (Bring Your Own Key): A setup where you use an intermediary platform but pay the underlying model provider directly with your own account. You pay a toll, but you fill your own tank with your own card. ↩
Pass-through: A billing method where the underlying cost is passed on without markup. Since revenue instead comes from a separate fee, when the underlying cost falls, the base on which that fee is calculated shrinks along with it. ↩
Interchange fee: The fee that flows from the merchant side to the card issuer whenever a card payment occurs. It’s unavoidable as long as a transaction runs through the card network, which is the single biggest motivation behind attempts to route around card networks altogether. ↩