Granted payments and associated ledges require a level of consistency that other systems do not require, but that is orthogonal to high volume.
(Not picking on you here, you just provided a well-written peg for a popular narrative. I'm aiming to sharpen my own thinking here & perhaps learn something.)
This seems like a stretch given the rise of local inference, especially the Prism Labs rumors from a few weeks ago.
One way I think about LLMs is they are akin to fancy databases in that they are software of which you can ask questions and get answers if you ask properly. Oracle & SQL Server are akin to OpenAI and Anthropic, and there are analogues for MySQL, PostgreSQL, SQLite, MongoDB, PlanetScale, etc. (This is an analogy, it's not going to be a perfect fit.)
In that view, would it make sense for someone to say that their credit card processor just bought the company that makes their ODBC driver? Would anybody suggest that the TPS of their RDS instance is a lightweight asset?
I don't see this as strategic beyond the obvious idea that Stripe wants to get closer to AI, and they haven't been able to get the market to care about their natural linkage to AI (Radar).
> they can start hosting their own models and competing as an AWS for tokens
The financial pressure of doing this has caused cuts to core product teams at the richest companies (which Stripe isn't!). I would not want my payments processor to go down this road and to get worse at processing payments.
Stripe is not a "credit card processor", and that's not the point of OP's comment. The point is that being the intermediary between merchants and processors is the valuable expertise in terms of an OpenRouter acquisition.
It's been a few years, but I used to work on almost exactly this (not at Stripe, but we processed hundreds of billions annually). If you're a big enough merchant (or Stripe itself), you can choose to send a credit card transaction to any one of potentially many processors, each of which have their own performance characteristics (one may approve at a higher rate, the other may charge better fees, etc). All of this subject to attributes of the transaction itself (ticket size, geography, card type, many others). You know quite a bit about the transaction itself before you send it out, so you can build up routing knowledge to optimize for whatever thing you care about (usually transaction success rates or fees).
See how this starts to look a lot like OpenRouter with money instead of tokens? I'm not sure I 100% believe that's how it'll shake out, but there is a transferable skillset.
I think the "money instead of tokens" is the important part. Money and tokens are fundamentally different was part of my argument (that I did not make well). It's not clear that it is a good thing for money movers to get into the token business (this apples to Ramp as well).
Banks are also intermediaries between parties (at scale, really between any parties). One could use similar logic to say that it therefore makes sense for Wells Fargo to start buying homebuilders because the home buyers will eventually be paying Wells anyway.
BankRate similarly processes volume of consumer mortgage quotes in real-time communication with lender APIs. They also do not have a reason to own a token router, even though their business involves similar processes to what you outline for Stripe.
Being an intermediary in a value chain does not mean you're critical path, or that it makes sense for you to be in the critical path.
Funnily enough, you argue that this is a natural fit for Stripe while a peer reply argues that it's a change in the business a la Amazon->AWS.
Anyway, I appreciate your thoughts.
Would is make sense to say that their online bookstore now sells Ethernet cables, bidets, and delivers groceries? Sells _cloud infrastructure_??
Amazon made two transitions:
1. Amazon the online bookseller => Amazon "The Everything Store"
2. Amazon the online retailer => Amazon the Cloud Services company
I believe what Stripe is doing here is closer to (1). "We are good at high-throughput APIs that wrap complexity with thin margins. We did it for credit cards, then ~all payment methods, now other digital bits."
In this context, tokens are much, much easier than international payments.
Now: Is it outlandish for their customers?
Not really. Every engineer knows stripe as an engineering company. I don't think Stripe is what it once was, but it's certainly a generational company. You're asking engineers—who broadly have a positive impression of Stripe—to use this product they already know. OpenRouter gets the positive brand association (trust) of Stripe; Stripe expands into a new domain whose technical needs are extremely similar.
re: the Amazon transitions, the first shareholder letter lays out the plan to go beyond books. As I remember it, books were always only supposed to be the entry point.
AWS transition was more around building the platform Amazon.com needed to grow, and also to monetize the same platform.
Neither of these really fit with Stripe.
Stripe can obviously operate OpenRouter, they have the tech skills. The risk to the core business is that OpenRouter's growth path will distract from the core financial business and/or require a very different capital stack. (Someone already suggested Stripe scale out first-party model running, which can get very expensive.)
Appreciate your taking the time to respond.
Stripe already has killer APIs and knows how to build API infrastructure at Extreme scale, it still has developer cred after all these years, they are already Kings of multiplexing to different sources -- so why not just build your own LLM gateway?
My only guesses are: they want to buy the traffic instead of trying to organically grow it (but I feel like Stripe would get the volume anyway because of name rec), they want to buy the revenue and think they can dramatically reduce COGs because of efficiencies, buying the relationships with the frontier labs (who all dogfood their LLMs in pseudo-private on OR), or... something else?
That's the part I'm confused about.
I’m sure payments are convoluted, but I’d still imagine they could be meaningfully easier for the bulk 80% of use case?
I worked on this at Stripe in 2022. We were the first teams to start building v2 APIs and data models to solve exactly this problem. The first target launch date (in Feb 2022) was November 2022. It was launched in May of 2025.
My reporting line, as an EM, was Netflix, Oracle, Oracle. No one had startup experience. It was drenched in politics. The engineers were largely brilliant, kind, and hardworking.
I still love the company and believe in Patrick. Believe me, he deeply understands what you're saying and wants it to be the best it can be. But it was clear to me, even then, that they'd lost a lot of what made them special. They could maintain it, but I wasn't sure they could do it again. Banking-as-a-Service was one opportunity, Link was another, and now this will be a third. We'll see. (I say this with a lot of love for Stripe and Stripes.)
Coincidentally, I had a conversation with a recruiter at Anthropic and saw them doing something very similar. They were starting a new team in a new vertical and wanted someone with experience running an org of 100+ people. I would bet real money that it will be a fraction of the product/impact it could be (though still probably make money!)
Stripe is a middleman. So is OpenRouter. So is OpenCode. Unless you own the data center and the hardware, how cheap can a middleman's tokens really be compared to the hyperscalers, without massive model compression?
Even DeepSeek itself is raising token prices. How much margin is there for a middleman like Stripe buying tokens in bulk from a data center and reselling them? Last i check Stripe do not run or own physical data center
I doubt Stripe can do much better here.
1.) LLMs are useful for programming
2.) Open models are excellent and will continue to improve
3.) Economies of scale and ease of access mean self hosting is out of the question for a large number of users
This means that even if the largest labs are not worth trillions and a large amount of the data center build out is not as valuable as the builders project and GPU/RAM prices plummet, it will not matter at all for this business. People will want to buy cheap open source tokens from a centralized trusted provider.
$7 billion for a business with little overhead that is already within their core competency and has strategic growth potential seems like a very good deal.
This seems like the kind of thing that you can have Claude write in an afternoon for whatever service you're running. I don't see the value.
people don't want to maintain infra (ie adding new models all the time)
people want their queries to work without thinking
if a provider goes down, openrouter queries dont (ideally)
it's really not that complex to understand
Broadly I agree with you. It seems like they have an in-demand product and there could be a sustainable business there at least in principle. But whether it's a $7B business or a $70M business I can't say.
openrouter is involved with a lot of scammy crypto personalities. maybe they are the tulips people have been manic about for too long. maybe the collisons are the tulips.