Martin Casado published a great piece on Stripe’s acquisition of OpenRouter, arguing that LLM tokens have become a new, universal medium of value exchange.
I agree with the larger thesis: OpenRouter + Stripe could become a critical piece of AI’s economic infrastructure.
But I think the monetary analogy gets one important thing backwards.
Tokens aren’t money. They’re what customers buy with money.
And that distinction points to something potentially much bigger about why Stripe wants OpenRouter.
Tokens aren’t really a medium of exchange
Consider a normal AI transaction:
The token is the thing being purchased, not the thing being used to purchase something else.
That’s fundamentally different from money.
A business receives dollars from a customer, then uses those dollars to pay AWS, employees, suppliers, or another business. The dollars keep circulating through the economy.
LLM tokens generally don’t.
If I buy 10 million tokens from an AI provider, I can’t take those tokens and pay AWS with them. I can’t use my Claude tokens to buy inference from OpenAI. And the model provider doesn’t receive my tokens and spend them somewhere else.
So tokens don’t behave much like a medium of exchange.
They behave more like a metering unit for a resource.
Think kilowatt-hours for electricity, GPU-hours for compute, or gigabytes for bandwidth.
A factory buys electricity measured in kWh and turns it into something valuable. But we wouldn’t say:
“Kilowatt-hours are the new dollars.”
AI tokens work similarly.
Electricity becomes compute. Compute becomes inference. Inference produces tokens. And those tokens represent consumption of machine intelligence.
That’s already economically important. We don’t need to turn tokens into money to make it transformative.
Tokens aren’t even a universal unit of value
There’s another problem with the dollar analogy.
A dollar is fungible.
$1 = $1.
A token isn’t.
One million tokens from a small open-source model and one million tokens from a frontier reasoning model can have dramatically different prices and produce dramatically different amounts of useful work.
Even the underlying measurement differs across tokenizers and models.
Most importantly:
One token does not represent a fixed amount of intelligence.
A weak model and a frontier model can both generate 1,000 tokens. The quantity tells us very little about the value of the result.
So I think the more useful framing is:
Tokens are becoming a standardized metering layer for machine intelligence, not a new form of money.
And once you make that distinction, the Stripe–OpenRouter acquisition actually becomes more interesting.
OpenRouter is building a market for intelligence
An application today might have hundreds of models available to it.
Some are faster.
Some are cheaper.
Some are better at coding.
Some are better at reasoning.
Some are good enough for 90% of requests at a fraction of the price.
The application increasingly shouldn’t care which company produced the model. It should care about getting the right amount of intelligence, at the right quality, latency, and price, for the task.
That’s where OpenRouter becomes interesting.
It can sit between demand and hundreds of suppliers of machine intelligence:
Application
↓
OpenRouter
↓ ↓ ↓
OpenAI · Anthropic · Google · xAI · open models · future providersOpenRouter can route demand while abstracting away much of the complexity underneath.
That starts looking less like a payment network and more like a market and routing layer for intelligence.
Stripe brings something different.
It can turn all of that consumption into money.
Stripe can connect metering directly to money
Imagine an AI application uses:
- $0.003 of one model
- $0.14 of another
- $0.02 of another
- $0.001 of an open model
OpenRouter can decide where the request goes and measure the resulting consumption.
Stripe can turn those microscopic consumption events into pricing, billing, credits, taxes, fraud detection, and ultimately settlement.
The stack starts to look like this:
OpenRouter → resource selection + routing
↓
Metering → consumption + pricing
↓
Stripe → billing + financial settlementThat is already a powerful infrastructure stack.
But I think there’s a bigger reason for Stripe to own it.
Stripe traditionally enters after the buying decision
Historically, Stripe sits relatively late in a transaction.
A person decides they want something.
They choose a product.
They choose a merchant.
Then Stripe helps move the money.
In simplified form:
Human decides what to buy
↓
Merchant
↓
Stripe
↓
Money movesStripe has built an enormous business by owning infrastructure around that last part.
AI agents change the sequence.
Increasingly, software itself can decide:
Which model should I use?
Is the better model worth 5x the price for this task?
Should I retry?
Should I buy more compute?
Which database should I query?
Which API should I call?
At that point, software isn’t merely initiating a payment somebody already decided to make.
Software is making the economic decision.
And OpenRouter already operates in one of the first markets where this is happening at enormous scale.
The important new primitive may be the machine buyer
Today, the machine buyer is mostly purchasing inference.
An application has a task, evaluates available resources, chooses a model, consumes some intelligence, and incurs a cost.
Tomorrow that decision could extend far beyond models.
An agent might choose between:
- Compute providers
- Databases
- APIs
- Data sources
- SaaS products
- Financial products
- Logistics providers
And eventually physical goods and services.
The flow becomes:
Agent has a goal
↓
What resource do I need?
↓
Which provider should I use?
↓
At what price?
↓
Consume
↓
Meter
↓
PayThis is a fundamentally different kind of commerce infrastructure.
Stripe no longer has to enter only after the purchasing decision.
With OpenRouter, it can potentially move upstream into the infrastructure where machines choose what to buy in the first place.
That may be what Stripe is really buying
This is why I think the Stripe–OpenRouter combination is more interesting than “tokens are the new dollars.”
Stripe doesn’t need tokens to replace dollars.
Dollars can remain the settlement layer.
Tokens can remain the metering layer.
OpenRouter can become the resource allocation and routing layer.
And agents can increasingly become the buyers.
Put those pieces together and you get:
The strategic opportunity isn’t creating a new currency.
It’s owning more of the economic stack around autonomous software.
Stripe already built infrastructure for an internet where humans decide what to buy and software moves the money.
The next internet may have software making more of those decisions itself.
If that happens, payments are only the final step.
The much more interesting question is:
Who owns the infrastructure between a machine deciding it needs something and the money eventually moving?
OpenRouter may give Stripe its first answer.
Tokens aren’t the new dollars.
The machine buyer is the new economic actor.
