The argument in brief· 3–5 min read· full book 30 pp
How AI compute gets priced, metered, and paid for.
A token is a meter reading, not a unit of account. Every AI credit that ships is a per-model conversion layer over dollars — which is why the position worth owning is the exchange desk rather than the mint.
In August 2026 Stripe agreed to buy OpenRouter for a price it never disclosed. Its chief executive explained the purchase in one line:
“Tokens are the central currency for companies building with AI.”
It is a good sentence. It is also contradicted by the documentation of the company Stripe was buying, which answers the only question that matters — in what unit is the balance held? — without ceremony:
“OpenRouter uses a credit system where the base currency is US dollars.”
Not tokens. Dollars. And the reason is measurement rather than caution. A unit of account has to mean the same thing to both parties at the moment they agree to a trade, and a token does not. Providers serving the identical open-weight model differ threefold in price. The router picks between them with a draw weighted by the inverse square of cost, so roughly one request in ten still goes to a costlier supplier. A cache write can cost more than an ordinary input token, and whether that adjustment is a discount or a surcharge depends on which vendor you asked.
Which means the price of a request is not a number when you press send. It is a distribution, and it does not resolve until the work is done. You cannot quote a fixed price for a request before it is routed — not because the tooling is young, but because there is no fixed price there to quote. The clearest admission of what that costs in practice sits in the documentation of LiteLLM, the open-source gateway thousands of teams use to track spend, under a heading it does not euphemise:
“Cost does not match your provider bill?”
So the interesting question is not why nobody minted a token currency. It is what the people closest to the problem built once they worked out they could not. The answer is already shipping, and almost nobody has noticed it.
When Claude is sold through the AWS and Azure marketplaces, Anthropic does not bill in tokens. A cloud marketplace needs one countable line item and cannot express “1.25x base input price for a five-minute cache write,” so the vendor rates the usage in dollars and then invents a unit to report:
“rates your token usage in USD at standard per-model, per-feature rates, applies any negotiated discount, converts the result to CCUs at $0.01 per CCU.”
A Claude Consumption Unit is worth a penny. Note the direction of travel: usage is priced in USD first, then converted. The dollars are not derived from the units. The units are derived from the dollars — which is the whole argument of this book, running in production as a billing integration, for reasons that have nothing to do with any acquisition.
Stripe’s own product does the same thing one layer up, letting a business define an “AI Credit” and configure how each model’s usage converts to it. That verb is carrying the argument. A credit is a synthetic unit whose exchange rate against real tokens is set per model and changed at will by whoever sells it. The complexity does not vanish; it moves behind an interface, where one business absorbs it so its customers never look at it.
It presents the strongest case against its own thesis — that routing is meant to become worthless, and that a buyer can afford zero routing margin because it monetises billing, tax, fraud and settlement around it — with no rebuttal appended, because the research could not build one.
It also marks its own limits: the mechanism is documented, the magnitude is not measured, and the regulatory section stops at a federal rulemaking.
30 pages. 38 sources captured and hashed, 77 claims bound to exact quotations, each machine-checked against a stored copy before the sentence citing it was written. Every source is archived alongside it, so you can check any claim yourself.