AI products are unprofitable per person, per month — visibly.
Inference cost turns pricing from an annual spreadsheet question into a per-user margin you can watch move.
Naru Pro shipped as an auto-renewing subscription metered at roughly two million AI tokens per user per month, which made per-user model cost a unit economics question rather than a line on an API bill.
Pricing a conventional app is mostly a question about volume. Servers amortise, marginal cost per user rounds to nothing, and the interesting numbers are acquisition and retention. Get the price wrong and you find out slowly.
Pricing an AI product is different in a way that is easy to say and hard to internalise: every active user carries a recurring inference cost that does not amortise. A heavy user is not a rounding error, they are a bill. Engagement — the thing every product instinct tells you to maximise — is also the thing that grows your cost of goods.
Putting a token ceiling on Naru Pro forced the question into the open before launch rather than after. You cannot set that ceiling without deciding what a user is worth, what a heavy user costs, and what happens at the boundary. That is a unit economics exercise wearing product clothes, and doing it early is the difference between a business model and a hope.
It also changes what failure looks like. A product that would once have been quietly unprofitable in aggregate is now unprofitable per person, per month, and you can watch it happen. That visibility is uncomfortable and it is the most useful thing about it — it is what let the decision to retire the product rest on arithmetic instead of mood.