Skip to content
Building an AI-Powered Product or ServiceLesson 6 of 13

6 · What it costs to run (tokens, subscriptions, unit economics)

An AI product has a cost every time someone uses it — and founders get burned by ignoring that. Before you price anything, know what one customer actually costs you to serve. That's your unit economics: revenue per customer minus what it costs to serve that customer.

How AI tools charge you (recap + the founder's lens):

  • Subscriptions (flat monthly). Predictable, but they stack — every "AI add-on" across your tools adds up.
  • Usage / pay-per-token (APIs). If your product calls a model's API, you pay per token — the chunk a model reads and writes, roughly ¾ of a word (F2). Both the input and the output cost tokens. Long prompts and long answers cost more on every single call — and you make that call for every customer, every time.
  • Per-seat or per-action. Some tools bill per user or per task.

Estimate your cost-to-serve BEFORE you price:

cost per use ≈ tokens per request × the model's per-token price (plus any per-use tool fees) monthly cost ≈ cost per use × uses per customer × number of customers

A feature that's "basically free" while three friends test it can become a real monthly bill at 500 customers. The math that felt optional in testing decides whether you make money at scale.

Control the cost like an owner:

  • Cap output length and keep prompts lean — you pay for both.
  • Use a cheaper/smaller model for easy tasks; save the expensive one for the hard ones.
  • Cache or reuse answers instead of regenerating identical work.
  • Audit subscriptions — cancel AI tools you stopped using; "I forgot I was paying" is a silent profit leak.

The SBA's startup-cost guidance is blunt about this: list your real one-time and ongoing operating costs so you can estimate profit and break even (SBA, n.d.). For an AI business, your per-use AI cost is an ongoing cost — it belongs in that math, not as an afterthought. NIST's GenAI Profile likewise flags resource and (literal) environmental cost as a real consideration of generative-AI use (NIST, 2024).

Check yourself. Why can an AI feature that's "free" in testing become a real monthly bill at scale — and what two numbers do you multiply to estimate your monthly cost-to-serve?

Sources

Sign in to track your progress