Monetization: the math behind a sustainable $9.99.

The standard structure has three tiers. Free covers the lowest-cost surface area and exists to seed word-of-mouth; Pro is the consumer-priced workhorse where you make your money; a Studio or Team tier captures the long tail of power users at 3–5× the Pro price. Designed well, Free is unprofitable per user but profitable per cohort, because Free users convert and refer.

A reverse trial inverts the usual flow. Instead of free → paywall, every signup gets the Pro experience for 7 days, then drops to Free. Superhuman, Krisp and Notion AI all report 2–3× conversion lift versus a flat free trial — losing the magic feels worse than never having it, and that loss-aversion is the conversion engine. The cost is one week of paid features per signup, which is trivial for cheap-to-serve software and expensive for heavy AI workloads. Size the trial usage cap accordingly.

ARPU is the blended monthly revenue per paying user. Gross margin is what survives after you subtract variable cost — for AI products that means inference, hosting and payment fees. Retention is how many months the average user stays before churning; for consumer SaaS, 14–22 months is the realistic band. LTV is the product of all three: ARPU × retention × gross margin. CAC is what you spend to acquire one paying user, all-in: ads, content, influencer seeds, tooling, the part of your founder time that actually moves users.

Two ratios matter. LTV:CAC ≥ 3× means the business has headroom to grow. Payback ≤ 12 months for consumer (≤ 18 for B2B) means you can reinvest revenue into more acquisition without burning a hole. Anything worse and you are subsidising your users — sometimes correct, but only if you know the bill.

The standard model assumes the median user is the average user. In a heavy-tailed product that's false: the top decile of usage often consumes 5–10× the median. Multiply that by the per-call cost of streaming STT, MT and TTS and the picture changes — a $10 subscription that gives unlimited inference is a put option the user can exercise against you. Once the cost per heavy user exceeds the price, you lose money the more successful the user becomes, which is the worst possible flywheel.

Three structural defences exist. First, gate truly expensive features (voice cloning, dual-stream dub) behind a higher tier whose price covers the worst case. Second, route cheap users to cheaper providers — every dollar of inference spend should have a cheaper substitute the moment quality permits. Third, set a soft fair-use cap high enough that 99% of users never see it but low enough that the long-tail abuser cannot exist.

Most consumer software at $10/mo is bootstrap-able to first signal. The AI infra cost on the first thousand paying users is a few thousand dollars per month — covered by revenue if your gross margin holds. Marketing in the early phase is content, Product Hunt, and influencer seeds, not paid acquisition. You raise when you hit a step-function expense that revenue cannot bridge: a second segment that needs a sales motion, infrastructure that needs SOC 2, or a geography that needs a local operator.

The discipline is to write the trigger before you have the option. "Raise at $X MRR" or "raise when we hire engineer #3" is a number a founder can defend; "raise when it feels right" is a number a founder regrets. Bootstrapping to clear PMF signal also doubles or triples your valuation when you do raise — investors price uncertainty, and revenue is the cheapest uncertainty-removal you can buy.