Unit economics when your cost of goods thinks

For thirty years, software pricing rested on one fact: serving the next customer cost approximately nothing. That single fact produced every instinct we now treat as obvious — unlimited plans, free tiers with no ceiling, land-and-expand, growth first and margin later.

If your product calls a model, that fact is no longer true. Your cost of goods thinks, and thinking is metered.

Most AI products are still priced by founders running the old playbook against the new cost structure. Here's what actually changes.

"Unlimited" is now a position, not a default

Unlimited used to be free to offer, so everyone offered it. Now it's a deliberate bet that your heaviest users won't be heavy enough to matter.

Sometimes that bet is right. Usage in most products follows a power law: the median customer uses far less than the mean, and the mean is dragged around by a handful of outliers. If your ceiling is low enough and your price high enough, unlimited is a marketing gift that costs you a rounding error.

But the bet has a specific failure mode, and it isn't the enthusiast. It's the customer who works out that your product is cheaper than the alternative at their volume — and tells people like them. Word of mouth among heavy users is exactly the growth you want and exactly the growth that kills an underpriced unlimited tier. You get punished by the mechanism you were hoping for.

Your gross margin is a variable now, and it moves

Traditional SaaS gross margin sits at 75–85% and barely moves. Yours moves with:

• Model choice. A more capable model can cost several times a cheaper one for the same task. That's a margin decision wearing an engineering costume, and it usually gets made by whoever picked the default.
• Usage mix. Two customers on the same plan can differ by an order of magnitude in what they cost you.
• Provider pricing. Which changes, in both directions, without asking you.

The practical consequence: you have to measure cost per unit of work, per customer, continuously — not once, during planning. If you can't answer "what did this customer cost us last month," you don't know your margin. You know your revenue, and those stop being the same thing the moment serving costs money.

Overage is the honest mechanism

The cleanest structure for a metered-cost product is the boring one utilities have used forever: a plan with a generous included allowance, and a per-unit price beyond it.

It works because it's legible. The customer knows what they get, heavy users pay more, and you can't be farmed. And the incentive points the right way — a customer using more is a customer getting more value, so charging them more isn't extraction, it's the deal working as described.

Two rules keep it fair rather than punitive. Set the allowance so the large majority never reach it, because a limit people keep brushing against generates support tickets and resentment in equal measure. And price the overage above your marginal cost, not at it — a unit price that merely recovers cost means growth in usage adds no margin at all, which is the same as not charging for it.

Three ways founders get this wrong

Pricing against a competitor whose cost structure is different. If their marginal cost is zero and yours isn't, matching their price means accepting a worse business. That comparison isn't a benchmark, it's a trap.

Racing to the bottom on a metered product. The bottom of a market where every customer costs real money to serve is a place where volume makes things worse. Growth accelerates losses instead of amortising fixed costs — the exact inverse of the instinct you inherited.

Treating the model bill as an infrastructure line. It's cost of goods. Put it above the gross-margin line where it belongs and the pricing decision starts making itself.

The number to know

One number tells you whether the business works: what it costs to serve one customer for one month, at the usage of your median paying customer.

Divide that into your price. If the answer isn't comfortably above 4×, you don't have a pricing problem to fix later. You have one to fix now, before volume makes being wrong expensive.