Classic SaaS trained a generation of founders to treat gross margin as a solved problem. Hosting was cheap, support was a ratio, and the 80-point margin was a personality trait. AI-native products broke that habit. The cost of goods sold now moves with every user action, every retry, and every time a human has to catch the model being confidently wrong.
Investors still want software multiples. The P&L is starting to look like a hybrid of software and services — except the “service” is a GPU bill that does not sleep, plus a review queue that does.
Three cost lines that matter
Inference. Caching, smaller models for easy tasks, and not sending the entire history on every turn are the difference between a toy and a business. If your architecture is “always call the biggest model,” you do not have architecture. You have a demo.
Evaluation. Quality is a cost center. Offline evals, production traces, and the people who read the traces. Skip this and you will “save” money until a customer files a ticket that is really a lawsuit.
Human review. The dirty secret of many “autonomous” products is a queue in Manila, Austin, or Eastern Europe. That is fine if it is designed. It is fatal if it is a surprise at 40% of tickets.
Price the 95th-percentile conversation, not the median demo.
Pricing has to follow the cost shape
Seat-based pricing on a variable-cost product is how you fund your power users with your idle ones until the power users eat the company. Usage tiers, hard caps, and product design that makes the expensive path rare are not “growth hacks.” They are survival.
The winning teams instrument contribution margin per workspace the way 2015 companies instrumented net retention. If a customer is beloved and negative, they are not a logo. They are a leak.
What “good” looks like
A path to 70%+ gross margin after the product has routing, caching, and a review policy — not on day one. A sales team that is not allowed to waive usage caps without finance. A roadmap that treats model spend as a first-class constraint, the way mobile teams treated battery.
The companies that fail will look busy. They will ship features. They will have impressive demos. They will not know, to the dollar, what a conversation costs. In this market that is not a rounding error. It is the business.