Growth used to be the scoreboard in software. In the AI era, it is table stakes, and the metric that actually separates businesses from bonfires is gross margin.
The pattern is visible across the datasets we have covered this week. Bessemer’s State of AI puts its fastest cohort, the Supernovas that reach $100 million ARR in about 18 months, at roughly 25 percent gross margins, with sometimes fragile retention. Its steadier Shooting Stars run near 60 percent. And the standouts in health AI clear 70 percent with software-like economics. Same revenue milestone, three completely different businesses.
Why AI margins run thin
Classic SaaS sold software that cost almost nothing to serve. AI products carry a cost of goods that scales with usage: every request burns inference. When the product is a thin layer over a foundation model, the margin belongs to the model provider and the cloud, not the application. Thicker products, with proprietary workflows, data advantages and the ability to route work to cheaper models, keep more of it.
The questions that now matter in diligence
Three replace the old growth worship. What is the fully loaded inference cost per dollar of revenue, and is it falling? Does retention hold when the novelty fades? And can the company swap underlying models without rewriting the product, the application-layer version of the lock-in problem enterprises face? Margins answer whether a company owns its economics or rents them.
The margin playbook, for builders and buyers
For companies building AI products, the levers are concrete. Route aggressively: most requests do not need the frontier model, and the spread between frontier and mid-tier pricing is wide enough that intelligent routing alone can move gross margin by double digits. Cache what repeats: providers price cached input tokens at a fraction of fresh ones, which rewards products architected so that context is reused rather than resent. And own the workflow, not just the wrapper: proprietary data, integrations and process depth are what let a product charge for outcomes while paying for tokens.
For enterprise buyers, the same scoreboard reads in reverse: a vendor’s margin structure predicts its behavior. Thin-margin vendors facing usage growth have two options, raise prices or degrade quality by quietly routing to cheaper models, and contracts should anticipate both, with model-version transparency and price-protection clauses. The pricing moves we covered at Cursor and OpenAI’s Codex are both vendors repairing unit economics in public, and the direction of travel is uniform: consumption pricing that passes cost risk to the customer. A buyer who understands the vendor’s margin math before signing negotiates from the only position that matters.
What to watch
Watch whether falling inference prices flow to application margins or get competed away, and whether the first Supernova to IPO gets priced like software or like a reseller. That listing will set the multiple for the whole cohort. Two recent moves show the margin question playing out in real time: Cursor pricing its in-house coding model at a tenth of frontier rates, and OpenAI shifting Codex to consumption billing, both bets that owning unit economics beats subsidizing usage indefinitely.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
