JUDGMENTLast week Yuanwang said AI rehiring is the opening price of a new labor market. This week I want to take apart something else: nearly every founder is lured in by the same line "we have users, we have revenue," yet most AI apps are simply not good businesses. The problem is not product or traffic. It is that the first line of the business model is wrong from the start: every AI app is, at bottom, selling tokens, and selling tokens is a business whose margin the upstream model layer has locked down through pricing power.

FACTFour datasets sit side by side. One: Bessemer studied 20 high-growth AI companies and found many hit close to $100M ARR in their first year of commercialization, yet average gross margin was only about 25%; mature cloud software runs around 70%. Two: Cursor is the oft-told "myth": in the quarter ending January 2026 its gross margin was -23%, about -31% once free-tier inference costs are counted — the more it sold, the more it lost; even at $4B ARR it could not fundraise independently and was ultimately bought by SpaceX for $60B, at its core an arbitrage of the "compute + scenario" loop. Three: Perplexity reported 60% gross margin, but only after booking ~$33M in free-trial costs as R&D; counted as COGS, the margin is negative. Four: Stripe's data — the top 100 AI companies took a median 11.5 months to reach $1M annualized revenue, but "reselling tokens, being a model distributor" ramps fast and churns far worse than traditional SaaS. All sources are public.

SPECULATIONI call it "AI apps' triple lock." Lock one, cost: every use incurs inference cost that does not decline with scale, so margin stays pinned low; being validated and being sustainable are two different things. Lock two, growth: growth itself drags the product toward losses (Kuse exploded when it handed free credits; OiiOii deliberately curbed growth; office-agent margins run as low as -200%). Fireworks' CEO said it outright: scaling to bankruptcy — the more you grow, the closer you get to death. Lock three, and the cruellest: pricing power is locked — the upstream model maker is simultaneously your supplier and your competitor, able to raise prices (Anthropic's Priority Tier), to cut supply (Claude pulled from Windsurf), and to ship its own tools (Claude Code). Cursor's share fell from 41% in June 2025 to 26% in May 2026 as upstream self-operated tools siphoned users away. Under this triple lock, "users and revenue" is exactly the most dangerous signal — it gives founders the illusion that lets them sink deeper into a wrong model.

OPINIONNilu looks past funding stories and watches the one basic account: when your biggest cost item, your largest supplier, and your strongest competitor are all the same entity, that is not a business; it is laboring for the upstream. But the inverse is the only structural opportunity of this round — not to fight the model layer for token distribution, but to seize the "workflow entry point" and "proprietary data" that the model layer cannot provide: the former is Cursor building Origin against GitHub, the latter is sinking a vertical scenario's user data into an asset models cannot replicate. The billion-dollar acquisitions are never buying "companies that sell tokens"; they buy a position that holds the entry point and the data, forcing the model makers to come to the table. See this, and you will not still be crowded at the token-selling table in September 2026.

---

Further Reading