This article is about AI applications, not core foundation models. None of the initial 3 companies mentioned (Cursor, Lovable, Gamma) train there own models from scratch, nor do they need to. E.g. I get tons of value from Cursor, but I also still pay for ChatGPT Plus.
There is also enough competition in the core model space that these apps don't need to have an Achilles heel by being reliant on a single vendor. E.g. I think Cursor was smart to let you "bring your own API key".
I get that the article is about applications, but the applications have a dependency on 3rd party models, and that dependency is costing them most (all? More than all?) of their revenue.
Put another way, if I make $100M in annual revenue, but am paying out $110M to the API I wrap, it’s not nearly as compelling a business as that top-line $100M number makes it out to be.
In the previous generation of startups, expenses were mostly dominated by headcount, and the cost of actually delivering the service tended to be small. The story was “keep growing revenue, and if you need to show a profit, stop hiring.”
An AI startup built on other people’s models has to hope that the foundational models end up being fungible commodities, otherwise any margins you might gain will get squeezed out by your LLM provider. Alternatively, you can train your own model.
I don’t know what Cursor’s userbase looks like. If everyone is paying for Pro but using their own API key, that’s obviously a high margin business.
There is also enough competition in the core model space that these apps don't need to have an Achilles heel by being reliant on a single vendor. E.g. I think Cursor was smart to let you "bring your own API key".