Legal AI startup Harvey just showed why renting frontier models can wreck your margins.
Harvey's gross margin fell from about 50% at the start of the year to minus 50% by June, as customer usage of its AI agents spiked and pushed up its model costs. The company returned to positive margins only after releasing its own model, built on top of Moonshot AI's open-weight Kimi K2. Abridge, Decagon, Ramp and Rogo are reportedly making similar moves, shifting away from paid third-party APIs toward open-weight models they can run themselves.
It's the clearest sign yet that leaning entirely on frontier-model APIs doesn't scale as a business model once usage grows. Open-weight models like Kimi K2 let startups cap their single biggest cost line, even if it means taking on the engineering work of running inference themselves.
Expect more self-styled AI-native startups to quietly become AI infrastructure companies too, whether that was the plan or not.