The startup Harvey, valued at $15.6 billion, saw its gross margin drop from 50% at the start of 2026 to -50% by June. This decline was driven by a 20-fold increase in token traffic, causing expenses for third-party AI models to exceed service revenue.
In August, Harvey launched its own model based on the Chinese Moonshot Kimi K3, which restored positive margins. While the company still utilizes Anthropic Opus for complex tasks, the shift reduced reliance on external providers. Other startups are following suit, with Abridge developing a model on Nvidia open weights and Decagon routing 80% of requests through its own systems.
Fintech firm Ramp, which raised $750 million in June, is now exploring training its own model, a move supported by venture capital firms Sequoia Capital and General Catalyst. CEO Karim Atiye noted that such development has become increasingly viable compared to a year ago.
The market landscape is shifting as major model providers face new challenges. Anthropic has highlighted risks regarding startups building their own models, while Mistral CEO Arthur Mensch warned clients about dependency on closed providers. Recent conflicts, such as OpenAI terminating its contract with Cursor, underscore the instability of relying on external AI services.
Despite the trend, experts remain cautious. Matt Kraning of Menlo Ventures argues that the choice between building or buying a model is a false dilemma, suggesting that developing an internal model may sometimes serve as a mere imitation of true independence rather than a strategic solution.
