Startups are accelerating efforts to develop their own models instead of using those from companies such as OpenAI and Anthropic. Analysts say the shift reflects a combination of cost-cutting, a push to reduce dependence on large AI model developers and a growing need for stronger control.

Legal AI startup Harvey, valued at $15.6 billion, launched its own AI model, Harvey Tenet, in August based on Chinese company Moonshot AI's open-weight model Kimi K3. Reliance on OpenAI or Anthropic models appears to have influenced the decision, as it is seen as making it difficult to improve profitability.

After Harvey updated its AI agent in March, customer usage surged, but gross margin fell sharply, Bloomberg News reported recently. Gross margin, which was about 50 percent at the start of the year, fell to minus 50 percent by June, Bloomberg reported, citing an official.

Gross margin is the percentage of profit left after subtracting the cost of making a product from total revenue. A drop from 50 percent to minus 50 percent means losses grew as sales increased.

Harvey adjusted how it uses AI, and gross margin turned positive again, Bloomberg reported, citing sources.

Healthcare tech startup Abridge is also developing its own foundation model tailored to clinical work based on an open-source Nvidia model. AI customer support startup Decagon has entered a stage where it handles 80 percent of all inquiries through its in-house model.

In fintech, startups such as Ramp and Rogo are reviewing training their own models. Ramp has been considering developing its own model since it raised $750 million in investment in June.

In the coding space, Cursor, now under SpaceX, and Cognition, valued at $48 billion, introduced customised models.

Design software company Canva is also switching to small models from OpenAI and Anthropic and to open-source models as part of cost-cutting, while moving to develop its own model. Cliff Obrecht (클리프 오브레히트), a Canva co-founder, said that before adopting AI, the cost of servicing free users was very low, but costs rose sharply after AI was introduced, changing the earnings structure. He said the company is now moving into a phase where it must prove profitability or a path to profitability from its AI business, The Information reported.

Developing AI models has been an area that startups could not easily attempt because it required large-scale funding. Now, using open-source or open-weight models has made it realistically possible for startups to develop customised models. That is a key reason more startups are developing in-house models.

Karim Atiyeh (카림 아티예), Ramp's co-CEO, said, "Even a year ago, developing our own model did not make any sense, but now it is increasingly becoming something worth trying."

Venture capital firms that fund startups, including Sequoia Capital and General Catalyst, are also supportive of the trend. By developing in-house models, companies can cut one of their biggest costs while strengthening control over their own technology without relying on OpenAI and Anthropic, Bloomberg reported.

Some investors still keep their distance from the in-house model development trend. Matt Kraning (매트 크래닝), a partner at Menlo Ventures, which invested in Anthropic, said, "It is not suitable for every company to develop a customised model. Building your own model requires specialised staff and also comes with higher upfront costs. For some companies, developing an in-house model is merely a marketing tool."