Artificial Intelligence
Startups train in-house AI models
Bloomberg reported that Harvey’s margins swung sharply as usage of its AI agents rose, and that the company and others have since shifted towards open-weight or in-house models to cut costs and dependence on frontier labs.
- Bloomberg reported that Harvey’s gross margin fell from about 50% at the start of the year to minus 50% by June after customer use of its AI agents increased.
- Harvey released its first in-house model in August, post-trained on Moonshot’s Kimi K3, and people familiar with the work said margins turned positive again after that launch and other changes.
- Abridge was building a clinical model on Nvidia’s open weights, Decagon said it routed 80% of customer queries through its own models, and Ramp said it was weighing model training for the first time after raising $750M in June.
- Sequoia Capital and General Catalyst were funding the shift, and Rogo was also making a similar move.
- OpenAI moved to end Cursor’s contract after SpaceX bought the company, while Legora said it had reached $100M in revenue in 18 months and served more than 1,200 firms.
Bloomberg reported that Harvey’s gross margin fell from about 50% at the start of the year to minus 50% by June as customer use of its AI agents rose. The company was valued at $15.6B and said its token usage had increased twentyfold this year.
Harvey released its first in-house model in August, after post-training it on Moonshot’s Kimi K3. People familiar with the work told Bloomberg that margins turned positive again after that launch and other changes to how the company used AI, and Harvey declined to comment on specific financials.
Abridge was building a clinical model on Nvidia’s open weights, Decagon said it was sending 80% of customer queries to models it owned, and Ramp said it was considering training a model for the first time after raising $750M in June. Rogo was also making a similar move, and Sequoia Capital and General Catalyst were backing the shift.
Anthropic showed an investor-forum slide about startups building their own models, while Mistral chief executive Arthur Mensch argued in July that closed models gave providers leverage. OpenAI later moved to end Cursor’s contract after SpaceX bought the company, citing past terms-of-service breaches by Musk businesses.
Legora, based in Stockholm, said it had reached $100M in revenue in 18 months and served more than 1,200 firms, but it did not say which models it used. Menlo Ventures partner Matt Kraning said model ownership was the wrong question and often amounted to cosplay.
Named in this story
People
- Arthur Mensch
- Mistral’s chief executive who argued that closed models give providers leverage
- Matt Kraning
- the Menlo Ventures partner who said model ownership was the wrong question
Companies
- Harvey
- the legal startup whose margins fell and later recovered after launching its own model
- Abridge
- the company building a clinical model on Nvidia’s open weights
- Decagon
- the company that said it routed 80% of customer queries through its own models
- Ramp
- the company that raised $750M in June and was weighing model training for the first time
- Rogo
- the company making a similar move towards its own models
- Sequoia Capital
- the investor funding the shift towards open-weight and in-house models
- General Catalyst
- the investor funding the shift towards open-weight and in-house models
- Moonshot
- the Chinese lab whose model Harvey used for post-training
- Nvidia
- the company whose open weights Abridge used
- OpenAI
- the company that moved to end Cursor’s contract
- Anthropic
- the company that showed a slide on startups building their own models
- Mistral
- the company whose chief executive argued that closed models give providers leverage
- Windsurf
- the company that Arthur Mensch said Anthropic had cut off
- Cursor
- the company whose contract OpenAI moved to end after SpaceX bought it
- SpaceX
- the company that bought Cursor
- Legora
- the Stockholm legal startup that said it had reached $100M in revenue
- Menlo Ventures
- the firm where Matt Kraning is a partner
Products and systems
- Kimi K3
- the Moonshot model Harvey post-trained on
How the source tells it
The source was mostly straight reporting, with a brisk, slightly stylised register and no sustained alarmism or boosterism.
No emotive framing was found in the original.