AI Startups Flee Frontier Labs for Open-Weight Models as Cost Crisis Bites
Summary
- • Harvey's gross margins collapsed from +50% to -50% in months on frontier AI costs
- • Harvey launched in-house model on Chinese open-weight Kimi K3; margins recovered
- • Abridge, Decagon, Ramp, and Rogo all reducing reliance on OpenAI and Anthropic
- • Trend signals structural financial decoupling from frontier AI lab providers
Details
Harvey gross margins: ~+50% in Jan 2026, crashed to -50% by June 2026
A March 2026 update to Harvey's AI agents caused frontier model token usage to spike twentyfold, collapsing gross margins from approximately +50% at year start to -50% by June — the direct financial shock that forced a strategic pivot away from frontier lab dependency.
Harvey launched in-house model post-trained on Kimi K3 (Moonshot AI)
Harvey released its first proprietary model in August 2026, built by post-training Moonshot AI's open-weight Kimi K3 (a Chinese model) for specialized legal work. Gross margins subsequently turned positive after the switch.
Abridge, Decagon, Ramp, and Rogo all reducing frontier lab reliance
Abridge is building a clinical AI model on Nvidia open weights; Decagon routes 80% of its queries through proprietary models; Ramp (raised $750M in June 2026) is evaluating training its own models; Rogo is pursuing a similar open-weight strategy.
Ramp raised $750M in June 2026; evaluating proprietary model training
Ramp, the fintech-AI company that closed a $750M round in June 2026, is now evaluating whether to train its own models as part of a broader cost reduction strategy — despite having significant capital to absorb frontier model costs.
Sequoia Capital and General Catalyst backing open-weight adoption
Major venture investors including Sequoia Capital and General Catalyst are backing startups pursuing open-weight model strategies, signaling institutional validation of the decoupling trend and investor appetite for margin improvement.
US startups adopting Chinese open-weight Kimi K3 for commercial products
Multiple US AI startups are now building commercial products on Kimi K3 from Moonshot AI (China), a development with potential regulatory and national security implications as the US focuses export controls on hardware rather than model weight distribution.
Harvey ($15.6B) originally built on OpenAI GPT-4; pivot is a major signal
Harvey built its $15.6B legal AI business by fine-tuning OpenAI's GPT-4 for specialized legal work. Its pivot to an in-house open-weight model marks a symbolic and commercial break from the frontier lab dependency model.
Details compiled via Grok live web research
Additional details sourced from Bloomberg (Sept 21, 2026), The Next Web, and Bloomberg Law via Grok live web research. Primary article text was unavailable from original source.
Sources: Bloomberg (Sept 21, 2026), The Next Web, Bloomberg Law. Details compiled via Grok live web research.
What This Means
Harvey's margin collapse from +50% to -50% in three months is the most concrete financial evidence yet that frontier model dependency creates existential business risk — and the AI startup ecosystem is responding at scale. The move to open-weight and self-trained models represents a structural decoupling from OpenAI and Anthropic that, if it accelerates, erodes the recurring revenue lock-in both companies have built and pressures their path to trillion-dollar IPO valuations. The adoption of Chinese open-weight models like Kimi K3 in commercial AI products also quietly introduces regulatory and national security questions that Washington has yet to address — the US tightens export controls on AI chips while leaving high-capability model weight distribution largely unregulated.
Sentiment
Pragmatic concern for frontier labs' margins with excitement over startups' shift to open-weight models
“Harvey’s AI costs got so high its gross margins plunged from 50% to -50% in 6 months. Its response: build its own model using open-weight AI. Startups from Abridge to Rogo are following suit to cut costs and reduce their reliance on OpenAI and Anthropic.”
“the top application layer companies have embraced building their own open source models OVER using OpenAI and Anthropic... Seems like Harvey is among the latest to remove the FMD — frontier model dependency.”
“This is the Deepseek moment again but even crazier. OpenAI and Anthropic IPOs might be derailed if this keeps going on. Also this proves how pointless Harvey’s second try at post-training is... They should just host the latest, best performing Chinese OSS models instead.”
“OpenAI has backed this company since 2022. It just shipped a model built on someone else's weights... Harvey released Tenet, its first purpose-built model for long-horizon legal work. It's post-trained on Moonshot AI's Kimi K3. Open weight. Not OpenAI.”
“JUST IN: Legal AI startup Harvey's gross margin fell to minus 50% paying OpenAI and Anthropic, then turned positive after it trained its own model on top of a free Chinese one. Source: Bloomberg”
Split
~65/35 split: startups & builders see cost relief & strategic independence vs. frontier labs facing revenue erosion and valuation pressure