Some startups, like Harvey, Abridge, Ramp, and Rogo, are embracing open-weight models or training their own models to reduce expensive reliance on frontier labs
Context & Ripple Effects
The move extends a cost-management shift already visible among companies using cheaper models to curb rising AI spending, including models from China, while earlier coverage described US startups adopting open-weight Chinese systems for their lower cost and customization. Startups’ earlier adoption of lower-cost open-weight models established an alternative to relying exclusively on frontier providers.
It also gives the open-weight ecosystem a more concrete customer case as investors question whether model builders can produce durable revenue. Investor scrutiny of open-weight model makers makes application companies’ willingness to train or adapt models commercially significant.
First-order effects
- Harvey, Abridge, Ramp, and Rogo can reduce their dependence on frontier-lab model access and pricing by using open-weight systems or internally trained models.
- Frontier labs face less locked-in demand from application startups whose workloads can move to cheaper or customized alternatives.
Second-order effects
- Model pricing pressure intensifies as software vendors can compare frontier APIs with open-weight deployments rather than treating a single provider as the default.
- Open-weight model suppliers gain a clearer route to commercialization through companies that need models adapted to their own products and cost structures.
Third-order effects
- If this pattern holds, the AI application layer will increasingly treat foundation models as substitutable infrastructure, reserving proprietary advantage for product data, integration, and model adaptation.
- The competitive divide may shift from access to the largest general-purpose model toward the ability to operate a capable model stack at sustainable unit economics.
The trend: AI software companies are moving toward multi-model and self-managed deployments as inference costs turn model sourcing into a core margin decision.