In August, White House officials met with leading AI companies to discuss voluntary safety testing limited to proprietary frontier models, leaving open models exempt. MIT research shows closed models capture 96 percent of industry revenue at six times the cost of open alternatives, even as open models close much of the performance gap.
In August, White House officials held a closed door meeting with leaders of the country's top AI companies to discuss voluntary safety testing for advanced models. The proposed scope would cover only proprietary systems judged a potential national security risk, excluding open source and open weight models entirely. Separate research by MIT scientists Frank Nagle and Daniel Yue finds that closed models account for roughly 80 percent of usage and 96 percent of industry revenue, despite costing on average six times more than comparable open models, which frequently close most of the performance gap within a few months of a new release.
One explanation comes from Decagon co-founder Jesse Zhang, who describes a lifecycle pattern: during early experimentation, companies pay a premium for the flexibility of large proprietary models because requirements are still unclear. Once an application matures and runs at scale, the calculation flips toward speed, cost and control, favoring smaller, often open, fine-tuned models. Yet the share of enterprise AI spend going to open models recently fell from 19 to 11 percent, running counter to the popular narrative that open source is winning.
A regulatory asymmetry that subjects only proprietary models to formal safety checks could further entrench their market position, while open alternatives scale into production without equivalent scrutiny. It remains unclear whether such rules will actually be adopted, and how companies will adjust their build or buy decisions if compliance burdens and cost advantages end up distributed unevenly.
What this means for decision-makers
- Review your model portfolio for cases where open models could replace costly proprietary systems in production.
- Track upcoming safety rules that may treat closed and open providers unequally.
- Plan a two-stage approach, frontier models for discovery, open models for scaled production.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made