How Chinese Companies Are Distilling US AI Models to Dominate Open-Weight Distribution
Key Takeaways
- ▸Chinese companies are systematically using model distillation to extract capabilities from American frontier AI systems and create competitive alternatives for local deployment
- ▸The AI competitive advantage has shifted from model development to distribution—determining which country's models become the default on billions of devices globally
- ▸US firms face legal and contractual constraints preventing them from employing the same distillation tactics that Chinese competitors can freely deploy
Summary
A new Foreign Affairs analysis examines how Chinese AI companies are systematically distilling frontier models from American AI providers to create competitive, lightweight versions for consumer devices. Model distillation—training smaller, more efficient AI systems to mimic sophisticated ones—enables these models to run on laptops and mobile devices without cloud infrastructure. While this practice promises to democratize AI access, it creates a significant competitive asymmetry: Chinese companies operate with few restrictions, while US firms are contractually prohibited by their terms of service from using model outputs to train competitors. The result has shifted the AI battleground from model development (where the US leads) to distribution of open-weight models on billions of consumer devices worldwide. The article argues that without targeted policy intervention, the United States could "win the AI training battle and lose the distribution war," and proposes adapting existing US-China policy strategies to promote American open-weight leadership while penalizing anticompetitive practices by Chinese entities.
- Without policy intervention, Chinese open-weight models may dominate global markets, creating a distribution-layer advantage even as American firms maintain frontier capability leadership
Editorial Opinion
The article identifies a legitimate competitive asymmetry worthy of policy attention: Chinese firms distilling US models while American firms cannot reciprocate due to contractual constraints. However, the prescribed government response carries significant execution risk. If applied too broadly, regulation could damage US AI leadership by restricting the international talent and model access that American companies depend on. The challenge lies in surgical policy calibration that targets clear anticompetitive practices without erecting barriers that ultimately impede American innovation.
