Flooding the AI Frontier (bturtel.substack.com)

Chinese companies — backed by an explicit government tilt toward “openness” in its July Global AI Governance Action Plan — are flooding the market with high-quality open-weight large language models: downloadable, runnable, and fine-tunable models released under permissive licenses. These Chinese open-weight models often match or beat expensive proprietary US models on benchmarks while costing a fraction to run, and can be hosted by US providers (DeepInfra, Cerebras) so users don’t need to send data to China. The result is rapid domestic adoption, faster iteration by academia and industry, and broad access for startups that would otherwise lack resources to train foundation models from scratch. Technically, open-weight LLMs change the economics and engineering of model choice: interoperability tools like OpenRouter make swapping models almost trivial (often a single parameter), reducing lock‑in and elevating fine‑tuning over end‑to‑end model development. That democratizes capabilities but compresses margins for frontier labs, undermining incentives to fund massive, risky training runs and potentially creating long‑term dependencies—an effect likened to China’s manufacturing playbook. The US response (open releases like gpt-oss, export controls, and industry warnings) acknowledges strategic risk, but whether open-weight proliferation will materially shift global AI leadership remains contested.