China's Open-Weights AI Strategy Is Winning

American AI companies are betting on proprietary models and locked-down APIs. China is betting on open weights — and it is pulling ahead.

The AI race has taken a sharp turn. While American frontier labs like OpenAI and Anthropic continue building walled gardens around their models, Chinese companies are flooding the ecosystem with open-weight alternatives — and the strategy is working.

Ben Werdmuller, writing at Werd.io, argues that AI models as products in themselves have very little moat beyond brand loyalty and superficial switching costs. The real defensibility sits in enterprise services — cloud deals, system integrations, and quality-of-life features — not in the models themselves. Developers can swap API endpoints with a single line change; users migrate from ChatGPT to Claude with near-zero friction.

The U.S. government's GPU export controls have backfired in an unexpected way. By restricting Chinese access to compute at scale, Washington inadvertently pushed Chinese AI firms toward open-weights distribution rather than centralized cloud services. China has enough compute to train competitive models but lacks the hardware to offer global-scale hosted services in the style of OpenAI. Open-weight release became the natural alternative — and it worked.

Open almost always wins in infrastructure adoption. Open-weight models can be used permissionlessly, hosted anywhere, and fine-tuned for specific use cases. That is the exact formula that made open-source infrastructure dominate every prior technology cycle, from Linux to Kubernetes to the web itself.

Stratechery's Ben Thompson echoes the concern: while American frontier labs will survive, the real vulnerability is that the U.S. lacks a healthy ecosystem of open alternatives. The pipeline that once turned academic research into startups into giants — Bell Labs to Silicon Valley — is at risk of being short-circuited by policies that inadvertently ceded the open-weights playing field to China.

The dynamic has not gone unnoticed. Chinese models from labs like DeepSeek and Alibaba's Qwen now compete at or near frontier performance while being freely downloadable. Developers worldwide are building on them rather than waiting for API access from San Francisco.

Sources: Werd.io — American AI is locked down and proprietary. It's losing. Stratechery — Who's Afraid of Chinese Models?