The Open-Source AI Race: Why China's Strategy Is Outpacing America's Locked-Down Approach

The AI model market is a commodity business. Switching costs between models like ChatGPT and Claude are negligible—swap an API key and you're done. The real moat isn't the model itself; it's the enterprise services, contracts, and integrations built around it. That's a thin wall, and Chinese open-weight model providers are exploiting it.

US export controls on GPUs and data-sharing restrictions have created a curious dynamic: Chinese companies can train competitive models but can't offer global-scale centralized services. So they've turned a compute disadvantage into a distribution advantage by releasing models openly. Open weights aren't true open source, but they're portable and permissionless—anyone can host, tweak, and experiment with them.

The result? According to a16z partner Martin Casado, there's an 80% chance any given startup is using Chinese models. Moonshot and Alibaba recently unveiled models that claim parity with OpenAI and Anthropic at a fraction of the cost. The US frontier model advantage is closing fast.

This isn't just about benchmarks. Open infrastructure wins because it sits at the center of more innovation. Every sector in China—manufacturing, scientific research—can plug in these models without vendor lock-in. Meanwhile, American companies chase first-order profits with proprietary APIs, and the government tries to compensate with export controls. That's a losing strategy for a technology with no real moat.

The irony is stark: China, a locked-down society, is releasing AI technology more openly than the US. The incentives in the US are misaligned—companies prioritize short-term revenue over ecosystem benefits. If AI spending, which currently drives a significant chunk of the US economy, collapses as commoditization accelerates, the fallout could be severe.

What's needed is a more nuanced strategy: support for public AI, federated services, and open research. The US built the open internet; it can do the same for AI. But that requires recognizing that locked-down, proprietary models are a dead end.