Washington just declared open season on open source—if it's trained in China.

The Summary

The Signal

The US government is putting Chinese AI models under a microscope, with Treasury Secretary Scott Bessent stating the administration will examine open source models for signs they were trained on stolen American intellectual property. The target isn't just any AI. It's the open source models that China has been releasing at a pace and price point that makes Silicon Valley nervous.

The timing tells you everything. Chinese AI companies are shipping lower-cost models that compete directly with OpenAI, Anthropic, and Google—often trained on similar data, sometimes built with similar architectures, always cheaper to run. DeepSeek. Alibaba's Qwen. Baidu's ERNIE. These aren't knockoffs anymore. They're competitive.

"Open source models from overseas are now a Treasury Department problem."

The accusation underneath Bessent's statement is clear: if you can build a frontier model for a fraction of the cost and release it open source, maybe you didn't start from scratch. Maybe you had access to training data, model weights, or architectural insights that weren't yours. The US has spent billions building these models. China is releasing comparable ones for millions.

Here's what makes this complicated:

  • Open source AI models are distributed, forkable, and often anonymous in origin
  • Proving IP theft in machine learning is harder than proving it in software—models don't copy-paste code, they approximate patterns
  • The US benefits enormously from open source AI, including models like Meta's Llama, which was trained partly on scraped web data that may not have been properly licensed

Bessent's scrutiny policy creates a new filter layer in the AI supply chain. If a Chinese lab releases an open source model, US companies and developers will now have to assess legal risk before integrating it. That's friction. That's also protectionism dressed up as enforcement. The line between defending IP and stifling competition isn't always clean.

The Implication

If you're building on open source models, the legal surface area just expanded. Models from non-US sources, especially China, now carry regulatory risk. That doesn't mean don't use them. It means document why you chose them, verify their lineage where possible, and understand that "open source" no longer implies "safe to integrate."

For the broader agent economy, this is a preview. As AI models become infrastructure, expect more governments to treat them like controlled goods. The era of borderless model sharing is ending. What replaces it will be messier, slower, and more political.

Sources

Bloomberg Tech