The Treasury Wants to Sanction Code You've Already Downloaded
The Trump administration wants you to believe it can lock down open-source software with executive orders. Good luck with that.
Treasury Secretary Scott Bessent dropped a bombshell this week, floating sanctions against Chinese open AI models over alleged intellectual property theft. The argument is simple enough on the surface. Washington claims Chinese labs use output distillation from top US models to steal proprietary trade secrets, releasing open weights that rival American frontier models for a tiny fraction of the cost.
Here's what most coverage misses: trying to sanction open weights is like trying to place trade sanctions on a torrent file. Once model weights are uploaded to Hugging Face, GitHub, or a thousand mirror servers across Asia and Europe, the math has left the building.
The reality is that Washington isn't just worried about stolen code. They're terrified that cheap, capable open-weight AI is destroying Silicon Valley's profit margins. We've seen how open-weight models panic Washington, but turning economic sanctions into a digital firewall for model weights is a fundamentally broken strategy.
Distillation, IP Theft, and the Open Source Trap
Let's look at how we got here. American companies like OpenAI spent billions running giant GPU clusters to build foundational systems. Then Chinese startups like DeepSeek and established tech heavyweights like Alibaba came along, using output distillation to train models at roughly 5% of the compute cost.
Is it distillation? Absolutely. Everyone in the AI industry knows Chinese teams regularly prompt US APIs to generate synthetic dataset training pipelines. But calling that straightforward IP theft under existing US law is legally murky at best. US courts are still wrestling with whether training on public web data violates copyright, let alone whether training on API outputs constitutes corporate espionage.
Take Alibaba's Qwen model as an example. It's widely used by developers across the globe because it performs exceptionally well on coding and reasoning tasks. If the Treasury Department bans US developers from downloading or hosting Qwen or DeepSeek weights, who actually gets hurt? It won't be the developers in Hangzhou or Beijing. It'll be American startups trying to run local models without paying high API bills.
An Enforceability Nightmare for American Developers
So how would the Treasury Department actually enforce sanctions against an AI model? They can easily target companies. They can sanction foreign corporate entities, block official API endpoints, and fine US cloud providers like AWS or Microsoft Azure for hosting those services.
But open weights don't need a cloud provider to run. They run locally on MacBooks, desktop rigs, and private corporate servers.
Imagine a developer team evaluating their stack by benchmarking ChatGPT vs Claude and deciding that a local, open-weights Chinese model gives them better privacy at zero marginal cost. If that team downloads a sanctioned weight file from a mirror site in Sweden, is the US government going to audit local Docker containers? It's completely unworkable.
That said, the corporate lobbying here is obvious. Closed-source American tech leaders have every incentive to push for these restrictions. They want high regulatory moats. They want open-weight competition neutered under the banner of national security and intellectual property protection.
Yet banning code has almost never worked in tech history. Banning model weights won't stop China from building elite AI. It'll just make American software developers regulatory outlaws for downloading a file.
Frequently Asked Questions
Can the US Treasury actually sanction software code?
The Treasury Department can place entities on the OFAC Sanctions List, making it illegal for US citizens and companies to conduct business with them or host their services. However, enforcing these restrictions against open-source model weights stored on local machines is nearly impossible from a technical standpoint.
What is model distillation in AI?
Model distillation is a process where a smaller or newer AI model is trained using the outputs generated by a larger, more powerful model. US officials claim Chinese companies use distillation on American AI models to skip expensive research and development costs.
Will American developers be banned from using DeepSeek or Qwen?
If formal sanctions are enacted, US cloud providers would be required to remove hosted versions of these models, and companies subject to US jurisdiction could face legal penalties for using or distributing them commercially.