The open source AI alliance just discovered it has a geopolitical fault line running through it.
The Summary
- Treasury Secretary Scott Bessent threatened sanctions against Chinese AI companies over alleged IP theft, specifically calling out "distillation" as a technical term for what he bluntly calls "theft"
- The probe targets popular Chinese models including Moonshot AI's Kimi K3, Minimax, and DeepSeek, which US companies like Anthropic have accused of copying their research
- The administration claims to support open source in principle, but will use sanctions against models "built on IP theft"
- The threat comes amid concerns that lower-cost Chinese competitors may displace top American models, signaling the economic stakes behind the IP claims
The Signal
The Treasury Department just weaponized intellectual property law to draw a red line around Chinese open source AI models. This is not about stolen code repositories or pirated software. It is about distillation, the practice of training a smaller, cheaper model to mimic the outputs of a larger, more expensive one. When DeepSeek releases a model that performs comparably to GPT-4 at a fraction of the training cost, American AI labs cry foul. When Kimi K3 from Moonshot AI delivers similar capabilities without the billion-dollar compute bill, Anthropic accuses them of ripping off its research.
Distillation is not technically theft. It is a standard machine learning technique. You query a big model millions of times, collect its responses, then train a smaller model on that dataset. No weights are copied. No source code is stolen. The student model learns from the teacher's behavior, not its internals. This is how OpenAI built GPT-3.5-turbo from GPT-4. This is how every AI lab optimizes for inference costs. But when Chinese companies do it at scale and release the results as open source, Bessent calls it theft on Fox Business and threatens sanctions.
"This administration supports open source models, but what we do not support is IP theft."
The real issue is cost, not theft. Chinese labs are distilling American frontier models and releasing them openly at price points that undercut the entire Western AI business model. If a $50,000 distillation run can replicate 90% of the capability of a $100 million training run, the economics of closed AI collapse. The Treasury Department explicitly framed this as concern about "lower-cost Chinese competitors" sweeping aside top American models. This is not about protecting innovation. It is about protecting margins.
The Trump administration claims it still supports open source, but only the kind that does not threaten American commercial dominance. The logic is circular: open source is good when Meta does it to pressure OpenAI, bad when China does it to pressure everyone. Sanctions give the Treasury a veto over which open source models Americans can use, which models can be deployed on US cloud infrastructure, which weights can be legally downloaded. The open source AI commons just got a border wall.
The Implication
If you are building on open source Chinese models, start planning your migration now. Sanctions mean those weights could become legally radioactive for US companies overnight. If you are a Western AI lab, you just got handed a playbook for turning competitive pressure into a national security issue. Call distillation "theft," get your trade association to call it theft louder, wait for Treasury to act.
Watch what happens to the open source AI movement when its flagship models get carved up by trade policy. The whole pitch was that openness would decentralize AI power away from a handful of labs. Instead, we are building a world where your model's country of origin determines whether it is innovation or espionage. The commons are closing.