US threatens sanctions on Chinese AI models over IP theft

If your app routes any traffic to a Chinese open-weight model because the tokens are cheap, the risk profile just changed. On Tuesday, Treasury Secretary Scott Bessent said the US is examining Chinese open source models for evidence of intellectual property theft and has the authority to sanction the companies behind them. Sanctions on Chinese AI models aren't an availability problem you solve with a retry — they're a compliance problem you solve with a lawyer.

What actually happened

Per CNBC, Bessent said the administration supports open source models but not IP theft: if overseas models are stealing from US companies, "we have the ability to sanction them." He also claimed the government is finding watermarks of US large language models inside many Chinese models, and said the administration would look at the issue "in the coming days or weeks."

The trigger is competitive, not just legal. Moonshot AI's Kimi K3, released earlier this month, beats leading US models on some benchmarks — and TechCrunch notes OpenAI and Anthropic are the American firms most directly named as being at risk. Reporting that the administration weighed a broader ban on Chinese open source models has been disputed, so treat that as unconfirmed. What's on the record is a Treasury Secretary saying the sanctions tool is available and being considered.

Why AI model sanctions matter for your business

An outage you can wait out. A sanctioned vendor you cannot. If a model provider lands on a restricted list, "we'll switch next sprint" stops being a roadmap item and becomes a same-day obligation — and every cached weight, fine-tune, and API contract you hold is suddenly a thing your accountant asks about.

The defense is boring and it works: don't hardcode a provider. Put every model call behind one internal interface with the provider, model name, and prompt format as configuration, not code. Keep a tested fallback on a US or EU host — not a theoretical one, an actual route you have run under load with your own evals. Know which of your workloads require the cheap model to pencil out, because those are the ones that break. And if you self-host open weights, note that possession isn't the same as permission; a sanctions designation reaches your deployment, not just your download.

We build the routing layer this way by default. Not because we predicted this specific headline, but because "your model vendor becomes politically radioactive" has been on the risk list all year.

Sources: CNBC, TechCrunch.