Treasury Secretary Scott Bessent dropped a bombshell this week when he announced the U.S. government is ready to sanction Chinese companies over AI model theft. Speaking on Fox Business, Bessent made it clear that Washington has seen enough. American investigators are detecting digital watermarks from top U.S. large language models embedded inside open-weight models originating in China.
It isn't a case of traditional industrial espionage where someone sneaks out with a thumb drive. It's a technical process called distillation. Chinese labs are systematically extracting intelligence from American AI models to build their own systems at a fraction of the cost. With Beijing heavily subsidizing electricity and computing infrastructure, those models are hitting the global market for dirt cheap, and Washington is preparing to fight back.
What AI Distillation Actually Means and How It Works
To understand why the White House is threatening sanctions, you need to look at how distillation works in modern machine learning. Building a frontier AI model from scratch requires hundreds of millions of dollars, tens of thousands of Nvidia GPUs, and months of continuous training.
Distillation bypasses almost all of that work.
A developer takes an existing top-tier model, like OpenAI's GPT or Anthropic's Claude, and bombards it with millions of automated prompts. The smaller model studies the complex responses, reasoning chains, and code outputs of the larger model. It learns to mimic the flagship system without having to discover the underlying logic on its own.
It's essentially copying off the smartest student in class after they spent all night studying.
Anthropic recently filed a complaint with federal lawmakers detailing how operators connected to Alibaba's Qwen AI lab ran over 28.8 million queries through 25,000 fraudulent accounts against Claude. OpenAI and Anthropic have flagged similar behavior from other Chinese AI developers including Moonshot AI, DeepSeek, and MiniMax.
Detecting this isn't easy because a single distillation prompt looks identical to a genuine human query. But when hundreds of coordinated accounts fire off tens of millions of structured requests around the clock, patterns emerge. More importantly, subtle watermarks embedded deep inside American models are now showing up inside Chinese open-source releases.
The Trigger Behind Washington's Sudden Action
This threat of sanctions didn't appear out of thin air. It was triggered by the release of Moonshot AI's Kimi K3 model and Alibaba's Qwen updates.
Kimi K3 shocked global tech markets when performance metrics showed it rivaling or beating flagship American models across key reasoning benchmarks. The real kicker? Running tasks on Kimi K3 costs roughly $0.95 per benchmark task, compared to well over $2.50 on leading closed American models.
That pricing gap sent shockwaves through Wall Street, causing a temporary sell-off in U.S. tech stocks as investors questioned whether Silicon Valley's massive infrastructure spending could withstand ultra-cheap open-weight competition from abroad.
American AI companies are pouring tens of billions into datacenters and power plants. If overseas competitors can simply harvest those outputs via API calls, repackage them into open-weight models, and distribute them worldwide for free or near-free, the financial model for frontier AI research crumbles.
What US Sanctions Could Look Like in Practice
Bessent indicated the administration is looking at several regulatory levers to stop the practice.
First, the Commerce Department could place offending Chinese AI labs on the Entity List, making it illegal for U.S. entities to supply them with software, hardware, or cloud computing resources.
Second, the government is considering mandatory disclosure rules. American businesses using foreign AI models might soon be required to report those integrations to federal regulators and clients. As Bessent put it during his broadcast, "You cannot use counterfeit goods."
There's also growing pressure to restrict U.S. cloud providers from serving anonymous overseas traffic that exhibits distillation patterns. By restricting API access and blocking proxy networks used to execute automated queries, regulators hope to cut off the data stream before it reaches foreign training servers.
Potential Surprises and Unintended Consequences
Sanctioning open-weight AI models carries serious technical and economic complications that policy makers are just beginning to grapple with.
Once an open-weight model like Kimi K3 is released online, the weights are out in the wild. Anyone can download them, host them locally, or run them on private infrastructure. Unlike physical trade or centralized SaaS platforms, enforcing sanctions against decentralized software weights is nearly impossible.
Imposing strict restrictions on foreign open-source AI could backfire on American developers. Thousands of U.S. startups rely on low-cost open-weight models to build niche applications, lower operating overhead, and avoid total reliance on a handful of tech giants. Banning or penalizing the use of foreign open-source models could force small American businesses into high-cost software monopolies.
Some policy advisors within the administration have warned that aggressive bans risk cementing a duopoly between the largest U.S. frontier labs while slowing down bottom-up innovation across the broader software industry.
Steps Companies Should Take Right Now
If your organization builds software or relies on third-party AI APIs, the sudden threat of federal sanctions requires immediate preparation.
- Audit your software stack to identify every foreign or open-weight AI model currently running in your pipeline.
- Track where your API queries are hosted and ensure your providers comply with U.S. export controls and IP rules.
- Establish clear data provenance policies so you can prove your internal models weren't trained on distilled or unauthorized outputs.
- Build flexible fallback options into your applications so you can swap out model backends quickly if specific foreign providers face sudden regulatory bans.
With high-level bilateral AI talks scheduled between Washington and Beijing for September, this issue is going to dominate tech policy for the rest of the year. Companies that act early to audit their AI infrastructure will avoid costly disruptions when the first wave of enforcement hits.
