Treasury Threatens Sanctions Over Chinese AI Distillation as Nvidia's Huang Tells Washington to Back Off

Treasury Secretary Scott Bessent went on television Tuesday and put China's AI industry on notice. Nvidia CEO Jensen Huang spent the same news cycle arguing the opposite position.

Bessent said the US will scrutinize Chinese open-source AI models for stolen intellectual property, and could hit the companies behind them with sanctions, according to Bloomberg. His evidence: watermarks. "We are finding watermarks of our US large language models on many Chinese models," Bessent said. "This administration supports open source models, but what we do not support is IP theft." He said action could come within weeks.

The technical term for what Bessent is describing is distillation, training one model on the outputs of another. Whether that counts as theft or standard industry practice is genuinely contested, and it matters because Bessent is threatening sanctions over it.

What Triggered This

The flashpoint is Kimi K3, a 2.8-trillion-parameter open-weight model released by Chinese startup Moonshot AI in mid-July 2026, according to Crypto Briefing. It has a 1-million-token context window, topped benchmarks on the Frontend Code Arena leaderboard, and scored well on GDPval-AA v2 tests. It performs comparably to GPT-5.5 and Claude Opus 4.8, per Tom's Hardware, while costing roughly a third as much to run.

Wall Street did not take it well. Semiconductor stocks fell more than 20% from their June 2026 peaks after the release, according to Crypto Briefing, and Nvidia briefly lost its title as the world's most valuable company.

Huang's Counter-Argument

Huang, speaking to Axios at a new chip plant in Texas and later at the China International Supply Chain Expo, called the Washington backlash overblown. "These Chinese models are excellent," he said, according to TNW. "Open-source models that are excellent should be used." Asked directly whether American companies should be allowed to run Chinese AI models, Huang told Axios co-founder Mike Allen: "Absolutely."

He also pushed back on the idea that downloaded Chinese models carry hidden vulnerabilities tied to Beijing. "There is a misconception that somehow there are backdoors that are somehow connected to China in some way," Huang said, per Tom's Hardware. "You download the models, you can fine-tune it, you can enhance it, you can guardrail it as you desire."

Huang's commercial logic is straightforward: cheaper, freely available models put AI in more hands, and more AI users means more demand for the chips, data centers, and power Nvidia sells. "Free AI should be great for chips," he said, according to TNW. He made the same point Nvidia made after DeepSeek's release rattled markets earlier: cheaper models don't shrink the pie, they grow it.

At the Chinese trade expo, Huang went further, name-checking DeepSeek, Alibaba, Tencent, MiniMax, and Baidu as evidence that China's open-source AI ecosystem is legitimate and globally competitive, according to Crypto Briefing. He argued open-weight models are actually safer than closed, proprietary systems because outside researchers can inspect them for flaws. "If everything just becomes one single model, one single point of attack, one single source of failure, I think the world is much, much more vulnerable," he said.

Huang is not arguing for a hardware free-for-all. He has endorsed limiting exports of Nvidia's most advanced chips, including Blackwell and Rubin architectures, to China, according to Crypto Briefing. His position is to restrict the silicon, not the software.

The Case Against Huang's Framing

Bessent's IP-theft concern is not fringe. If Chinese firms are training on US model outputs to shortcut years of expensive research and then undercutting American labs on price, that is a competitive and legal grievance that Anthropic and other US labs have been making similar accusations about for months, according to TNW.

But the evidence offered so far, watermarks showing up in Chinese model outputs, establishes that distillation happened. It does not by itself establish theft in a legal sense, and no charges, indictments, or formal sanctions have been announced. Bessent's own framing treats this as settled wrongdoing before any enforcement action exists.

Hugging Face CEO Clem Delangue disputed the scale of the problem entirely. "We know distillation to be a very small factor," he told TechCrunch, adding it is "a practice that everyone is doing, including companies in the US."

There's an added wrinkle: Anthropic, among the most vocal companies alleging Chinese IP theft, recently had its own $1.5 billion settlement approved over pirated books used to train Claude, according to TNW. Huang used that fact to argue Washington should stop restricting Anthropic's Mythos model too. "Let Anthropic run," he said.

What's Next

Bessent is set to lead the US side of AI talks with China scheduled for September, where the distillation dispute will reportedly be on the agenda, according to TNW. Whether Treasury follows through with actual sanctions within weeks, as Bessent promised, or whether the threat cools once formal talks begin, remains an open question. So does whether any US company actually tests Huang's "absolutely" by deploying a model like Kimi K3 while Washington is actively threatening to sanction its maker.