Treasury Secretary Bessent: US considering sanctions on China over AI model theft

Mitchell Duran Tue, July 21, 2026 at 12:43 PM PDT | 3 min read

• NVDA +2.30%
• BABA -1.20%
• GOOG -1.24%

The Gist

Treasury Secretary Scott Bessent said the U.S. is considering sanctions on China over its alleged "theft" of U.S. large language models.

On Bessent's list are Chinese open-weight AI models such as Moonshot's Kimi, DeepSeek, and GLM of z.ai, as well as the tech giant Alibaba. The specifics of these proposed sanctions remain unknown.

If imposed, the sanctions could restrict or ban Chinese open-weight models and push U.S. developers toward domestic budget-friendly intelligence alternatives.

What Happened

Bessent on Tuesday told Fox Business that frontier AI companies like Anthropic and OpenAI face what he called "IP theft" from Chinese AI models that "distill" their products.

"If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft," Bessent said.

Model distillation is the process by which one AI lab copies or mimics another lab's higher-level model. This is done by training its own model with billions of questions while observing the better model's reasoning.

In other words, smaller AI labs like Moonshot and z.ai are said to be extracting knowledge from larger frontier AI labs like Anthropic and OpenAI, which performed the original heavy lifting.

Sanctions could be imposed in the coming weeks. Potential measures include:

• A Specially Designated Nationals (SDN) financial sanction on designated Chinese AI companies, freezing any U.S.-based assets, bank accounts, or IP they hold.
• Orders for U.S. companies and cloud providers to audit and remove Chinese models found to contain stolen weights or watermarks.

Bessent framed this as an anti-piracy measure: "You can't use counterfeit goods."

Why It Matters

Bessent told Fox Business he believes the sanctions would create stronger domestic competitors. Models including Meta's Muse Spark 1.1, Google's Gemma, Gemini, or even xAI's Grok could become the default options for U.S. companies if they are forced to suddenly remove Chinese AI open-weight models.

In late June, OpenRouter found that nearly 50% of all weekly API tokens processed for U.S.-based users on their platform came from Chinese open-weight models. Abruptly removing these models could impact companies' bottom lines—unless Bessent's "competitors" can fill the gap.

Google's Gemini 3.6 Flash closed model could be one such alternative. Shipped on Tuesday, Gemini 3.6 Flash ranks:

• #7 in Artificial Analysis' Intelligence Index (Moonshot's Kimi K3 ranks #3)
• #1 in speed

On cost, GLM 5.2 is priced at $0.47, $0.03 cheaper than Gemini 3.6 Flash at $0.50. If Bessent labels Chinese open-weight models as "counterfeit goods" and effectively bans U.S. corporations from using them, Gemini 3.6 Flash is positioned as a ready replacement. However, Meta's Muse Spark 1.1 and xAI's Grok 4.5 both score higher on Artificial Analysis' Intelligence Index than Gemini 3.6 Flash and are similarly priced, so they remain strong contenders.

What's Next

If Bessent—presumably with the president's blessing—decides to impose sanctions on Chinese open-weight models before the late-September AI Summit between China and the U.S., it would create an awkward diplomatic situation. This is especially notable because Bessent is leading the talks on the U.S. side.

For now, Bessent and Trump need to appear tough on television. U.S. companies will continue using their current tools, though some may begin shifting to similar domestic alternatives to lock in their operations.