Treasury proposes banning US investment in China AI

If you put US money into AI, the map just got a lot smaller.

The Treasury Department proposed a rule on Oct. 22 that would bar US persons from investing in the development of any AI system by companies tied to China and other countries of concern. Not big models. Not frontier labs. Any.

The old rules drew the line at training compute. This draft erases the line.

The threshold is gone

The proposal implements the COINS Act, which Congress passed in the 2026 defense bill. Under the current outbound investment rules, AI deals were sorted by how much computing power trained a model and what it was meant to do. Some were banned, some only had to be reported, some were not covered.

The proposed rule moves AI out of the "notify us" bucket entirely. "Development of any SI system" becomes a prohibited technology. SI stands for super intelligence: a September executive order told agencies to use that term in place of AI in certain documents.

Treasury's reasoning is blunt. It says models now gain ability through synthetic data, better algorithms, new post-training methods and more compute at answer time, so "no single measure, benchmark, or threshold is currently sufficient to distinguish advanced models."

The countries of concern are China, including Hong Kong and Macau, plus Cuba, Iran, North Korea and Russia, per the draft. It also lists Venezuela under the Maduro regime, then notes that regime no longer exists: the US recognized Delcy Rodríguez as Venezuela's head of state on March 5.

Building a wrapper counts

Here's the thing most founders will miss.

The draft defines an AI system to include any software or hardware that runs on one, and names "a harness for agentic capability" and robots by example. It says a company that builds an agent harness on someone else's model is developing an AI system, because the harness extends what the model can do: running action loops, managing memory and context, self-correcting.

So a startup tied to a covered country that never trained a model, and only ships an agent on a third-party API, would be off-limits to US investors under this text.

There is one carve-out. Using an off-the-shelf model strictly for internal, non-commercial purposes, including fine-tuning it for internal use, does not count as development on that basis alone.

The $50,000 small-deal exception does not help either. The draft limits it to deals that would only need a notice, so a banned AI deal stays banned at any size.

What it means if you invest or partner

If you are a US angel, fund or company with stakes in, or a joint venture with, an AI startup that has ties to these countries, read the draft now. The old test asked how big the model was. The new one asks whether the company builds with AI at all.

Treasury says it "particularly welcomes" comment on defining the ban this broadly, and asks for alternatives that hit the same risk with less damage to US businesses. It also points to a process for asking to do a deal judged to be in the US national interest.

"Under President Trump's leadership, we are committed to putting America's national security first," Treasury Secretary Scott Bessent said in the release.

This follows the federal government's order to strip DeepSeek from civilian agencies earlier this week. Both moves point the same way: wider bans, fewer exceptions.

What to watch

Comments are due:

December 7, 2026.

Nothing takes effect until Treasury publishes a final rule, which will set the date. Watch who files comments, and whether the agent harness language survives them. If it does, "we just build on top of the models" stops being a defense.

Read the draft before your next term sheet.