Treasury proposes banning US investment in super intelligence work in China

What Treasury proposed

On 22 October 2026 the US Treasury Department announced a proposed rule to update its Outbound Investment Security Program, the system that limits how Americans can invest in companies working on sensitive technologies in China and a handful of other countries. The proposal implements the Comprehensive Outbound Investment National Security Act, known as the COINS Act, which Congress passed in the defence bill for fiscal 2026 and which became law on 18 December 2025 as a new Title VIII of the Defense Production Act.

The headline change for the AI industry is blunt. Under the existing rules, some AI investments were banned and others only had to be reported, depending on how much computing power a model was trained with and what it was for. Under the proposal, a US person would be prohibited from investing in a company in a country of concern that develops any super intelligence system at all, with no technical threshold below which the deal becomes allowed.

Treasury Secretary Scott Bessent said the administration is committed to putting America's national security first and that the proposal would stop adversaries from using US investment to develop critical technologies that threaten it, while giving investors clarity and a chance to comment. Christopher Pilkerton, assistant secretary for investment security, said the rules are designed to address risks that arise when foreign adversaries try to benefit from US investment in key technology sectors.

From AI systems to super intelligence systems

The first thing a reader of the proposed rule notices is the vocabulary. What the current regulations, known as Part 850, call AI systems would become super intelligence systems, or SI systems, in a new Part 851. Treasury says the change follows Executive Order 14434, Inaugurating the Era of Super Intelligence, signed on 29 September 2026, which directed agencies to use the new term.

The definition itself is wide. An SI system is a machine-based system that can, for a given set of objectives, generate content or make predictions, recommendations or decisions that influence real or virtual environments. It also covers any system, software or device that runs partly on such a model, including what the rule calls an SI agent harness, and embodied systems such as industrial, humanoid and four-legged robots, autonomous vehicles and unmanned systems.

In practice that sweeps in nearly every modern AI product, from chatbots to robot software, rather than only the largest frontier models.

Why Treasury dropped compute thresholds

Treasury explains why it is abandoning the training-compute approach the Biden-era rules relied on. Over the past two years, it writes, models have gained capability through synthetic training data, more efficient algorithms, post-training and extra computing at inference time, not just bigger training runs. As a result it concluded that no single measure, benchmark or threshold is currently sufficient to tell advanced models apart from others.

The department also says end uses are hard to pin down. A general model built for business tasks can be turned to military or surveillance uses, and Treasury points to open-source models as carrying fewer safeguards against that kind of repurposing. Rather than draw a line it believes it cannot defend, it proposes to cover all SI system development in countries of concern, and it invites comment specifically on that choice.

The rule says Treasury did not propose any technical parameters under which investment in SI systems or quantum computing could be permitted. Those two areas become prohibited categories without the notifiable tier that exists for other technologies.

What counts as developing an SI system

Because the ban turns on development, the rule spends time on what development means. It says building an agentic harness layer on top of a third-party model counts as developing an SI system, since it extends what that model can do. A Chinese start-up that wraps an agent framework around someone else's model would therefore fall inside the ban, even if it never trains a model of its own.

There is a carve-out for ordinary users. Using an off-the-shelf third-party SI system strictly for internal, non-commercial purposes would not count as development, and neither would customising, configuring or fine-tuning a third-party model for that internal use, as long as the result is not sold or licensed. A Chinese manufacturer that uses a chatbot in its own back office would not, on that reading alone, become a prohibited investment target.

The other technologies on the list

The COINS Act names five technology areas: semiconductors and microelectronics, artificial intelligence, quantum information technologies, high-performance computing and supercomputing, and hypersonic systems. Treasury's proposal adds biotechnology as a sixth. Reuters reported that pharmaceutical licensing deals would be excluded from the biotechnology rules.

For AI hardware, the proposal treats high-performance computing interconnects, the scale-up and scale-out networking that links accelerators into large clusters, as notifiable. That means investments involving them would have to be reported to Treasury rather than banned outright. Semiconductor rules keep a mix of prohibited and notifiable categories, depending on the technology node and the type of product.

The approach mirrors controls on the other side of the Pacific. China's commerce ministry has itself proposed export controls on 23 AI technologies, including large models and agents, so both governments are now writing rules around the same building blocks.

Who and what the rule covers

The countries of concern listed in the proposal are China, including Hong Kong and Macau, Cuba, Iran, North Korea, Russia, and Venezuela under the regime of Nicolas Maduro. Treasury notes that this Venezuelan regime no longer exists, since the United States recognised Delcy Rodriguez as the country's leader on 5 March 2026, and Reuters reported that Venezuela was dropped from the list in practice.

The rules apply to US persons, a term that covers American citizens and permanent residents, US companies and their foreign branches, and anyone in the United States. They reach equity investments, certain debt that converts to equity, joint ventures, greenfield projects and some limited-partner stakes in funds that invest in covered companies.

Exceptions, penalties and divestment

The proposal keeps exceptions for some public-market investments and sets a de minimis threshold of $50,000, measured across the preceding 12 months. It also includes exceptions for things such as employee equity awards and transactions that are secondary to a covered deal.

On enforcement, the COINS Act gives Treasury civil penalty powers up to the maximum allowed under the International Emergency Economic Powers Act. Because the statute does not include language on criminal penalties, Treasury proposes to omit the criminal-penalty language found in the current regulations, while noting that criminal and other penalties under other laws may still apply.

The Secretary would also be able to compel divestment of a prohibited transaction. A voluntary self-disclosure route remains, and Treasury says it will weigh cooperation, including divestment or voluntary disgorgement of profits, when deciding on penalties.

How this fits the wider super intelligence push

The rename is part of a broader shift in Washington's vocabulary since the September executive order. A State Department cable has told staff to write super intelligence instead of AI, and an IMF committee dropped the term AI after a US push. Treasury's rule is the first place where the new term carries direct legal consequences for investors.

It also shows how the label changes policy, not just wording. By defining super intelligence to include every machine-based system that generates content or makes predictions, the rule treats all AI development in China as if it were frontier work, which is what removes the space for compute thresholds.

What happens next

This is a notice of proposed rulemaking, not a final rule. Treasury says written comments must be received by 7 December 2026, and it has posted background on its outbound investment program page. The current Part 850 rules stay in force until a final rule takes effect, and the final rule will set its own effective date.

Investors, law firms and AI companies are likely to push on the definitions, especially the treatment of agent harnesses, the line between internal use and development, and whether any technical threshold could work after all. Treasury said it particularly wants views on the proposal to prohibit investment in any SI system.

Why it matters

For US venture funds that still hold or consider stakes in Chinese AI companies, the proposal would close the notify-and-proceed path for most AI deals. Under the existing rules, a fund could often invest in a Chinese model maker below the compute thresholds after filing a notice. Under the new definition, almost any company that builds or sells AI software, robots or agents would be off limits.

The rule also sets a template other agencies may follow. Its argument that compute thresholds no longer separate advanced models from ordinary ones contrasts with US export controls and some state AI laws, which still lean on compute and benchmark measures. If Treasury's reasoning survives the comment period, it may push other regulators to define AI risk by who is building it and where, rather than by how large the model is.