A Treasury Department proposal scheduled for Federal Register publication on October 26 would change the way some U.S. outbound investments in AI are sorted. Within the rule's defined transaction, entity and country scope, development of any system Treasury calls a “Super Intelligence” system would become prohibited technology. The current framework divides AI activity among prohibited, notifiable and uncovered categories using factors such as training compute and intended use.
That shift would remove the comfort of a single technical cutoff. An investor could not settle the question by checking a model's training budget or benchmark score. The file would have to show who the counterparty is, what it develops, where it is connected, what kind of transaction is planned and whether a specific exception or exemption applies.
What changed
Treasury's current outbound-investment rules, known as Part 850, took effect in January 2025. They cover certain transactions involving semiconductor, quantum and AI activity connected to China, Hong Kong and Macau. Treasury says it has received about 100 notifications involving semiconductor and AI investments during the program's first 21 months.
Congress created a broader statutory framework in 2025. Treasury's proposed Part 851 would implement it by adding high-performance computing, supercomputing and hypersonic systems, while expanding the list of countries of concern to include Cuba, Iran, North Korea, Russia and Venezuela under Nicolás Maduro's regime.
The proposal's sharpest AI change is categorical. Treasury would place development of any “SI system” in the prohibited-technology category rather than the notifiable category. “SI” is Treasury's terminology in this proposal. Its use here does not endorse the label or imply that every covered system has superhuman capability.
The proposal is not a blanket ban on AI investment or on every transaction involving the listed countries. It applies through defined terms for U.S. persons, covered foreign persons, covered national security transactions and specified technology activity. It also contains exceptions for some public securities, funds, limited-partner interests, derivatives, routine business activity and financial services, each with its own conditions.
Why it matters
Treasury says model capability no longer fits one dependable threshold. Training compute remains relevant, but the proposal points to other ways performance can improve, including synthetic data, more efficient algorithms and architectures, post-training methods and additional inference-time compute. It also argues that general-purpose systems may display sensitive capabilities even when those uses were not part of the original design.
That reasoning moves the operating question away from “Is this model above the line?” The harder question becomes “What, exactly, is this company developing, and how does this transaction connect to it?”
For an investment team, that means the technical review and the ownership review have to meet. A clean diligence record would need to connect the legal entity to its country relationships, products, research activity, controlled affiliates and transaction structure. It would also need to preserve the evidence used to decide that the transaction was prohibited, notifiable, excepted or outside the rule.
The same burden can reach beyond a direct investor. The proposal addresses transactions by controlled foreign entities and knowledge learned after a transaction. It would also let Treasury consider whether a U.S. person appropriately reviewed information the department had already provided when judging whether the inquiry was reasonable and diligent.
Verification bottleneck
A categorical rule can look simpler because it removes a numeric dividing line. The evidence problem becomes more complicated. Company descriptions, product claims, ownership records and research activity may not use the regulator's categories. A general-purpose model may sit inside a larger tool, application, robotic system or agent harness. The proposal's definition reaches those arrangements when they operate in whole or in part using the covered machine-based system.
Treasury proposes a confidential non-binding feedback process for potentially prohibited transactions, along with a path for anonymized public guidance. Those routes may help parties surface uncertainty before acting, but non-binding feedback is still not a legal safe harbor. A proposed public database of covered foreign persons would also be discretionary and non-exhaustive, so its absence of a name would not settle the issue.
The transition creates another recordkeeping problem. The proposed Part 851 would not be retroactive. Existing Part 850 rules would continue to govern activity from January 2, 2025 until the effective date of a later final rule. A useful transaction file will therefore need the applicable rule, dates and evidence preserved together.
Watch next
Comments are due December 7, 2026. The most consequential questions concern the breadth of “any SI system,” the treatment of systems built with third-party models, the limits of exceptions, and the evidence Treasury expects from a reasonable and diligent inquiry.
The final rule could narrow, revise or retain the proposal. Until then, this document is a map of Treasury's intended direction, not a final prohibition under Part 851 and not a finding about any company or transaction.
Source
- Treasury Office of Investment Security, “Outbound Investment Security Program Regulations,” Federal Register document 2026-21786, proposed rule scheduled for publication October 26, 2026; comments due December 7, 2026.
