China AI Investment Proposal Would Bar More Venture Deals in App Developers
Treasury's COINS Act proposal would bar more AI app investments by removing compute and end-use tests, including for developers using third-party models.
Primary lensTrade policy
Sub-topicPolicy monitoring
Evidence base5 records used
Use casePolicy monitoring
Treasury's October 22 proposal to implement the Comprehensive Outbound Investment National Security Act, or COINS Act, would bar more U.S. venture investments in Chinese AI application developers. Current rules already address development using third-party models. The proposed change would remove the compute and end-use tests that divide AI transactions among prohibited, notifiable and uncovered categories, placing all covered AI development in the prohibited category. The current rule's notes to sections 850.217 and 850.224 and Treasury's proposed section 851.226(h), pages 103-104 show the distinction.
That would make the target's application architecture more consequential for U.S. corporate venture counsel. The proposal expressly includes certain agentic applications built around third-party models, while preserving a boundary for internal use of off-the-shelf systems. An investment memo needs to explain what the target adds to the model and how that capability is used. A statement that the company trains no foundation model would leave that question unanswered.
The Compute Threshold Would Disappear
Current rules distinguish prohibited, notifiable and uncovered AI transactions through training-compute thresholds and specified end uses. Among the prohibited categories are systems trained using more than 10^25 computational operations, or more than 10^24 when primarily using biological sequence data. Those thresholds sit alongside military, intelligence and surveillance provisions in the 2024 final rule, section 850.224.
Treasury now argues that training compute cannot adequately separate advanced models. Synthetic data, changes in model architecture, post-training techniques and additional inference compute can improve capability without a corresponding increase in the original training run. General-purpose models may also acquire sensitive capabilities that developers have neither evaluated nor advertised. These are the agency's reasons for proposing a prohibition encompassing all AI development, explained on pages 30-31 of the proposed rule.
Treasury presents the removal of technical thresholds as a way to simplify compliance. It would also change the evidence counsel needs. An investment committee would still need to establish whether the target develops a system, even if it trains no model itself. The diligence question would increasingly concern the company's engineering work and product design.
An Application Can Be the Developed System
The proposal's definition extends beyond the underlying model to applications, tools and other systems that operate wholly or partly using it. Treasury uses the term "super intelligence" in this text, following an executive-order terminology change. The definition is not confined to systems with superhuman capabilities. Proposed section 851.232, page 109, and the terminology note on page 39 establish the relevant scope.
Section 851.226(h)(1) then explains when work on a third-party model counts as development. The target must enhance or extend the third-party system's capability. Its explicit example is an agentic harness that enables autonomous action loops, manages memory and context, or uses intermediate reasoning and self-correction to perform complex tasks. Owning or training the underlying model is unnecessary to that example.
Consider a prospective investment in a Chinese vendor selling an agent that completes customer workflows using a licensed model. The investment memo would need to explain the vendor's execution layer, memory functions and control over the model's actions. An answer limited to the model provider's identity would not resolve the development question. This follows from the proposed application and development provisions, sections 851.226(h) and 851.232.
Internal Use Has Its Own Boundary
The same proposed paragraph excludes use of an off-the-shelf third-party system strictly for internal, non-commercial purposes from development solely on that basis. It includes customization, configuration and fine-tuning for that use, giving sale or licensing as examples of commercial exploitation. Section 851.226(h)(2), page 104, supplies the qualification.
The following examples apply that language to three possible technical descriptions. They illustrate the development inquiry, not a complete determination of whether an investment is prohibited.
Target's activity
Treatment under the proposed development provision
Evidence counsel would need
Configures an off-the-shelf third-party assistant strictly for internal, non-commercial use
Does not constitute development solely on that basis
Users, purpose and limits on deployment
Builds an agent for sale or licensing around a third-party model, adding autonomous execution, memory or reasoning capabilities
Falls within the proposal's express example of development
Architecture, added functions and the target's engineering role
Uses a third-party tool internally while separately developing an AI product
The internal-use exclusion does not resolve the separate development activity
A description of each activity and the product roadmap
The comparison derives from proposed section 851.226(h). An internal deployment label alone would not settle whether work on the system falls within the limited exclusion.
The internal-use language is limited. It does not clear the company's other AI development or its activities in another covered technology. Nor does building an original system for internal use automatically bring it within an exception specifically framed around off-the-shelf third-party systems.
Counsel should ask the technical team to connect its architecture description to the product roadmap and licensing arrangements. That recommendation follows from Treasury's proposed diligence standard, which considers representations, non-public information, public claims and warning signs together. A contractual assurance is one part of that inquiry. Proposed section 851.104, pages 87-89, does not make it a substitute for the other evidence.
The Investment Still Needs Its Own Classification
A finding that the target develops AI would not prohibit every financial exposure to it. The proposal still requires a covered transaction involving the relevant covered foreign person, with applicable knowledge and exceptions. Equity, certain debt with equity-like rights, joint ventures and specified fund investments are among the categories in proposed section 851.209, pages 94-97.
Publicly traded securities and qualifying limited-partner investments would remain excepted. For the relevant LP category, the proposal retains a $2 million committed-capital threshold, aggregated across the fund's investment and co-investment vehicles, or a qualifying binding contractual assurance. Those exceptions can be lost when the investment affords governance rights beyond standard minority shareholder protections. They cannot be applied to an ordinary direct venture investment merely because its size is small. Section 851.501(a), pages 118-119, sets these limits.
The proposed $50,000 exception offers no general relief for prohibited AI investments. It applies only to transactions that would otherwise require notification, with funds and other consideration to the relevant covered foreign person or joint venture aggregated over the preceding 12 months. Non-cash consideration counts at fair market value. Its restricted scope is explicit in section 851.501(g), page 120.
This investment analysis also addresses a different obligation from China's interim AI export-registration requirement. A transaction involving both investment and technology transfer needs an assessment of each applicable regime.
December Comments Should Test the Development Boundary
For a live investment review, counsel can preserve the current-law conclusion while adding a separate assessment under the proposed development test. The useful comment would explain a specific architecture, the capability added by the investee and its intended use. Treasury expressly requests feedback on defining all AI development as prohibited technology. A final rule that narrows that definition or changes the treatment of third-party systems would change the investment analysis. Counsel should revisit the assessment when Treasury settles those provisions or the target changes its development plans.
Daily Trade Brief
Get the next one in your inbox.
New U.S. trade actions and the analysis that changes an import, customs, or counsel file, on weekday mornings after you confirm your address.
From reading to review
Run the numbers on your lane.
The duty calculator runs the current stack for any HTS code and origin. A free account opens full tool output, AD/CVD detail, Chapter 98 processing, and available exports.