US Treasury Proposes Rule to Bar US Investment in Any Chinese AI System Development, Scrapping Compute Thresholds

On October 22, the US Treasury released a Notice of Proposed Rulemaking (NPRM) implementing the Comprehensive Outbound Investment National Security (COINS) Act, marking a fundamental shift in AI investment rules toward China. **From compute-tiered screening to blanket prohibition.** Previously, investments were judged by training compute (notification above 10²³ FLOPs, prohibition above 10²⁵ FLOPs) and specific end uses; the proposal would bar US investors, in principle, from investing in the "development" of any Chinese AI system. Treasury argues that over the past two years AI capability gains have depended less and less on raw training compute — algorithmic advances, synthetic data, post-training and inference-time compute matter just as much — so fixed compute thresholds no longer work. **A much wider net.** Covered sectors expand from three (semiconductors, quantum, AI) to six, adding high-performance computing/supercomputing, hypersonics and biotechnology; the targets go beyond foundation models to explicitly include AI agents, embodied AI, robots and unmanned systems — skipping base-model work is no longer a safe harbor. **"Development" defined very broadly.** It covers not just writing code and training models, but R&D through affiliates and participation in R&D decisions (setting technical roadmaps, approving R&D plans, directing and supervising). An offshore parent that never writes code could still be deemed a developer. Countries of concern grow from China (incl. Hong Kong and Macau) to add Russia, Iran, North Korea, Cuba and Venezuela. The rule is a proposal; public comments are due December 7. > Source: [Read the original article](https://news.qq.com/rain/a/20261023A031YU00)