China AI Export Controls: MOFCOM Announcement No. 48 Explained
On October 16, 2026, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) issued Announcement No. 48, introducing significant new export controls on artificial intelligence. The announcement establishes an immediate interim registration requirement for technology exports and proposes a revised catalogue of restricted technologies that includes frontier model weights, agent architectures, and AI-enabled cyber capabilities. While the registration requirement is currently in effect, the broader catalogue additions remain open for public comment until November 15, 2026.
Scope of Control and Technical Thresholds
The draft regulations target frontier AI capabilities using specific technical benchmarks. A model is considered restricted if it meets any of the following criteria: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or a top-10 global position in public rankings, call volume, or similar metrics.
Beyond foundation models, the announcement introduces independent controls for AI agents. Technologies involving multi-agent coordination, complex task decomposition, and large-scale tool integration are subject to restriction. High-performance agents specifically designed for critical infrastructure, defense, or core government use are prohibited from export entirely. Furthermore, AI-assisted vulnerability discovery and malicious code generation tools are subject to strict licensing requirements or prohibitions.
Deployment and Enterprise Exposure
MOFCOM distinguishes between service access and physical custody of technology. API, SaaS, and lightweight on-device usage generally fall outside the scope of these export controls. However, the local deployment of restricted models on infrastructure outside of China is explicitly captured, requiring an export license.
Multinational corporations face significant compliance risks regarding internal operations. The registration requirement is triggered not only by commercial sales but also by the licensing of technology for overseas use, changes in corporate control, and potentially intra-group code sharing. Organizations with China-based R&D teams must now evaluate whether their internal transfers of model checkpoints, agent frameworks, or security tools require registration with MOFCOM.
Strategic Implications and Industry Response
Industry observers note that this move aligns China’s regulatory framework with existing U.S. export controls, creating a dual-jurisdiction compliance environment for frontier AI. Commentators on Hacker News have highlighted the tension between these restrictions and the previous trend of Chinese labs releasing open-weight models. As one user noted, "Restricting any powerful Chinese model that ranks in the global top 10 is going to make them very difficult to advocate for, and very hard for global users to adopt."
While the official stance from MOFCOM encourages open-source collaboration through "automatic licensing" and exempts technologies already in the public domain, the dynamic nature of the "global leading" threshold creates uncertainty. Because a model's regulatory status can shift based on its performance rankings, organizations relying on Chinese-origin models must treat them as contingent operational dependencies.
Recommended Actions for Organizations
- Inventory and Classification: Identify all Chinese-origin AI assets, including models, agent frameworks, and security tools. Determine if they are consumed via API or local deployment.
- Review Internal Transfers: Audit code repositories and R&D workflows to ensure compliance with the new interim registration requirements for cross-border data or code sharing.
- Contractual Safeguards: Update vendor agreements to mandate that suppliers handle necessary registration and licensing, and include provisions for service continuity in the event of a license denial.
- Monitor Regulatory Updates: Prepare comments for the November 15, 2026, deadline and track the finalization of the catalogue, as the scope of restricted items may evolve following the consultation period.