Joint Announcement No. 48: China Codifies AI Export Restrictions
On October 16, 2026, China's Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology (MOST) published Joint Announcement No. 48 of 2026, initiating a public consultation on proposed revisions to the nation's Catalogue of Technologies Prohibited or Restricted from Export. The consultation period runs through November 15, 2026.
The proposed framework adds 40 strategic technologies across three sectors—classifying 14 as prohibited and 26 as restricted. Notably, 23 of these technologies are dedicated explicitly to artificial intelligence, spanning frontier foundation models, autonomous agents, and AI-enabled cybersecurity capabilities.
Technical Criteria: The 2-Trillion Parameter and 10^26 FLOPs Thresholds
Under the prohibited category, the draft explicitly bars the export of AI-driven malicious code generation and mutation technologies targeted at critical infrastructure, national defense, and government systems. The restricted export category includes AI-assisted automated vulnerability discovery, alongside frontier foundation models and advanced agent workflows.
For frontier foundation models, export restrictions attach directly to model weights, training checkpoints, structural architectures, and underlying source code if the system meets specific technical baselines: models trained using 10^26 FLOPs or greater, architectures with 2 trillion parameters or more, or systems scoring at 'globally leading' performance benchmarks.
Immediate Pre-Export Filings and Critical Open-Source Carveouts
Critically, while the broader regulatory catalogue remains subject to public commentary, a mandatory pre-export registration mechanism took effect immediately on October 16, 2026. This requirement covers in-flight technology transfers, mandating corporate disclosures detailing recipient entities, destination jurisdictions, stated end-uses, and estimated commercial valuations prior to weight transfers or offshore intellectual property licensing.
Conversely, the framework includes definitive safe harbors. Technologies already released into the public domain via academic papers or open-source repositories are explicitly exempt. Furthermore, serving model inference to international users through standard application programming interfaces (APIs), web interfaces, or software-as-a-service (SaaS) platforms is expressly excluded from classification as a technology export. New open-source contributions and Chinese enterprises operating proprietary overseas cloud infrastructure are routed into a streamlined, automated licensing pathway upon filing.
Practitioner Reactions, Regulatory Ambiguities, and Operational Skepticism
Among software practitioners and machine learning engineers, early discussions reflected acute concern that China was preparing to restrict its internationally relied-upon open-weight ecosystem. These anxieties moderated as developers analyzed the explicit open-source safe harbor provisions. Nevertheless, technical discourse has pivoted toward the compliance hurdles facing future frontier releases exceeding 2 trillion parameters originating from top Chinese research laboratories.
Concurrently, speculative suggestions that high inference traffic alone across global model routers might retroactively subject existing systems like DeepSeek or GLM to export bans remain ungrounded inferences. The regulatory text clearly insulates already-public model weights and external API calls. However, ambiguity remains regarding how regulatory bodies will systematically operationalize benchmarks defining 'globally leading performance' once final revisions take effect.
Strategic Implications for Enterprises and System Integrators in Thailand
For Thai enterprises, technology consultancies, and system integrators deploying AI solutions, China's policy announcement provides short-term operational continuity. Because open-source distributions and commercial API access are formally exempted, active enterprise deployments, customer support automation, and proprietary fine-tuning pipelines built on public Chinese weights face no immediate regulatory interruption.
For long-term IT procurement and architectural roadmaps, however, Thai technology leaders must evaluate dependencies on upcoming frontier releases. If next-generation models crossing the 2-trillion parameter threshold undergo extensive licensing reviews before overseas release, Thai firms could experience latency in accessing state-of-the-art open weights. Enterprise engineering teams should adopt modular, model-agnostic orchestration layers to ensure workloads can dynamically route across varied open and proprietary providers without friction.
For Thai enterprises and developers reliant on Chinese open-weight foundations, the framework guarantees ongoing access to existing open-source releases and cloud APIs, but introduces regulatory friction and potential delays for next-generation frontier releases exceeding 2 trillion parameters.