The tables have turned. On Friday, China’s Ministry of Commerce and Ministry of Science and Technology proposed adding 40 technologies to its export control catalogue, targeting the country’s most advanced AI models and chip designs. This draft rule, currently open for comment until November 15, aims to restrict the export of large AI models, agents, computing clusters, and semiconductor designs that meet specific strategic thresholds.

The 2-Trillion-Parameter Gate

For developers and infrastructure teams, the most critical detail is the definition of a 'large model' under these new rules. The draft restricts exports of model weights, source code, and core technology for any model with at least 2 trillion parameters, or those ranked in the global top 10 for performance, usage, and safety. Bloomberg identifies Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8-Max as examples already exceeding this size threshold. If your stack relies on deploying these weights internationally, you are now navigating a complex licensing landscape.

Chip Design Restrictions Kick In

The controls extend beyond software to the hardware blueprints themselves. Design technologies for AI chips manufactured at 14 nanometers or more advanced, as well as those utilizing 2.5D or 3D advanced packaging, are included in the restricted list. This move effectively treats chip design IP as a national security asset, preventing easy transfer of know-how for advanced memory chips and logic dies to foreign entities. For infrastructure builders, this means that even if you can access the compute, the underlying design methodology for next-gen accelerators is now subject to state oversight.

Open Source and API Exemptions

Despite the heavy restrictions, Beijing is not closing the door entirely. Technologies already public via open source or academic papers are excluded from the export ban. Furthermore, access through APIs, websites, apps, and SaaS platforms remains largely unrestricted, allowing overseas users to continue interacting with Chinese models without triggering export licenses. However, deploying a covered model on servers abroad counts as an export, requiring simplified or automatic licensing for Chinese companies operating their own overseas computing centers.

Immediate Registration Requirements

While the full catalogue is still a draft, a temporary pre-export registration requirement took effect immediately. This rule applies to deals already underway but not yet completed, including IP transfers, overseas licensing, and changes in control of Chinese tech owners. This creates an immediate compliance hurdle for any ongoing negotiations, forcing companies to file paperwork before the final rule is even ratified. The ministry states these measures aim to 'prevent misuse and abuse' and ensure AI remains under human control, though the geopolitical subtext is clear.

Key Takeaways

  • Models with 2T+ parameters or top-10 global rankings face export restrictions on weights and source code.
  • Chip designs for 14nm+ nodes and advanced packaging are now controlled technologies.
  • Open source releases and API-based access remain exempt from the new export controls.
  • A temporary registration rule is already live, affecting ongoing IP transfers and licensing deals.

The Bottom Line

China is mirroring US strategy by treating frontier AI weights and advanced chip designs as sovereign assets, forcing global infrastructure teams to navigate a new web of licensing and registration requirements for open-source and API-based access alike.