China’s new export curbs draw a line around its AI crown jewels

The draft rules broadly permit overseas users to access Chinese AI via web and cloud platforms. However, they subject transfers of underlying technology or overseas deployments of models, data, or expertise to government approval. This framework aligns with a business model, already favored by Beijing and major industry players, that keeps intellectual property and research within China while selling access to operations and services abroad, according to Ankura managing director Alfredo Montufar-Helu.

To determine which AI models fall under export restrictions, the draft specifies three tests: a computing work threshold of 10 to the power of 26 calculations for training, a size of at least 2 trillion parameters, or ranking among the world’s top 10 in certain assessments of performance, use, or efficiency. Parameters are defined as the numerical settings a model learns during training.

These criteria introduce practical ambiguities for developers and researchers. Beijing Dacheng Law Offices senior partner Liu Jing, who participated in consultations with policymakers, noted that companies seek more clarity on the tests, such as which specific rankings would count for the "top 10" assessment. Additionally, Chinese law firm Han Kun questioned how regulators would count the size of models that activate only some components for each task, suggesting the current definitions may not fully account for diverse model architectures.

This uncertainty presents challenges for Chinese AI developers like DeepSeek, Moonshot, and Z.AI, who offer models to overseas customers. Researchers are seeking clarification on whether they can train or run their most advanced models on foreign-owned overseas servers, a practice sometimes necessary to access advanced chips restricted from export to China. An anonymous source indicated that a "worst-case scenario" could force developers to maintain separate domestic and overseas versions of the same downloadable model.