China’s AI Export Curbs Favor Access Over Model Transfer Rules

By Tom Whitfield

China’s proposed export rules would add 23 artificial-intelligence technologies to a list banned or restricted from export, according to an October 16 draft. The measure covers advanced models, agents that carry out tasks with limited human direction, chip-design technology and tools for running vast computing facilities. A temporary registration requirement for certain transfers took effect immediately on October 16.

What does the draft allow?

The draft draws a commercial line between foreign customers using Chinese AI services and foreign companies obtaining the models, data or expertise needed to operate or reproduce them independently. It would broadly permit overseas access through websites and cloud platforms while requiring government approval for some transfers of underlying technology or overseas deployments.

DeepSeek, Moonshot and Z.AI already offer models to overseas customers through websites or cloud platforms. Some Chinese developers also use foreign computing facilities because U.S. export restrictions limit their access to advanced chips. U.S. rules bar Nvidia’s most advanced processors from export to China.

The practical result is that Chinese suppliers could keep intellectual property and research onshore while selling access abroad. That structure limits direct technology transfers without closing the service channel.

When will the rules take effect?

A source familiar with the discussions, who requested anonymity because the timetable was not public, said China is expected to adopt the proposals by the end of the year. Implementation guidelines would follow later.

The proposals remain open for public comment until November 15. China’s Ministry of Commerce did not respond to a request for comment.

For companies, year-end remains a target rather than a confirmed legal date. Implementation guidance will determine how much separation the rules require between domestic and overseas operations.

Which models fall within the controls?

Liu Jing, a senior partner at Beijing Dacheng Law Offices who participated in policy consultations, said companies also wanted clarity on the tests used to determine model coverage. The draft identifies three possible tests:

• Training compute: the computing work required to train a model reaches a threshold of 10^26 calculations.
• Model size: the model contains at least 2 trillion parameters.
• Assessment ranking: the model places among the world’s top 10 in certain assessments of performance, use or efficiency.

Parameters are the numerical settings a model learns during training. The ranking test could capture smaller systems that excel at specialized tasks, but the draft does not identify which rankings qualify.

Chinese law firm Han Kun also questioned how regulators would count the size of models that activate only some of their components for each task. The firm said a proposed chip-design provision could reach beyond the most advanced products because it lacks a minimum performance requirement.

That leaves model developers without a reliable way to judge which downloadable releases require approval. The ambiguity will persist until regulators define qualifying rankings and partial-model sizing.

Why are labs asking for clarification?

Researchers at major Chinese AI labs are seeking clarity on whether they may train or run their most advanced models on foreign-owned servers overseas. A person briefed on the nonpublic discussions said a worst-case outcome would require separate domestic and overseas versions of the same downloadable model.

The rules could also affect how leading Chinese AI firms structure overseas data-centre deployments, international partnerships, and foreign investment or ownership of AI assets. The practical effect is separate maintenance for identical downloadable models across two markets if the split becomes mandatory.

How do ownership rules affect deals?

Registration requirements, effective from October 16, cover transfers that put Chinese technology under foreign ownership or control, including qualifying transactions already underway, according to an analysis by law firm Fangda. China’s earlier move to order Meta to unwind its more than $2 billion acquisition of Chinese-founded AI startup Manus shows the risk facing foreign buyers.

For computing centres, the draft provides simplified approvals for Chinese companies that build and operate their own facilities abroad. It does not expressly authorize transferring technology to foreign customers to build equivalent systems, Han Kun said.

The controls also cover methods for coordinating more than 10,000 AI chips. Companies must separate approval to build from the registration required when ownership or control changes hands.

What is the policy’s commercial purpose?

“The structure the leadership prefers, with IP and R&D kept onshore and operations and access sold abroad, is the one the market was already converging on,” said Alfredo Montufar-Helu, Beijing-based managing director at consultancy Ankura.

As technology rivalry with the United States intensifies, security concerns accompany the commercial aim to keep powerful cyberattack capabilities and sensitive data in China. The draft therefore seeks to retain valuable technological advantages at home.

Yang Jingru, a partner at law firm Jincheng Tongda & Neal, called the approach a counterpart to U.S. restrictions on China’s access to advanced technology and investment. He wrote: “The US controls the inflow of computing power, hardware and capital, while China is beginning to systematically manage the outflow of models and technological assets.”

Market participants will watch how regulators define qualifying rankings and foreign server deployments. The coming guidelines will show whether Chinese model vendors can expand hosted access abroad without surrendering control of their most advanced models.