China proposes new export curbs on advanced AI technology

Defining the new export boundaries

China has released draft rules intended to curb the export of its most advanced artificial intelligence technologies. As reported by WSAU, these regulations aim to create a clear distinction between how Chinese companies sell AI services and how they share the underlying technology. The goal is to allow the government to keep valuable intellectual property and research and development efforts onshore while still permitting commercial sales to foreign customers.

Under the proposed framework, overseas users would generally be allowed to access Chinese AI through standard websites and cloud platforms. However, the rules would subject the transfer of the actual models, data, or specific expertise needed to reproduce those models to government approval. This distinction is significant for developers who currently rely on international computing facilities to bypass hardware limitations.

The draft specifically identifies 23 AI-related technologies for restriction or ban. These include advanced models, agents capable of performing tasks with limited human direction, and the specific technologies required to design chips or operate massive computing centers. A temporary requirement to register certain technology transfers involving foreign ownership or control went into effect on October 16.

Technical thresholds for restrictions

To determine which AI systems fall under these new export curbs, the draft outlines three specific tests. One test looks at the sheer scale of the model, setting a threshold of at least 2 trillion parameters. Parameters are the numerical settings a model learns during its training process, and this high number ensures only the most massive systems are caught in the net.

The second test involves the amount of computing work required to train a model, specifically looking for a threshold of 10 to the power of 26 calculations. The third test is a performance-based ranking, which would cover systems that place among the world's top 10 in specific assessments of efficiency, use, or performance. This ranking could potentially capture smaller systems that are highly specialized for certain tasks.

Industry experts have already raised questions about how these tests will be applied in practice. Some analysts have noted that it is unclear which specific rankings would count for the performance test. Others have questioned how regulators will count the size of models that only activate certain components for specific tasks, as well as provisions for chip design that lack a minimum performance requirement.

Impact on global AI partnerships

The new rules could significantly alter how leading Chinese AI firms structure their international partnerships and handle foreign investment. Companies like DeepSeek, Moonshot, and Z.AI already offer models to overseas customers through cloud platforms. These firms may now face more complex hurdles when trying to run advanced models on overseas data centers or coordinating more than 10,000 AI chips.

Because Nvidia's most advanced processors are restricted from being exported to China by the US, some Chinese developers have turned to foreign-owned servers to train their models. Researchers at major Chinese labs are currently seeking clarification on whether they can continue to run their most advanced models on these foreign servers. If the rules are strict, developers might be forced to maintain two completely separate versions of the same downloadable model for domestic and international use.

The regulations also extend to corporate deals that put Chinese technology under foreign control. This follows a trend of increased scrutiny, such as a previous order for Meta to unwind its acquisition of a Chinese-founded AI startup. These moves suggest that Beijing is increasingly wary of foreign entities gaining control over the core assets of its AI industry.

Strategic goals and next steps

The proposed curbs are viewed by some analysts as a strategic countermeasure to the restrictions placed by the United States. While the US focuses on controlling the inflow of hardware, capital, and computing power to China, these new rules represent a systematic effort by China to manage the outflow of its own models and technological assets. The move serves both a commercial aim to retain technological advantages and a security aim to keep sensitive data and cyberattack capabilities within Chinese borders.

The Chinese government is expected to officially adopt these proposals by the end of the year. Once adopted, the government will issue specific guidelines on how the rules will be implemented across the industry. Until then, the proposals remain open for public comment until November 15.

For the broader tech industry, these rules signal a move toward a more regulated and segmented global AI market. Companies operating in both regions will need to navigate a complex web of approvals to ensure they do not inadvertently transfer restricted technology while attempting to provide services to a global audience.

The path forward for developers

Chinese AI developers currently face a period of uncertainty as they wait for the final implementation guidelines. The primary challenge will be distinguishing between a service—like a chatbot accessible via a website—and a technology transfer—like providing the weights of a model to a foreign partner. Companies will need to audit their current international operations to see if they fall under the new registration requirements.

Furthermore, the lack of clarity on how " }