China Weighs Export Curbs on Advanced AI Models and Chip Designs

China is considering tighter export controls on advanced artificial intelligence and semiconductor technologies, potentially restricting foreign access to Chinese model weights, training data and chip designs as Beijing moves to protect capabilities it increasingly regards as strategic national assets.

Regulators led by the ministry of commerce have consulted leading Chinese AI and semiconductor companies on measures to restrict overseas transfer of strategic technologies and foreign acquisitions of promising domestic start-ups, the Financial Times reported on July 21, citing two people involved in the discussions.

The proposals would widen Beijing’s control over advanced AI models to the data used to train them, the chips needed to run them and foreign investments in companies developing strategically important technologies.

The ministry has spoken with AI companies including Alibaba, ByteDance and Zhipu AI (now Z.ai) about limiting the transfer of key training data overseas and whether foreign users should be allowed to download the weights of their most advanced models, according to the FT.

Model weights are the parameters produced during training that determine how an AI system operates. Restricting their distribution would not necessarily prevent overseas customers from accessing Chinese models through cloud services or application programming interfaces. But it would make it difficult for developers to run, modify or build products directly on the underlying systems.

That distinction matters because Chinese AI developers have expanded internationally partly by releasing capable, relatively inexpensive models whose weights can be downloaded and customised. Alibaba’s Qwen family and models from Zhipu and other domestic groups have emerged as alternatives to closed systems offered by US companies.

Earlier this month, Reuters reported that Chinese authorities had held meetings with Alibaba, ByteDance and Zhipu about potentially restricting overseas access to the country’s most advanced and upcoming models.

Those discussions also covered tougher penalties for the theft or unauthorised transfer of proprietary AI technology, as well as possible restrictions on who could finance Chinese AI start-ups, Reuters reported.

The latest deliberations suggest the scope under consideration has broadened. The commerce ministry has also sought industry views on limits that could prevent overseas semiconductor companies, including Qualcomm and Taiwan Semiconductor Manufacturing Co, from making or commercialising advanced chips based on designs developed by Huawei, Alibaba and ByteDance, according to the FT.

Such measures could complicate the international arrangements used by Chinese chip designers. China has invested heavily in domestic semiconductor capacity, but its companies continue to rely on overseas manufacturing facilities, intellectual property and supply chains for many of the chip-development process.

Potential controls could also cover foreign acquisitions of strategic Chinese technologies, particularly in agentic AI, which refers to systems designed to plan and perform multi-step tasks with limited human intervention.

The proposals could be incorporated into the next revision of China’s catalogue of technologies prohibited or restricted from export. Technologies listed as ‘restricted generally’ require government approval before they can be transferred abroad, allowing regulators to review transactions individually rather than impose an outright ban.

No final decision has been made, and authorities are weighing feedback from the companies involved, according to the FT. It is also unclear whether any eventual controls would apply to models and technologies already available overseas or focus primarily on future releases.

The debate has become more pressing as Chinese developers close the capability gap with leading US laboratories. Last week, Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter, natively multimodal model designed for long-horizon coding, knowledge work and complex reasoning.

Its release followed a series of increasingly competitive Chinese models from companies including DeepSeek, Alibaba, ByteDance and Zhipu, strengthening Beijing’s position in a technology industry whose most advanced products are increasingly being treated as economic and national-security assets.

For China, tighter controls would require a difficult balance. Its companies have benefited from making lower-cost models widely available, helping Chinese AI gain users and influence overseas. Broad restrictions could slow that expansion and make it harder for domestic developers to attract international customers, capital and research partners.

The proposals come as Beijing promotes open-source AI as an alternative to the largely closed ecosystems built by major US developers. President Xi Jinping last week presented wider access to Chinese AI as part of the country’s technology offer to developing economies.

The deliberations indicate that Beijing may be moving towards a more selective version of that strategy: continuing to promote Chinese AI services and less sensitive models abroad while drawing tighter boundaries around the data, model weights and chip technologies it considers strategically important.

Reuters said it could not independently verify the FT report. The commerce ministry, Alibaba, ByteDance, Zhipu, Huawei, Qualcomm and TSMC did not immediately respond to its requests for comment.