Chinese defense universities bypass US chip bans via local suppliers

Bypassing export restrictions

Chinese research institutions that are specifically blacklisted by the U.S. government are successfully acquiring high-performance AI semiconductors. Despite strict export controls designed to limit China's access to advanced military AI technology, these entities are finding ways to secure the hardware. As reported by The Chosun Ilbo on MSN, these acquisitions are often made through local Chinese suppliers rather than direct international channels.

One notable example involves Northwestern Polytechnical University in Xi’an. This institution is a leading defense research center that specializes in the development of missiles, satellites, and military aircraft. Despite being on the U.S. Department of Commerce's export control list since 2001, the university recently purchased two NVIDIA H200 AI semiconductors. These chips are high-performance components that outperform the H20 versions specifically designed with reduced capabilities for the Chinese market.

The procurement records for these chips were made public on the university's own website. The transaction was facilitated by a local company called Shanxi Boson Network Technology. This indicates that even when direct sales are blocked by international policy, domestic networks can still move restricted technology into the hands of sensitive organizations.

Diverse acquisition methods

The methods used to bypass these restrictions are varied and include both direct purchases and complex leasing agreements. Northwestern Polytechnical University's Cybersecurity College signed contracts in January to lease servers equipped with eight H200 chips for a period of seven months. These servers were specifically intended for training AI models and performing reinforcement learning, which is a type of machine learning where an agent learns by trial and error.

In addition to the H200 chips, the university also leased servers containing eight A100 and eight A800 chips. Both of these models are also subject to U.S. export restrictions. By using a leasing model, the university can access high-end computing power without necessarily owning the hardware directly, which may complicate the tracking of the equipment's ultimate destination.

Other universities across China have shown similar patterns of procurement. Shenzhen University was found to have purchased three NVIDIA H100 chips in September, while the Wuhan University of Technology acquired a server with H100 chips in June. These instances suggest a widespread trend of academic and defense institutions seeking out the most powerful chips available on the market.

The role of local suppliers

A Reuters analysis of 26 different procurement and corporate records in China this year revealed a consistent pattern. Multiple universities and research institutions, including those with defense ties, are acquiring restricted U.S. chips. Most of these transactions are facilitated by lesser-known local Chinese suppliers rather than major international distributors.

This reliance on local intermediaries creates a significant hurdle for international regulators. Because the procurement records do not confirm the original source of the chips or the specific distribution channels used, it is difficult to prove a violation of U.S. export laws. The physical location of the leased servers also remains unknown, making it hard to verify if the hardware is being used in prohibited ways.

The complexity of these supply chains allows restricted entities to maintain a steady flow of high-end hardware. While the U.S. government requires case-by-case reviews for some high-end exports, the existence of these local resale networks provides a workaround. This allows Chinese defense-related entities to continue their research into advanced AI technologies despite official trade barriers.

Challenges in enforcement

The situation highlights the difficulty of controlling the actual usage of semiconductors once they enter a domestic market. An industry source noted that these cases demonstrate how challenging it is to manage hardware usage through local resales and leasing. Even with strict export controls in place, the movement of technology through secondary markets remains a persistent issue for policymakers.

The university's ability to acquire these chips also affects the broader goal of limiting China's military AI development. By securing H200 and H100 chips, these institutions can train more sophisticated models than they could with the lower-tier hardware intended for the Chinese market. This creates a technological gap where restricted entities can still access high-end performance through non-traditional means.

Moving forward, the focus of export controls may need to shift toward the service and leasing models used by these institutions. Since the university's procurement records do not confirm NVIDIA's direct involvement, the focus remains on the local companies facilitating the deals. The persistence of these bypasses suggests that current restrictions may not be sufficient to block the acquisition of high-performance AI tools by sensitive actors.