Chinese universities and state-linked research institutions are employing creative workarounds to preserve access to Nvidia's most advanced chips, according to procurement documents reviewed by researchers. Despite Beijing's aggressive push for technological self-sufficiency, tender records from late 2025 show that domestic neural processing units (NPUs), including Huawei's latest Atlas 350 accelerator, still cannot match the performance demands of cutting-edge artificial intelligence research .

Why Are Chinese Institutions Still Chasing Nvidia Chips?

The tension between China's official technology policy and its actual procurement practices reveals a critical gap in the domestic AI chip ecosystem. Beijing has spent years promoting "xinchuang," a policy that pushes public sector organizations toward indigenous information technology products. Yet when domestic alternatives fall short, institutions appear to be finding creative ways around the restrictions .

The evidence is embedded in the technical specifications themselves. Universities including Jilin University and Zhejiang University of Technology explicitly requested Nvidia H200-class hardware in their procurement documents. Others used more subtle approaches, specifying technical benchmarks that only the H200 could meet. A particularly telling requirement was 141 gigabytes of video random-access memory (VRAM), the high-speed memory built into graphics processors that is essential for training large AI models. Nvidia's earlier H100 chip offered only 80 gigabytes, and most Chinese accelerators currently deployed remain below that threshold, making the 141-gigabyte specification a clear signal of H200-class hardware .

Some institutions employed even more politically adaptive language. Southern University of Science and Technology and Jiangsu University described their procurement needs using terms like "domestic chips," while Henan University of Economics and Law specified export-compliant alternatives such as the H20. Yet all of these documents listed technical benchmarks associated with substantially more powerful systems, creating plausible deniability while preserving access to restricted foreign technology .

How Are Institutions Circumventing Export Controls?

• Documentary Ambiguity: Compliance checks in China often focus primarily on whether contracts, invoices, and customs declarations match each other, rather than verifying the actual end use of the hardware, creating room for misrepresentation between politically acceptable descriptions and real deployment.
• Coded Technical Specifications: Institutions specify performance benchmarks that only advanced chips like the H200 can meet, while avoiding explicit product names that would trigger regulatory scrutiny or raise red flags during procurement reviews.
• Pre-positioned Inventory: Some tender deadlines appear too short for newly sourced high-end chips to move through compliant export channels, suggesting that inventory is already circulating through underground networks and ready for deployment once procurement procedures are completed .

The pattern extends beyond universities. Procurement documents linked to China National Nuclear Corporation, a state-owned enterprise, show similar strategies of obfuscation, indicating that the workaround approach is not limited to academic institutions alone .

What Does This Reveal About the State of Domestic AI Chips?

Huawei's March 2026 launch of the Atlas 350 neural processing unit, powered by its Ascend 950PR chip, was presented as a major milestone. The company claimed the chip could deliver 2.87 times the computing power of Nvidia's H20. Yet the procurement behavior tells a different story: despite aggressive promotion of domestic substitution from political leaders, Chinese institutions continue to seek access to Nvidia's frontier hardware, indicating that U.S. export controls continue to impose meaningful constraints on technology development .

If domestic alternatives were truly sufficient for frontier AI research, procurement practices would have shifted decisively toward cheaper, domestic options where stable supply is assured. The fact that institutions are still finding workarounds suggests the technological gap remains real. This does not mean Chinese firms are failing; rather, it indicates that the most demanding workloads still require the performance envelope that only Nvidia's latest chips can provide .

The implications are significant for the broader technology competition between the United States and China. Continued access to advanced U.S. chips allows Chinese institutions to maintain momentum in AI research while domestic alternatives mature. This buys time for firms like Huawei to develop without sacrificing progress at the technological frontier. In AI development, time matters enormously. More computing power enables larger models, faster iteration, and stronger research ecosystems, all of which can compound over time .

The White House lifted its ban on exporting H200 chips on January 20, 2026, which may reduce the need for such workarounds going forward. However, the procurement documents from before that decision offer a window into the real state of the technology competition. They suggest that despite official rhetoric about self-sufficiency, U.S. frontier chips remain valuable enough that institutions continue to find creative ways to secure access, and that the domestic AI chip ecosystem, while advancing, has not yet closed the performance gap with the world's most advanced hardware .