Procurement Documents Reveal AI Chip Workarounds
Executive Summary:
- Procurement records suggest that U.S. export controls on frontier artificial intelligence (AI) chips imposed meaningful constraints until January 2026, as institutions in the People’s Republic of China’s (PRC) appeared to be adapting procurement practices to preserve access rather than replacing foreign hardware with domestic alternatives.
- Tender documents from universities and state-linked entities show repeated efforts to obtain Nvidia H200-class computing power. Some mention H200s explicitly, others specify capabilities that can only refer to H200s, and others still obfuscate by appending politically acceptable labels such as “domestic chips” or “H20” to specifications that indicate they actually refer to H200s.
- The persistence of these workarounds indicates that despite visible progress by domestic firms such as Huawei, PRC alternatives remain insufficient for the most demanding frontier AI workloads, making continued access to U.S. hardware strategically consequential.
At Huawei’s China Partner Conference 2026 (华为中国合作伙伴大会) in March, the company unveiled a new neural processing unit (NPU), the Atlas 350. This latest artificial intelligence (AI) accelerator card is powered by Huawei’s Ascend 950PR chip, which Huawei claims can deliver 2.87 times the computing power of Nvidia’s H20, marking a new stage in the commercialization of domestic AI compute (Guancha, March 21). The launch reinforced an official narrative that the country’s indigenous computing stack is maturing, and that a more self-reliant, high-performance domestic compute ecosystem is beginning to take shape (People’s Daily, March 11).
Public procurement records from before the U.S. government’s decision in early 2026 to lift its restriction on buyers in the People’s Republic of China (PRC) purchasing Nvidia’s H200 chips tell a different story. Despite aggressive promotion of domestic substitution from political leaders, in recent months Chinese institutions have sought to preserve access to Nvidia’s frontier hardware, indicating that U.S. export controls continue to impose meaningful constraints on PRC technology development.
Obfuscation Provides Workarounds to Procure H200s
A review of PRC tender documents reveals repeated efforts to acquire computing hardware with capabilities equal to Nvidia’s highest-performance tier, the H200. Some institutions embedded in the PRC’s broader AI and advanced computing research ecosystem, including Jilin University (吉林大学) and Zhejiang University of Technology (浙江工业大学), identify the chip explicitly in their procurement documents (Zhiliao Tender Information, December 16, 2025, November 11, 2025; Jilin University, accessed April 2; Chinese Association for Artificial Intelligence, April 23, 2025).
In other cases, tendering entities do not name Nvidia directly but instead specify technical requirements that mirror the H200’s performance profile. Specific references to benchmarks of 141 gigabytes (GB) of video random-access memory (VRAM)—the high-speed memory built into graphics processors that is particularly important for training AI models—are notable. This figure stands out because Nvidia’s earlier flagship H100 offered 80GB of VRAM and most PRC accelerators currently deployed remain below that threshold. A requirement for 141GB of VRAM functions as a strong indicator of H200-class hardware, even when the product itself is not named (Zhiliao Tender Information, December 17, 2025; Ascend Developer Community, January 30; Nvidia, accessed April 3; Metax Tech, accessed April 3; Cambricon, accessed April 3; Moore Threads, accessed April 3).
Some institutions employ a more subtle and politically adaptive strategy. Southern University of Science and Technology (SUSTech; 南方科技大学) and Jiangsu University (江苏大学), for instance, describe procurement needs using language more consistent with domestic policy preferences, such as referring to “domestic chips” (国产芯片). Henan University of Economics and Law (河南财经政法大学), meanwhile, specifies export-compliant alternatives in its procurement documents, such as the H20. In all cases, however, the documents list technical benchmarks associated with substantially more powerful systems (Zhiliao Tender Information, November 3, 2025; November 25, 2025; December 18, 2025). [1] Similar patterns have also appeared in tenders linked to China National Nuclear Corporation (中核集团), suggesting that this strategy of obfuscation is not limited to universities alone (Zhiliao Tender Information, October 31, 2025).
Adaptation Involves Legal Gray Areas
These procurement patterns highlight a tension in Beijing’s technology policy. Beijing has spent years promoting self-reliance in core technologies through policies such as “xinchuang” (信创), which pushes public sector users toward indigenous information technology products (Sina, June 20, 2025). When domestic substitutes cannot fully meet the requirements of frontier AI research, institutions appear to adapt by finding workarounds that preserve access to restricted foreign technology.
According to an interview conducted with two industry participants, compliance checks in the PRC often focus primarily on documentary consistency, such as whether contracts, invoices, and customs declarations match (Author interview, February 21). This creates room for ambiguity between politically acceptable descriptions and actual end use, allowing for plausible deniability from compliance teams and authorities.
Delivery timelines in some tenders also raise questions about the nature of the supply chain for high-performance AI hardware. The White House lifted its ban on the export of H200 chips on January 20, but the deadlines specified in some tender documents appear too short for newly sourced high-end chips to move through fully compliant export channels. This suggests that inventory is already circulating through underground networks and ready to be deployed once procurement procedures are completed.
Conclusion
The set of tender documents reviewed in this article suggest that the PRC’s AI ecosystem is still reliant on Nvidia chips. Although domestic firms, and Huawei in particular, have made visible progress in AI accelerator development, foreign chips apparently continue to be preferred for the most demanding workloads (21st Century Economy, September 18, 2025). If this were not the case, procurement practices would have shifted decisively toward cheaper, domestic alternatives where stable supply is assured. If PRC actors are yet to solve the frontier chip problem and instead are attempting to work around it, this suggests that the technological gap remains real and that controls on advanced semiconductors continue to impose meaningful constraints on the PRC’s technological progress, even if those constraints are uneven and imperfectly enforced.
For now, continued access to advanced U.S. chips can help PRC institutions maintain momentum until domestic alternatives mature. This fails to divert the PRC from its long-term push for self-sufficiency, and simply alleviates near-term constraints, buying time for PRC firms to develop without sacrificing progress at the technological frontier. This is critical, because time matter in AI development. More computing power enables larger models, faster iteration, and stronger research ecosystems, all of which can compound over time.
The PRC’s procurement behavior therefore offers a useful empirical window into the state of the technology competition between the two global leaders. It suggests that, notwithstanding official rhetoric about self-sufficiency, U.S. frontier chips remain valuable enough that institutions continue to find workarounds to secure access to them. This evidence reflects a continuing U.S.–PRC capability gap, as well as diverging realities between Chinese indigenization interests and industry needs.
Notes
[1] Southern University of Science and Technology participate in national defense-related research programs under PRC’s military-civil fusion framework (ASPI, accessed April 7).
