business 5 min read

China's Blacklisted Universities Still Buying Nvidia Chips

Chinese military-linked universities continue procuring Nvidia AI chips through shadowy local resellers, exposing a critical enforcement gap that undermines the US strategy to restrict Beijing's artificial intelligence capabilities.

  • Semiconductors
  • China Tech
  • NVIDIA
  • Export Controls
  • US-China Relations

The Leak in the Wall

The United States built its AI containment strategy around a single assumption: if you control the chips, you control the game. Nvidia’s H200 is supposed to be the gatekeeper — a chip so powerful that owning it means deciding who trains what models and how fast. In October 2025, Washington required case-by-case licensing for every H200 shipment to China. The logic was straightforward.

It doesn’t appear to have worked.

According to Reuters’ analysis of 26 procurement records published this year, Chinese universities linked to the military — institutions already on the US Commerce Department’sEntity List since at least 2001 — are regularly acquiring Nvidia AI chips through local resellers. The chips move from authorized distributors into the hands of restricted buyers via middlemen whose names don’t appear on any public blacklist. The servers follow the same path, leased by the month rather than purchased outright, making attribution even harder.

The most recent example surfaced on November 21, when Northwest Polytechnical University in Xi’an posted a procurement record showing it bought two H200 chips for 874,700 yuan — roughly $120,000 — from a company called Shanxi Bosom Network Technology. That university is one of China’s “Defense Seven,” the cluster of institutions designated by Beijing to lead military research in areas like aircraft, missiles, and satellite systems. It has been on the US export control list since 2001.

The H200 is not a cut-rate product. It outperforms the H20, which is Nvidia’s deliberately downgraded chip for the Chinese market. The H200 is the real thing — the kind of hardware the US government specifically tried to restrict.

How the Loophole Works

Northwest Polytechnical didn’t just buy the chips. In January alone, its cybersecurity school signed two contracts to lease servers loaded with eight H200 units each, for periods of seven months. The servers were destined for fine-tuning large language models and reinforcement learning training — exactly the work the export controls were designed to slow.

Other universities ran similar plays. Shenzhen University purchased three H100 chips for 769,800 yuan in September. Beihang University, another Defense Seven member, bought four A800 chips for 489,000 yuan in April. Wuhan University of Technology acquired a server loaded with H100s for 355,000 yuan in June. The Chinese Academy of Sciences University contracted for cloud-based compute time at 10 yuan per hour to access H800 chips.

The pattern is consistent enough to matter. Almost all of these transactions flowed through small, relatively obscure Chinese companies rather than direct orders from the universities themselves. That’s the feature, not the bug. The middlemen absorb the compliance risk and provide plausible deniability. When a university lists a reseller on a procurement record, the record shows a local networking equipment supplier — not a US chipmaker, not a restricted end-user, not anything that triggers an automated alert.

American officials can’t trace the physical chips after they enter China. Procurement records don’t reveal origin points or distribution routes. They can’t confirm whether the leased servers sit on university grounds or in commercial data centers. They can’t confirm whether Nvidia itself facilitated any of these transactions. The gap between policy intent and observable reality is wide and growing.

Why Enforcement Fails Where It Matters Most

The deeper problem isn’t that individual universities found a workaround. It’s that the workaround is structurally inevitable given how the current rules are written.

China’s chip market operates on two layers. On the surface, licensed distributors sell restricted hardware to vetted buyers. Beneath that, an informal network of resellers, system integrators, and cloud providers re-moves hardware to anyone who can pay. The H200 purchased by Northwest Polytechnical may have originated through legitimate channels and then been diverted domestically. Or it may have entered China through a different route entirely. There’s no way to tell from available data.

Server leasing compounds the problem. Even if the US blocked every sale of individual chips to restricted entities, a Chinese company could buy a server containing those chips from an authorized distributor, install it in a commercial data center, and lease compute capacity to universities without ever transferring the hardware itself. The chip never changes hands. The capability still flows.

This is not a theoretical concern. Reuters found evidence of it happening now.

What Changes Next

Washington has limited tools left. Extending end-use monitoring deeper into China would require on-the-ground inspections that Beijing would almost certainly reject. Expanding the Entity List further is repetitive — most of the relevant institutions are already on it. Criminalizing the resellers might deter some actors but wouldn’t eliminate the network; it would only push it further underground.

The more realistic path is to pressure the first link in the chain: the Chinese distributors and resellers that handle the paperwork. If the US makes it legally risky for a Chinese company to process transactions involving known restricted entities — even indirectly — the middlemen may start declining deals that look too exposed. But that requires evidence the US currently lacks, and diplomatic leverage that depends on trade relationships extending well beyond semiconductors.

For now, the leak persists. The H200 continues moving through the cracks in a system designed to catch it. Chinese military universities keep getting the hardware they need. And the US approach — build a wall, monitor the gates — keeps failing at the point where it matters most.