Publicly available procurement documents show that multiple institutions with ties to China's People's Liberation Army have sought Nvidia AI chips even after Washington imposed export controls designed to keep that hardware out of Chinese military hands, according to research reported by Tom's Hardware.

The findings put fresh pressure on the effectiveness of the U.S. export control regime, which the Commerce Department has tightened repeatedly since 2022 in an effort to cut off China's access to advanced AI accelerators. Nvidia's data center GPUs — including the A100, H100 and, later, the China-specific A800 and H800 variants — sit at the center of those restrictions. Washington has also moved to close loopholes that allowed Nvidia to ship cut-down parts to Chinese customers.

According to the research, the procurement solicitations come from a range of Chinese institutions linked to the PLA, not a single actor. The documents reviewed are public records of purchase requests, meaning the evidence covers demand-side behavior: institutions asking for the chips. The reporting does not establish how many units, if any, were actually delivered after the relevant export controls took effect.

That distinction matters commercially. Nvidia cannot legally ship its flagship accelerators to military end users in China, and the company has repeatedly said it complies with all export rules. What the documents demonstrate is sustained Chinese military demand for Nvidia's hardware — demand that persists through third-party channels, gray-market resellers, or front companies, routes that U.S. enforcement agencies have repeatedly warned about.

For Nvidia, the disclosure cuts two ways. China has historically accounted for roughly a fifth of the company's data center revenue, and each successive round of U.S. controls has forced Nvidia to write down that business — most visibly when Washington revoked licenses for the A800 and H800 in 2023, and again when the Blackwell-generation B200, and later the China-tuned H20, fell under restriction. Continued PLA-linked demand underscores both the strategic value of Nvidia's silicon and the enforcement gap between American rules on paper and chip flows in practice.

For Beijing, the documents reinforce a policy already visible in its procurement rules and national champion funding: reduce dependence on U.S. accelerators. Chinese AI chip developers such as Huawei, with its Ascend product line, have pushed to fill the gap, though yield and software-ecosystem constraints have limited substitution at the high end.

The geopolitical context directly changes the commercial picture here. Every documented attempt by a PLA-linked institution to acquire restricted Nvidia hardware strengthens the argument in Washington for tighter enforcement — broader entity list additions, stricter licensing, and pressure on intermediaries in the supply chain. That, in turn, raises the risk that Nvidia's remaining compliant China business gets squeezed further.

It also sharpens the case in Beijing for accelerating domestic alternatives, which could eventually erode Nvidia's long-term position in one of the world's largest AI markets, even as near-term demand for smuggled or gray-market parts suggests Chinese buyers still see no equal to Nvidia's accelerators for training large AI models.

The research draws only on publicly available documents, so it likely undercounts the full extent of demand. As U.S. agencies step up enforcement and China ramps up domestic accelerator programs, the gap between restricted-on-paper and available-in-practice will determine how much of this demand Nvidia — and its would-be Chinese replacements — actually captures.