Nvidia H200 China Licenses Went to Firms With Empty Offices

Estimated reading time: 7 minutes · Last updated: 9 November 2026

The Nvidia H200 China licenses issued by the US Commerce Department include two companies that seem to exist mostly on paper. Visits to the registered addresses of New Matrix and Maginfra found vacant rooms, and both firms connect to businesses Washington has sanctioned for supplying China's military, as first reported by The New York Times. Neither company is the final buyer. Each is licensed to build the chips into AI servers for customers nobody has named. A third licensee, Zhejiang Yindun Cloud Technology, is tied by family and premises to a data center group whose clients have included military contractors. Together, the cases put a hard question to the licensing system: who checks where the chips actually go?

I have a ton of concerns that neither the Commerce Department or Nvidia is doing the due diligence needed to ensure that the licensing terms are being followed.

Liza Tobin, former China director, White House National Security Council

What Visitors Found at New Matrix and Maginfra

The listed addresses for New Matrix and Maginfra led to offices that were empty apart from some furniture. Receptionists in the buildings did not recognise either name. People working nearby said they had never heard of the companies and had never seen anyone use the deserted spaces.

New Matrix's website adds little. It reportedly carries no information about who owns the business or how long it has existed, which leaves public company filings as the main way to trace it.

Neither firm is supposed to keep the hardware. Their licences cast them as assemblers that fit H200 GPUs into AI servers and pass those machines on to an end user. The identity of those end users is unclear, so an approval that appears to cover two small firms in practice covers an unknown set of customers further down the chain.

How New Matrix Leads Back to Wise Road and Tsinghua Unigroup

Public company records name Li Bin as the founder of New Matrix. Li and the company's two other principal owners have connections to Wise Road, a Chinese private equity firm under US sanctions for helping Beijing obtain sensitive chip technology.

Li holds other posts that compound the problem. He chairs Tsinghua Unigroup, a technology conglomerate in which the Chinese government holds part of the equity. He also owns New H3C Semiconductor Technologies, which the Commerce Department placed on its Entity List for working with the Chinese military. The Entity List is the department's register of foreign companies facing added US export restrictions.

That produces an awkward picture. The Commerce Department restricts a company Li owns, yet the same department has cleared a separate company he founded to buy H200 chips. Whether licence reviewers spotted the ownership overlap, or spotted it and approved the application anyway, has not been made public.

Why Maginfra's Overlap With Inspur Matters

Maginfra is thought to have ties to Inspur, a company the US has kept on a blacklist since 2023 because it supports the People's Liberation Army. Staff and patents overlap between the two companies, and the state holds its stakes in each through a similar ownership structure.

If Maginfra functions as an extension of Inspur, the licence would let a related business buy American AI accelerators that the blacklist is meant to keep from Inspur. That remains a suspicion. No published evidence shows that chips sold to Maginfra would reach Inspur.

The server-assembly arrangement makes outside scrutiny harder. Because Maginfra's finished servers go to customers that have not been identified, even a Maginfra with no Inspur connection would leave the final home of its H200 chips unaccounted for. Export controls built around end users only work when somebody knows who those users are.

The Range Data Center Link at Zhejiang Yindun Cloud

The doubts reach beyond the two empty offices. Executives at Zhejiang Yindun Cloud Technology, which also holds an H200 licence, are reportedly relatives of the founder of Range Intelligent Computing Technology Group. Range operates data centers, and Chinese military contractors are among the past buyers of its computing capacity.

Yindun reportedly runs part of its business out of a Range data center. That places licensed chips inside a facility belonging to a company with military customers, even if Yindun itself is the formal licensee.

The case differs from New Matrix and Maginfra in one respect: nobody has described Yindun as an empty shell. Its exposure comes from family ties and shared premises rather than from a missing business. Compute sold as a service is difficult to police, because the party running workloads on the processors need not be the party that bought them.

From Outright Ban to Licences With a 25% Cut

President Donald Trump banned exports of Nvidia AI GPUs to China in April 2025, including versions tailored to stay beneath the performance ceiling Washington had set for AI processors. He reversed that decision in late 2025, with the US taking a 25% cut of each sale.

Shipments still stalled. Through the opening months of 2026, Beijing refused to let Chinese tech companies purchase the chips, and the first H200 deliveries took more than eight months to arrive in China. When China eventually dropped its import ban, Chinese firms submitted licence applications, of which roughly two dozen have been approved.

Most approved buyers are household names: Baidu, BYD, Tesla Shanghai, Alibaba, Tencent and ByteDance, plus ZTE, a company the US previously blacklisted. Liza Tobin, former China director at the White House National Security Council, said she has "a ton of concerns" about due diligence. She questioned whether either the Commerce Department or Nvidia is checking that buyers follow the terms of their licences.

Can the H200 licensing system hold?

The bottom line

The H200 policy rests on a trade: Chinese buyers get Nvidia hardware, and Washington takes 25% of each sale. The cases of New Matrix, Maginfra and Zhejiang Yindun Cloud Technology test whether the control side of that bargain works. Two licensees operate from vacant offices, a third partly runs from a data center owned by a company that has served military contractors, and the end users of all three are hard to see. None of this proves chips have reached the People's Liberation Army. It does show that, among around two dozen approvals, at least three involve ownership, staffing or premises that overlap with sanctioned or military-linked businesses. The next question is whether the Commerce Department or Nvidia explains what checks were run.

What to watch

• Watch for the Commerce Department to say whether the New Matrix, Maginfra or Zhejiang Yindun Cloud Technology licences will be reviewed; no date has been set.
• Watch for Nvidia to explain what checks it applies to the downstream customers of server assemblers; no timetable has been given.
• Watch for disclosure of who buys the AI servers New Matrix and Maginfra assemble; nothing has been scheduled.