Nvidia H200 approvals face questions over Chinese buyer screening | Magica
Empty offices and blacklist ties put U.S. Nvidia H200 approvals under scrutiny
Editorial Team
••📖7 min read
Recent checks found empty registered offices and blacklist ties among Chinese firms previously cleared to handle Nvidia’s H200 AI chips. The findings challenge buyer screening, but do not establish military diversion; licensed sales have also been constrained by China.
Recent checks found empty registered offices at two Chinese firms approved to handle Nvidia’s H200 chips.
Corporate links to restricted entities raise screening questions; the investigation did not establish military diversion.
Nvidia says licensed shipments were less than 1% of quarterly data-center revenue, while it absorbed a 25% tariff.
Chinese companies approved by the United States to buy or handle H200 data-center AI processors from U.S. chip designer Nvidia have corporate links to restricted entities, and recent visits found empty registered offices at two approved handlers. The investigation raises questions about how Washington identified the businesses receiving access to controlled technology.
The findings test a policy that reopened exports to approved Chinese customers while requiring customer screening and independent hardware testing. They do not establish that approved processors reached China’s military. Meanwhile, Nvidia’s financial disclosures show that U.S. permission has translated into only limited sales under the program, with Chinese restrictions constraining demand.
Nvidia’s HGX H200 platform in its November 2023 launch materials. Image: Nvidia. Source: Nvidia.
What the approved companies’ records showed
New Matrix, a Chinese company approved earlier in 2026 to assemble Nvidia chips into AI computers, had been established only 17 months before the report. When investigators visited its listed address in an industrial park on Beijing’s northwest outskirts, they found an empty office. The building directory omitted the company, and property managers said they had no record of it.
Chinese corporate records linked New Matrix’s owners to a private equity fund blacklisted by the United States in 2024. Two analysts who examined those records and other Chinese-language information believed the company might be a front for businesses associated with the fund. That is an assessment of possible control, not an established finding about its customers or chip destinations.
A detailed account of the New York Times investigation identified the fund as Wise Road. The reason for its restrictions matters: a December 2024 added Wise Road Capital and two other entities to the Entity List for helping Beijing acquire businesses with sensitive semiconductor manufacturing capabilities critical to U.S. and allied defense industries, with the objective of relocating them to China.
Founded in 2025 by a team with autonomous-driving data experience, GenRobot AI says it raised $100 million in a Series A+. Its robot-data business reports million-hour quarterly deliveries, while funded rivals and undisclosed commercial metrics complicate the growth story.
The Entity List imposes additional export-license requirements on transactions involving listed entities. For Wise Road, the rule required licenses for all items subject to U.S. Export Administration Regulations, with a presumption of denial. The reported ownership connections therefore warrant scrutiny, but are not themselves proof of an unauthorized transfer.
The detailed account described three other approved companies whose connections drew attention:
Maginfra: Its registered office also appeared deserted. Analysts associated the relatively new company with Inspur, a Chinese technology company placed on the Entity List for involvement with China’s military. Corporate records indicated shared personnel and patents, and government ownership in both businesses.
Zhejiang Yindun Cloud Technology: It had connections to Range Intelligent Computing Technology Group, a Chinese cloud provider serving military contractors. Range supplied cloud services to a satellite program operated by Hebei province’s defense bureau and described that program as supporting national defense development.
Shanghai Jiatang: It originated from state-owned Datang Telecom, a technology supplier to China’s military. Datang is controlled by the China Academy of Telecommunications Technology, a state-owned enterprise involved in military cryptography.
The visits document conditions at registered addresses after the approvals. They do not establish what operations existed elsewhere, or where any delivered processors ultimately went.
Nvidia’s H200 announcement and subsequent U.S. export-policy milestones. Sources: Nvidia and the Bureau of Industry and Security. Source: Nvidia.
An established chip, a newly reopened export route
Nvidia is the U.S. chip designer that expanded from PC graphics into AI computing and data-center infrastructure. Its H200 is a data-center graphics processor built on the Hopper architecture for generative AI and high-performance computing. Nvidia announced it in November 2023, with systems expected to begin shipping in the second quarter of 2024.
This is access to an established product generation. In its July-quarter filing, Nvidia said Blackwell systems accounted for most system shipments and that its next-generation Vera Rubin architecture began production shipments in the following fiscal quarter.
U.S. President Donald Trump announced on December 8, 2025 that H200 and similar products could be shipped to approved Chinese customers, easing earlier restrictions. On January 13, 2026, the Commerce Department’s Bureau of Industry and Security, which administers export licensing, announced case-by-case review for Nvidia’s H200, AMD’s MI325X and similar chips, subject to security requirements.
Applicants must demonstrate that exports will not reduce global semiconductor production capacity available to U.S. customers, that Chinese purchasers have adopted export-compliance procedures including customer screening, and that products have undergone independent testing in the United States to verify performance and security.
Hardware testing and purchaser screening answer different questions. A processor’s verified performance does not establish who controls the company receiving it. The corporate connections identified in the investigation bear directly on that second safeguard.
The detailed account said roughly two dozen Chinese companies had received approval to purchase or handle H200 processors, including Baidu, BYD and Tesla Shanghai. People familiar with the approvals told investigators that companies could buy tens of thousands of chips each but had not reached their limits. Those permissions are not shipment totals.
Limited revenue, a tariff and competing routes to compute
Nvidia said in its quarterly filing that U.S. H200 licenses began arriving in February 2026. Chinese government restrictions prevented it from selling all the licensed products, and it shipped only a fraction of the permitted amount.
Those licensed shipments accounted for less than 1% of Nvidia’s $89.0 billion in Data Center revenue for the quarter ended July 26, 2026. The percentage measures revenue, not the share of approved chips delivered or the share reaching particular end users. The company also recorded a $400 million charge for H200 excess inventory and purchase obligations in the first half of fiscal 2027 as demand diminished.
The inspection requirement carries a separate cost. Licensed H200s must pass through the United States before shipment to customers, triggering a 25% tariff on importation. Nvidia said it had been unable to pass any of that tariff to customers and did not expect to do so on future licensed sales to China.
Nvidia has a commercial incentive to preserve Chinese market access beyond these limited sales. In the same filing, it argued that restrictions had helped competitors build developer and customer ecosystems capable of challenging it worldwide. That is the supplier’s assessment of the competitive consequences, not an independent measure of rivals’ gains.
Nvidia spokesman John Rizzo, quoted in the detailed account of the investigation, defended exports by arguing that China’s military had little incentive to depend on American technology. He said Beijing restricted H200 imports to promote domestic semiconductor manufacturers and that Nvidia’s limited exports supported American employment and manufacturing.
The same account said Chinese businesses had also sought advanced U.S. processors through overseas data centers, particularly in Southeast Asia. Domestic suppliers and overseas computing access complicate any assumption that direct H200 imports alone determine Chinese access to AI computing.
The next test is ownership and delivery evidence
The Bureau of Industry and Security defended its review of national-security risks. In the detailed account, the agency cited $324 million in export-violation penalties in 2025 and more than 90 enforcement officers hired in 2026. Those figures describe broader enforcement activity; they do not explain how the companies identified in the investigation passed screening.
The remaining question also extends beyond the initial license. Nvidia acknowledged in its filing that it does not physically control products after sale and must rely on customers’ and partners’ compliance programs.
Resolving the screening question requires evidence of who controls the approved intermediaries, where licensed processors were delivered, and how subsequent access was checked. The available reporting identifies corporate connections and empty addresses, but does not trace an approved shipment to China’s military. Those records would allow scrutiny of both how the approvals were granted and whether any prohibited transfer followed.
AI safety advocates gained experienced lawmakers and battleground seats in the midterms, but rival technology-backed groups also elected allies, leaving federal legislation far from assured.
A federal judge found Anthropic’s Claude Max usage advertising plausibly misleading because subscribers could read the multipliers as monthly promises. Contract and misrepresentation claims survive, but several other claims were dismissed with leave to amend.
Thailand’s proposed data-centre framework would scale requirements by power demand, screen sites for electricity and water access, and assess domestic benefits. It also covers existing facilities, while conflicting application and approval schedules remain unresolved.
PJM’s independent market monitor attributes about $11 per megawatt-hour of first-half 2026 wholesale costs to data-center demand in capacity auctions. Energy costs drove more of the overall increase, while construction barriers and an October regulatory rejection complicate the response.
DataVita now says its proposed Lanarkshire AI data centres will be fully grid served, with direct renewable connections only a future possibility. The change in public presentation raises questions about promised power costs and community benefits, although the DV4 planning application already assumed grid supply.
AMD chief Lisa Su calls the chipmaker a data-center-first company as server and AI sales surge. Gaming revenue fell 40%, but PC-chip sales grew, and her remarks leave the supply outlook for expensive Radeon cards unresolved.
A reported U.S. intelligence assessment links Megaspeed executive Alice Huang to China’s Ministry of State Security, raising questions about access to Nvidia’s leadership and chip-export policy. Whether the alleged operation succeeded remains unclear.
Ocient’s $100 million Series C backs tailored AI analytics on an established database platform, with national-security investors and deployment economics shaping its expansion.
Anthropic plans IPO marketing next week and a listing in the week of November 16, according to people familiar with the plans. Buyers are discussing valuations of $1.2 trillion–$1.3 trillion, with some weighing $1.1 trillion, as its filing reveals faster-growing compute costs and extensive infrastructure commitments.
India's IT lobby chief warns that AI could make technology jobs harder for graduates to secure. Hiring is below earlier peaks, but some firms favour AI-equipped beginners over costly experienced staff. The unresolved question is how junior workers will build expertise.
Internet scans and leaked databases expose the spread of ARTEX, a security-testing framework used in cyberattacks. Researchers documented largely autonomous compromises, but the records also show failed cashouts and uncertain counts of independent deployments.
Meta's strategy chief and former CFO is joining Fireworks AI ahead of a reported fundraising push targeting at least a $30 billion valuation. The inference company announced a $17.5 billion valuation in July, making the hire part of a much larger expansion than its older funding figures suggest.
Sand.ai says its $200 million Series A+ round lifts its valuation above $1 billion. Co-lead investor Wangsu plans inference and deployment cooperation, while its filing shows substantial losses behind the startup’s video AI ambitions.
Onboard fisheries AI and faster satellite fire alerts show how conservation groups are shortening response times. Human oversight and deployment choices matter, while project costs and environmental footprints remain unquantified.
An advocacy group estimates Interior Secretary Doug Burgum’s AI holdings at $18.1 million–$64.4 million as his department advances data centers on public land. His ethics agreement allows retained investments with recusals, while pending permits will determine both the application of those safeguards and the public’s return. Interior denies his decisions are driven by personal investments.
Microsoft is evaluating fewer neocloud partners while expanding its own infrastructure. Delivery-dependent contracts underpin supplier financing, while IREN reports rising revenue alongside a $6 billion funding target for its broader expansion.
Physical Superintelligence says its AI system found a proof of a conjecture limiting quantum query advantages on random inputs. Public proof files enable scrutiny, but do not establish the private system’s cost or repeatability.
OpenAI is expanding its relationship with Hacktron, whose researchers previously reported an attack chain into employee accounts, to test third-party connections with greater access to cybersecurity models.
Google has acquired Crusoe’s southeast Wyoming data center site, where it plans a long-term $17 billion investment. Crusoe describes completed pre-development work; independent reporting points to customer setbacks. Power delivery and carbon capture remain key tests.
