China’s military seeks Nvidia chips as U.S. tightens access | Magica
China’s military seeks Nvidia chips as U.S. targets overseas computing access
Editorial Team
••📖7 min read
A congressional commission links China’s military AI ambitions to demand for advanced Nvidia chips. Procurement records show several routes to American computing, while memory costs and new remote-access rules complicate the alternatives.
China’s military is seeking advanced Nvidia chips for AI capabilities, a congressional commission says.
Procurement records reveal requests for processors, cloud computing and civilian AI models; they do not establish total deliveries or military computing capacity.
Domestic substitutes face memory costs, while new U.S. controls target computing supplied through overseas data centres.
China’s People’s Liberation Army is seeking advanced processors from Nvidia, the U.S. supplier of graphics chips used for AI, to develop military capabilities despite U.S. export restrictions, according to the U.S.-China Economic and Security Review Commission’s annual report released November 18. The finding puts military demand for American computing alongside a new U.S. effort to restrict remote access through third countries.
The commission, created by Congress in 2000 to monitor the national-security implications of U.S.-China economic relations, cites an April assessment of procurement documents spanning several years. Its security review brings that evidence into a current congressional assessment; it does not by itself demonstrate a new shipment or establish that a particular weapon uses Nvidia hardware.
The stakes extend beyond autonomous weapons. The PLA calls its integration of AI into weapons, targeting, command and logistics “intelligentized warfare.” The commission describes exercises involving AI-assisted target recognition, firepower allocation, command decision support and cooperation between crewed and uncrewed systems. These efforts form part of military modernization toward a 2027 readiness milestone for a Taiwan contingency—a capability goal, not a scheduled invasion date.
The underlying April 20 analysis came from researchers at Georgetown University’s Center for Security and Emerging Technology. Drawing on thousands of AI-related PLA procurement documents from the preceding three years, it identifies two routes to American computing: acquiring chips directly and procuring civilian AI models developed using U.S. hardware.
The ransomware gang threatens a November 28 data release, but Nvidia has not confirmed an intrusion. Earlier listings of a supplier and a partner raise questions about data provenance without establishing a connection.
One May 2025 request sought 16 export-controlled Nvidia H100 graphics processors for a PLA research laboratory’s modeling work. A July 2024 request specified access to Nvidia V100 or A100 computing resources through a supplier. The researchers also found requests for commercially developed Chinese language models and noted that many documents required domestic technology. Their public-document analysis cannot establish the scale of PLA use of American chips or comprehensively measure military AI adoption.
That distinction between a request and a completed transaction matters. An earlier January 2024 tender review found more than 100 tenders procuring Nvidia A100 chips across Chinese state entities. Most purchases were small. The cited military unit’s requests were heavily redacted, and the review could not establish their winners or purposes, although completion notices documented purchases by universities.
Nvidia had developed slower A800 and H800 processors for China after restrictions on its A100 and H100 products; those replacements were themselves restricted in October 2023. Small purchases can support machine-learning tasks and improve existing models, the earlier review noted, even when they fall short of the large clusters needed to develop sophisticated language models from scratch.
Domestic substitutes face more than a processor gap
Nvidia’s appeal also rests on software: the commission identifies Nvidia’s CUDA software ecosystem among the obstacles China has yet to overcome, alongside advanced chip fabrication, memory and semiconductor-manufacturing equipment. In an account of the commission’s findings, the panel credits U.S. export controls with creating a substantial computing gap. Continuing demand does not establish that restrictions have been ineffective.
The commission places the Ascend processor family from Chinese technology company Huawei and the Siyuan family from Chinese chipmaker Cambricon behind more advanced products from Nvidia and fellow U.S. chip supplier Advanced Micro Devices, according to reporting on the assessment. That comparison supplies no common benchmark from which to calculate a performance ratio.
Domestic alternatives also face a component constraint. A September 10 pricing report, citing two people familiar with the quotations, said Huawei had raised the indicated price of its Ascend 950DT accelerator card above 250,000 yuan. The card combines a processor, memory and other components. Depending on contract terms, that was 20%–50% above quotations two months earlier.
Huawei had said the 950DT would become available in the fourth quarter of 2026. These were advance customer quotations, not evidence of military purchases or a like-for-like price comparison with Nvidia. Two sources also described a 20%–30% increase in indicated prices for Cambricon’s next-generation 690, which had not formally launched. Earlier Huawei products had become more expensive too.
The report attributed the increases to a worldwide shortage of high-bandwidth memory, the stacked memory that feeds processors data at high speed. Its sources said Chinese chipmakers had increasingly turned to grey-market supplies after Washington tightened controls on certain advanced memory exports in December 2024, paying several times the prices available outside China. Huawei said its 950 series would use proprietary memory technologies but did not detail where or how that memory was manufactured. The manufacturers did not respond to requests for comment.
More chips would not resolve every military AI constraint. The commission says PLA researchers have identified insufficiently diverse real-world training data as an obstacle to reliable perception by uncrewed systems.
Nvidia faces scrutiny over access beyond its sales channels
The earlier tender review found neither Nvidia nor its approved retailers among the identified suppliers and could not establish how those suppliers obtained the chips. Nvidia said it complied with applicable export-control laws, required customers to comply and would act if it learned of unlawful resales. Those statements are the company’s account, not independent verification of downstream compliance.
Nvidia has said export controls effectively excluded it from China’s AI data-centre market, according to the current report. That account also cites a separate review finding more than $3 billion in Chinese computing projects using restricted Nvidia chips. That figure describes projects; it is neither a military procurement total nor a valuation of completed chip deliveries.
The commission’s findings arrive amid intensified scrutiny following the departure of Jensen Huang, Nvidia’s former chief executive. The report links his departure to an affair with a former executive of a Singapore-based business partner and says U.S. authorities suspected her of spying for China. A separate account identifies the partner as Megaspeed and says U.S. investigators are examining potential chip diversion to China. That investigation does not establish that Nvidia knowingly supplied a prohibited military buyer.
The commission also describes Chinese companies obtaining remote access to Nvidia chips through third-country data centres, including in Malaysia. Commercial remote access is a route to computing capacity; it does not establish that a particular customer or facility serves the PLA.
On November 16, the Commerce Department’s Bureau of Industry and Security announced an interim final rule addressing remote access to advanced chips and servers. The bureau said the rule was written by and at the direction of the Treasury Department and National Security Council, with the aim of preventing circumvention through overseas data centres.
The published rule text sets restrictions on chip exports, reexports and transfers to overseas data centres supplying significant computing volumes to companies headquartered in designated countries of concern or Macau. It requires customer due diligence, including checks on indirect ownership or control, while exempting commercial recommender systems, such as retail or social-media recommendation services, from that license requirement.
It also adds a licensing requirement for advanced semiconductor shipments to Indonesia, Malaysia, Singapore, Thailand and Brazil, with an exception for U.S.-headquartered companies or their subsidiaries subject to reporting requirements. These provisions put suppliers’ knowledge of customers and their users at the centre of enforcement.
The evidence needed next is whether restricted users can still obtain and use meaningful computing capacity: completed deliveries, verified customer identities and actual workloads. Procurement requests reveal demand. The test for the new controls is whether they constrain access through intermediaries while China develops substitutes with their own cost, software and data limitations.
Google is promising faster Gemini Live conversations on home speakers and displays, with timers, reminders and smart-home controls available inside the chat. The upgrade is still forthcoming and requires Google Home Premium.
Humyn Labs combines $26 million in equity with $21 million in venture debt to expand human demonstration data and planned robot testing centres. Its reported sales show demand, but deployment economics remain unproven.
Nvidia’s compact Windows desktops bring local agents and creative tools to a dedicated PC starting around $4,400. Competing desktops and Nvidia’s earlier AI hardware narrow the novelty, while conflicting configuration reports complicate price comparisons.
Maria Cantwell and Mike Rounds want a national challenge to measure how quickly AI can turn software discoveries into reliable protection. Earlier federal trials already produced patches, but safe deployment across essential services remains the test; officials say recent outages show no indication of an attack.
ByteDance is reportedly negotiating about $500 million in additional Azure spending, mostly for OpenAI models. The talks highlight different customer-access policies at Microsoft and OpenAI, against a history of scrutiny over ByteDance's use of the technology.
Corvex plans to fund a $78 million Blackwell deployment with customer prepayment, a $38 million USD.AI debt commitment and $14 million of its own money. The contract could more than double its Mid-Atlantic 1 annual contracted revenue, but loan closing, delivery and customer acceptance remain ahead.
Hyundai Card and JPMorgan tested AI agents that compare currency quotes and request transfers under bank-enforced mandates. No real money moved, and authority across multiple banks remains an unresolved hurdle.
Industry sources say TSMC requires cash for orders exceeding customer credit limits, putting custom AI chip designers' ability to finance production at the center of the capacity race. The requirement remains unconfirmed by TSMC, and customer prepayments have a long precedent.
Binance Wallet says subscriptions for its Anthropic IPO xStocks campaign reached the 500 million USDC limit. Final allocations, token delivery and applicable refunds remain pending; the cap does not establish proceeds received by the AI company.
Tencent says US restrictions on remote computing will likely slow its next AI model while leaving cloud revenue largely protected. Southeast Asian operators face licensing and financing uncertainty, though US cloud access and domestic demand offer partial alternatives.
Microsoft is rolling out Agent 365 Business Analytics through its Frontier early access program, pairing AI spending with estimates of assisted work. The dashboard expands its existing analytics as agent workloads add usage charges, but its dollar-value estimates do not establish financial returns.
Adobe’s Firefly update saves recurring visual subjects and creative preferences, extending an existing push into repeatable production. Memory has narrower reach than the assistant, while campaign demonstrations leave reliability and savings unmeasured.
Microsoft’s Ignite upgrades package enterprise knowledge, skills and tools into reusable business plugins. The expansion builds on IQ’s June release and competes with existing graph retrieval services; total savings and portability remain unproven.
The European Central Bank warns that debt and interlinked financing could carry an AI investment reversal into credit markets and the European economy. Resilient earnings, limited direct bank lending and stronger bank buffers temper its comparison with past bubbles.
Reported introductions helped Alice Huang secure Alibaba as a chip-rental customer and obtain financing. The relationship raises questions about Nvidia's overseas sales network, but does not itself establish illegal exports.
KPMG says AI agents pursued some issues more deeply in a parallel UK audit, though humans still had to correct their work. Client fee savings remain unconfirmed, even as KPMG International has negotiated lower fees from its own auditor by citing AI.
Payhawk announced a $280 million all-equity Series C led by Growth Equity at Goldman Sachs Alternatives, valuing it at $2 billion. The funding backs an expansion from managing business spending into broader AI finance operations, where competitors already combine automation, payments and controls.
A 16-country survey of 15- to 22-year-olds finds a median of 33% have used AI as a companion or friend. Users have higher adjusted odds of loneliness, but the research cannot establish whether AI helps, harms or replaces human connection.