Nvidia Chip Restrictions Expose China’s Advanced AI Manufacturing Weakness

Training advanced AI models and operating increasingly complex AI systems require enormous computing resources. Nvidia’s H100, H200 and Blackwell processors power data centres used by leading American technology companies, including Microsoft, Meta, Amazon and Google.

Their AI infrastructure investments run into hundreds of billions of dollars, covering processors, networking equipment, data centres and electricity supply. Demand for Nvidia’s hardware has grown as developers expand model training and deploy AI systems that require more computing power to operate.

On November 12, Nvidia announced the resignation of co-founder and chief executive Jensen Huang, with Chief Financial Officer Colette Kress taking over temporarily.

His departure followed revelations of an affair with Alice Huang, a former executive at Singapore-based Megaspeed whom US intelligence officials reportedly suspected of working for China.

Megaspeed has faced allegations involving the diversion of restricted Nvidia processors to Chinese customers. Singaporean authorities investigated Alice Huang but reported insufficient evidence to establish criminal offences. Nvidia has said it has no reason to believe its technology or national security was compromised.

The controversy has renewed scrutiny of Washington’s restrictions on advanced AI chip exports to China. The controls target processors used in large computing clusters, although certain Nvidia products have remained eligible for export under licensing arrangements.

Chinese technology companies have developed domestic alternatives, including Huawei’s Ascend processors, but American-designed hardware continues to attract demand despite the restrictions.

China’s dependence on these processors raises a manufacturing question beyond the investigation surrounding Nvidia’s former chief executive.

Huawei and other Chinese chipmakers have developed competing products, yet producing advanced AI hardware in sufficient quantities remains difficult. The investigation is whether China’s semiconductor industry can close that production gap before American developers extend their advantage in the next generation of AI computing.

China’s Chip Production Falls Behind Its AI Ambitions

Huawei has developed alternatives to Nvidia’s AI processors, but domestic production has not kept pace with demand. At its September 2026 technology conference, the company acknowledged that orders for its AI computing equipment exceeded available supply.

Huawei reported delivering more than 1,000 large computing systems to over 370 customers, confirming commercial adoption despite its manufacturing limitations.

The Ascend 910C is already used in Huawei’s Atlas 900 A3 SuperPoD, which connects 384 processors. Its newer Ascend 950 series is designed for AI training and inference. Huawei lists 144 gigabytes of high-bandwidth memory and memory bandwidth of 4 terabytes per second for the 950DT.

ByteDance and Alibaba were reportedly preparing orders for the 950PR in March, following testing. The reported orders do not establish how many processors have entered commercial service.

China’s Semiconductor Manufacturing International Corporation (SMIC) has produced advanced chips using deep-ultraviolet lithography. However, restrictions on extreme-ultraviolet equipment prevent Chinese foundries from using the manufacturing technology employed in leading-edge semiconductor production.

SMIC can produce smaller chip features through repeated patterning, although the additional manufacturing steps increase complexity and production costs. Comparable public yield figures are insufficient to determine the precise effect on AI processor output.

High-bandwidth memory is another production difficulty. Huawei has announced its own memory technology for the Ascend 950 generation, but access to sufficient advanced memory remains uncertain.

Reuters reported in September that Huawei raised the indicated price of its Ascend 950DT accelerator card to more than 250,000 yuan, approximately $37,255.

The reported increase of 20% to 50% over earlier quotations was partly attributed to higher memory costs. Chinese chipmaker Cambricon also increased indicative prices for its next-generation processors.

Huawei’s SuperPoD architecture connects large numbers of Ascend processors to provide computing capacity for demanding AI workloads. Manufacturing these systems requires advanced chip packaging, memory integration and reliable high-speed connections.

Huawei has demonstrated working large-scale systems, but publicly available disclosures do not establish how much packaging capacity Chinese suppliers can provide or whether it is sufficient for the company’s planned expansion.

The company has brought forward its processor roadmap. The Ascend 960DT is scheduled for the first quarter of 2027, followed by the 960PR in the third quarter.

Huawei plans to introduce Ascend 970 in 2028 and Ascend 980 in 2029, with improvements in processing performance, memory and interconnect bandwidth. These are announced development targets, not confirmed production volumes.

Huawei’s commercial deployments establish that China can design and operate advanced AI computing systems. The remaining uncertainty concerns manufacturing output.

Restrictions on semiconductor equipment, expensive memory and limited disclosed production capacity make it difficult to determine when Chinese suppliers will be able to satisfy domestic demand without relying on foreign accelerators.

Nvidia’s AI Computing Advantage Extends Beyond Chip Design

Nvidia’s H100, H200 and Blackwell B200 processors have increased the computing capacity available to developers training and operating advanced AI models. The H100 provides 80 gigabytes of high-bandwidth memory, compared with 141 gigabytes in the H200 and 180 gigabytes in the B200.

Memory bandwidth reaches approximately 3.35 terabytes per second in the H100, 4.8 terabytes in the H200 and 8 terabytes in the B200. The higher bandwidth is particularly important for moving data during large AI training and inference workloads.

Nvidia’s HGX B200 platform combines eight Blackwell processors and delivers up to 144 petaflops of stated peak computing performance under specified precision settings.

Its fifth-generation NVLink technology provides up to 1.8 terabytes per second of communication bandwidth per GPU. These figures describe hardware capabilities rather than guaranteed performance improvements across every AI model.

Meta’s Prometheus AI supercluster in Ohio illustrates the scale of commercial deployment. In October 2026, the company reported completing the facility’s first phase, with plans to expand toward approximately one million GPUs.

Meta is deploying Nvidia’s GB200 and GB300 systems, including configurations connecting 72 Blackwell-generation processors through NVLink. Its GB300 racks provide more than one exaflop of stated peak FP4 computing performance.

Nvidia’s CUDA software supports the development and operation of AI workloads across its processors. The company also supplies networking equipment and software libraries used in large computing clusters.

AMD produces competing AI accelerators, Google deploys its own tensor processing units, and Huawei has developed Ascend-based systems. Replacing Nvidia hardware can require changes to software, networking and workload optimisation, particularly in facilities already designed around its computing platforms.

US export controls have restricted sales of advanced Nvidia processors to China, including the H100 and B200. The Commerce Department revised its policy for the H200 in January 2026, allowing case-by-case licence reviews subject to security requirements and safeguards for American customers.

Commerce Under Secretary Jeffrey Kessler confirmed in July that a small number of H200 processors had reached China under the revised arrangements.

Restricted Nvidia hardware has also appeared in reported diversion investigations. A July 2025 Financial Times investigation identified at least $1 billion worth of advanced Nvidia processors allegedly smuggled into China over three months, including Blackwell B200 chips.

Trade records subsequently examined by Reuters showed that a subsidiary of Singapore-based Megaspeed imported H100 processors into Indonesia. The records establish the shipments to Indonesia but do not independently confirm that those particular processors were illegally transferred to China.

Nvidia’s fiscal 2026 annual report identifies Taiwan Semiconductor Manufacturing Company and Samsung among its semiconductor foundries.

SK Hynix, Micron and Samsung supply memory, while advanced packaging and assembly involve additional contractors, including Hon Hai and Wistron. Nvidia controls the design of its processors and computing platforms but depends on suppliers outside the United States for much of their production.

Washington’s export restrictions extend to certain foreign-produced semiconductor products containing or developed using controlled American technology.

The January 2026 H200 licensing framework also requires exporters to demonstrate that eligible Chinese sales will not reduce manufacturing capacity available to US customers. American authorities can therefore restrict access to Nvidia’s advanced processors even when those chips are manufactured overseas.

Nvidia’s advantage comes from supplying processors, software and networking systems already deployed across major AI infrastructure projects.

US export controls give Washington additional influence over access to that technology, although the manufacturing network remains dependent on companies operating across several countries.

China’s Computing Constraints Threaten Its Next AI Expansion

The United States received an estimated 72% of global AI computing capacity shipped between 2022 and June 2026, compared with 10% for China, according to research published by Epoch AI on October 27.

The study estimated approximately 19.8 million Nvidia H100-equivalents for the United States and 2.6 million for China, using standardised computing performance rather than physical processor counts.

China’s estimated share of AI chips was higher, at approximately 18%, indicating that the hardware supplied to the country delivered less computing capacity per chip on average.

Epoch AI placed China’s computing capacity between 1.9 million and 4.2 million H100-equivalents under alternative assumptions. Its estimates include domestic accelerators and foreign processors, including hardware obtained through permitted exports and estimated unauthorised shipments.

The research does not establish how much computing capacity Chinese developers can access at any given time or how effectively individual companies use their processors.

DeepSeek has pursued more efficient model development while adapting its software to Chinese hardware. Huawei reported that its Ascend processors supported part of the training of DeepSeek’s smaller V4-Flash model.

The companies have also worked on compatibility with domestic accelerators, including products from Cambricon. These developments demonstrate progress in using Chinese processors, although they do not establish that DeepSeek’s largest models can be trained entirely on domestic hardware.

The scale of future training requirements remains substantial. In September, The Information reported that DeepSeek expected Huawei training processors during the fourth quarter.

Chief Executive Liang Wenfeng had estimated that training a model comparable with the largest American systems would require approximately 50,000 Nvidia GB300 processors or 200,000 Huawei Ascend 950 chips. Those figures were estimates rather than results from comparable completed training runs.

Alibaba is developing its own computing equipment. At its September 2026 Apsara Conference, the company introduced the Zhenwu V900 accelerator and described a computing architecture designed to accommodate as many as 500,000 cards. Alibaba also disclosed continued development of its Qwen models.

The announced architecture establishes its intended computing scale, but not the number of accelerators already installed or available for commercial workloads.

AI agents are creating additional demand for inference capacity. An April 2026 Stanford Digital Economy Lab study found that agentic coding tasks could consume roughly 1,000 times as many tokens as simpler coding interactions in the workloads examined, without consistently delivering better results.

Microsoft researchers separately found that longer reasoning and agentic queries could increase inference energy consumption by more than an order of magnitude compared with standard requests.

These workloads place different demands on computing infrastructure. Training a frontier model requires large concentrations of processing capacity, while agents performing software engineering, research and other extended tasks can generate substantial computing demand after deployment. Their operating costs depend on how many tasks they complete, how long they run and the efficiency of the underlying models.

Epoch AI estimated in October that high-bandwidth memory shipped through 2027 could support approximately 30 million to 170 million concurrent frontier-model agents under different assumptions.

The projection describes potential capacity rather than expected deployment. It nevertheless indicates the scale of hardware that could be required if developers expand the commercial use of autonomous systems.

American technology companies are committing substantial capital to computing infrastructure. Meta projected capital expenditure of $135 billion to $145 billion for 2026, while Amazon’s latest outlook was approximately $220 billion.

These budgets include infrastructure beyond AI accelerators and cannot be compared directly with Alibaba’s proposed cluster capacity. They establish the financial resources available to two major American operators as Chinese developers seek to expand domestic computing supply.

China’s developers have demonstrated that model efficiency and software adaptation can reduce dependence on the most advanced imported processors. DeepSeek’s training methods and Alibaba’s accelerator programme provide evidence of that approach.

The larger constraint is deployment capacity: insufficient advanced computing resources could limit how frequently Chinese companies train new models and how widely they operate inference-intensive services, even when their models achieve competitive results.

Advanced AI Computing Will Test China’s Technological Independence

Huawei’s planned Ascend 960 release in 2027 and Ascend 970 introduction in 2028 will test China’s ability to expand domestic AI computing capacity. The important figures will be commercial production volumes, high-bandwidth memory availability and deliveries to major Chinese technology companies.

Adoption in large training clusters would provide evidence that domestic processors are replacing some of the computing capacity previously supplied by Nvidia. Announced performance improvements alone cannot establish that shift.

AI developers are increasingly deploying systems that carry out extended software engineering, research and operational tasks. These applications can require substantially more inference computing than conventional chatbot interactions, particularly when agents repeatedly generate code, use external tools and evaluate their results.

Automated scientific research and more capable autonomous systems could increase those requirements further. Their contribution to artificial general intelligence remains uncertain, with no established computing threshold guaranteeing its development.

The effectiveness of American export restrictions will depend partly on how quickly Chinese suppliers expand production and how successfully Washington limits access to controlled processors.

DeepSeek, Alibaba and other Chinese developers can continue improving model efficiency, but those advances do not eliminate the need for computing infrastructure capable of supporting commercial deployment.

China does not need to manufacture identical replacements for every Nvidia processor. Its competitive position will depend on whether domestic suppliers can deliver advanced computing systems at the performance, volume and cost required by its AI industry.

Until they can, restrictions on foreign accelerators will continue to constrain China’s ability to develop and deploy increasingly demanding AI systems.