What are China’s Nvidia Alternatives After New U.S. AI Chip Restrictions?

What are China's Nvidia Alternatives After New U.S. AI Chip Restrictions?
Share

China’s biggest technology companies are turning to domestic AI chips as new U.S. restrictions make Nvidia’s advanced processors harder to access. Huawei, Cambricon, Biren and several smaller chipmakers are competing to fill the gap, but replacing Nvidia for training powerful AI models remains a major challenge.

China’s efforts to replace Nvidia with domestically developed AI processors are entering a new phase as Washington tightens restrictions on the computing resources available to Chinese technology companies.

On November 20, 2026, Reuters reported that new U.S. restrictions were affecting companies including Alibaba, Tencent, MiniMax and Z.AI. The measures extend beyond physical chip shipments by restricting access to advanced computing resources through certain overseas data centers.

The development has increased pressure on Chinese AI developers to reduce their dependence on American semiconductor technology. Companies that previously relied on Nvidia accelerators or rented computing capacity abroad must now consider how to build and operate more of their AI infrastructure using domestic equipment.

Tencent has indicated that its existing inference workloads are relatively well protected. However, the company acknowledged that restrictions could slow the development of its next generation of AI models.

That distinction explains much of China’s current challenge. Domestic chips are becoming increasingly useful for running existing AI models, but training the largest and most advanced systems requires substantially more computing power, memory bandwidth and software support.

Huawei, Cambricon, Biren, Moore Threads, MetaX and other Chinese semiconductor companies are competing to meet that demand. Some have established commercial products, while others are still scaling production or improving compatibility with widely used AI frameworks.

The question is no longer whether China has alternatives to Nvidia. It is whether those alternatives can provide enough computing capacity, at a competitive cost, to support the country’s growing AI industry.

Why China Is Accelerating Its Shift Away From Nvidia

Nvidia has dominated the market for processors used to train and operate advanced AI systems.

Its GPUs are widely used by technology companies because they combine high computing performance with a mature software ecosystem, including CUDA, the programming platform used by many AI developers.

For Chinese companies, access to Nvidia’s most advanced products has become increasingly uncertain because of U.S. export restrictions.

Washington has imposed controls on certain high-performance semiconductors and related technologies, citing national security concerns about advanced computing capabilities.

The latest restrictions increase the pressure by addressing access to computing infrastructure outside China, not simply the direct shipment of physical chips.

According to Reuters’ November 20 report on U.S. AI-chip restrictions, Chinese technology companies are responding by accelerating efforts to develop and adopt domestic computing systems.

This creates a commercial opportunity for local chipmakers, but it also introduces technical challenges.

Switching from Nvidia to another supplier can require changes to model code, software libraries, infrastructure management and the way computing resources are distributed across servers.

A domestic chip may be capable of running an AI model without delivering the same speed, efficiency or ease of deployment as an Nvidia-based system.

That makes the transition more complicated than simply replacing one processor with another.

ALSO READ: Chinese AI Companies Rush to Secure Nvidia Chips as U.S. Restrictions Fears Grow

1. Huawei Ascend: China’s Most Established Nvidia Alternative

Huawei has emerged as one of China’s most important suppliers of domestic AI accelerators.

Its Ascend family includes processors such as the Ascend 910C, 950PR and 950DT, alongside server systems designed to combine multiple chips into larger computing clusters.

Huawei also develops the CANN software platform, which provides tools for running AI workloads on Ascend hardware.

This gives Huawei an advantage over smaller competitors because it offers a broader combination of chips, servers, software and technical support.

The Ascend 910C has become an important product for Chinese companies seeking alternatives to Nvidia’s earlier data center accelerators.

Huawei has also introduced newer products intended to improve performance and compatibility with existing AI development tools.

In March, Reuters reported that ByteDance and Alibaba were considering orders for Huawei’s Ascend 950PR.

Sources familiar with the plans said the newer chip offered improvements in response speed and compatibility with software commonly used in Nvidia-based environments.

The development matters because Chinese companies have historically faced difficulties moving large AI workloads from Nvidia’s CUDA platform to alternative hardware.

However, Huawei still faces manufacturing constraints and performance limitations compared with Nvidia’s latest flagship accelerators.

A November report from the U.S.-China Economic and Security Review Commission compared the Ascend 910C with Nvidia’s B200 and found substantial differences in processing capability and memory bandwidth.

Huawei’s advantage is therefore not necessarily having the fastest individual chip. It is offering a domestic computing platform that Chinese companies can obtain and develop around with less dependence on U.S. suppliers.

2. Cambricon: A Major Chinese AI Accelerator Competitor

Cambricon Technologies is another important company in China’s domestic AI semiconductor industry.

The Beijing-based chipmaker develops processors for AI training and inference, including its Siyuan family of accelerators.

Its Siyuan 590 has been one of the company’s main commercial products, while the newer 690 has been under development.

Cambricon competes with Huawei in supplying computing hardware to Chinese cloud providers, technology companies and other organizations.

The company’s products use a separate software environment, which means developers may need to adapt workloads that were originally created for Nvidia hardware.

Cambricon has benefited from growing demand for domestic AI processors as export restrictions make foreign alternatives harder to obtain.

However, its expansion has also been affected by the rising cost and limited availability of advanced memory components.

In September, Reuters reported that Cambricon had increased its indicated pricing for its next-generation 690 accelerator by approximately 20% to 30%.

At that time, the company had not formally launched the 690, and the Siyuan 590 remained its flagship commercial product.

The price increases demonstrate that domestic alternatives are not automatically cheaper than imported accelerators.

China’s chipmakers face many of the same cost pressures as their international competitors, including expensive memory, manufacturing capacity and the need to invest heavily in research.

Cambricon’s longer-term opportunity depends on improving performance while expanding production and strengthening its software ecosystem.

3. Biren Technology: A Chinese GPU Developer Challenging Nvidia

Biren Technology is another Chinese semiconductor company developing processors for high-performance computing and AI workloads.

The company attracted attention with its BR100 architecture, which was introduced as a high-performance accelerator for data centers.

Biren’s approach differs from companies that focus primarily on specialized AI processors because it has pursued GPU-style computing hardware intended to support a wider range of parallel processing tasks.

That makes it a potential alternative for workloads that require large amounts of mathematical computation.

However, Biren has faced major obstacles in bringing its technology to market at scale.

U.S. export restrictions have affected access to advanced semiconductor manufacturing and related technologies, limiting the company’s ability to produce some of its most ambitious designs.

The company has continued developing its hardware and pursuing opportunities in China’s domestic computing market.

Biren’s processors were also included in China’s 2026 security certification framework for certain government technology purchases.

That recognition could support adoption in government-related and regulated environments.

However, certification does not establish that Biren’s products match Nvidia’s newest GPUs in performance or that they are available in comparable quantities.

For customers, the practical questions remain whether the required hardware can be obtained, which AI frameworks it supports and how much engineering work is necessary to move existing applications.

4. Moore Threads: Building a Chinese GPU and Software Ecosystem

Moore Threads is another company seeking to develop domestic GPU technology.

Its products include the MTT accelerator family, with hardware intended for graphics, general-purpose computing and AI workloads.

The company has also developed MUSA, its own GPU computing software platform.

MUSA is important because the software used to operate AI accelerators can be as valuable as the hardware itself.

Developers working with Nvidia products often rely on CUDA libraries, tools and optimized code.

Moving those workloads to a different GPU platform can require substantial software changes.

Moore Threads has therefore focused on improving compatibility with existing computing tools and making it easier for developers to adopt its hardware.

Its MTT S5000 has been positioned for AI computing applications, including inference and other demanding workloads.

The company was among the domestic suppliers included in China’s 2026 AI processor certification framework.

However, Moore Threads still faces the challenge of proving that its products can deliver reliable performance across a wide range of real-world applications.

The ability to run a model successfully is only one part of the equation. Businesses also need predictable performance, stable software and technical support over time.

5. MetaX: Another Domestic AI GPU Supplier

5. MetaX: Another Domestic AI GPU Supplier

MetaX Integrated Circuits is developing GPUs for AI and other high-performance computing applications.

Its C-series processors are intended to serve data center workloads, giving Chinese companies another possible source of computing capacity.

MetaX has attracted attention as demand for domestically produced accelerators increases.

The company was also included in China’s 2026 certification framework for AI processors used in government procurement.

However, MetaX operates in a market where established software support and access to manufacturing capacity can determine whether a processor becomes commercially successful.

Developers need to know that models can run reliably on the hardware and that required software tools will remain supported.

MetaX also faces competition from Huawei, Cambricon and other domestic companies pursuing similar customers.

Its opportunity is likely to depend on offering a useful combination of performance, availability and software compatibility rather than competing only on headline processor specifications.

6. Baidu Kunlunxin: AI Chips Developed for Cloud Computing

Baidu has also invested in developing its own AI processors through Kunlunxin.

The company has worked on accelerators intended for AI training, inference and cloud computing.

Its Kunlun family includes processors used in Baidu’s computing infrastructure and related services.

Unlike some independent chipmakers, Baidu has an established AI and cloud business that can provide demand for its hardware.

This creates an opportunity to develop chips around specific workloads and improve their performance through real-world deployment.

For example, a cloud provider may optimize its processors for the models and services it operates most frequently.

However, hardware developed for one company’s infrastructure may not be as easy for outside customers to adopt.

Businesses considering Kunlunxin processors must evaluate software support, availability and compatibility with their existing systems.

Baidu’s investment demonstrates that China’s push for domestic AI chips extends beyond semiconductor specialists.

Major technology companies are also developing their own processors to reduce dependence on outside suppliers.

7. Alibaba T-Head, Hygon and Iluvatar CoreX

China’s domestic AI chip industry includes several additional companies pursuing different approaches.

Alibaba’s T-Head develops processors for cloud computing and AI workloads, including its Zhenwu family.

Like Baidu, Alibaba has a large cloud business that can support the development and deployment of specialized hardware.

Hygon Information Technology develops data center processors and DCU accelerators intended for high-performance computing and AI applications.

Its products provide another option for organizations that want computing hardware from a Chinese supplier.

Iluvatar CoreX has also expanded its position in the domestic AI accelerator market.

According to Reuters’ September investigation into Chinese AI chip supply, Iluvatar CoreX increased shipments to ByteDance as the technology company sought additional computing capacity.

The report said Huawei was ByteDance’s largest domestic AI chip supplier, followed by Cambricon and Iluvatar CoreX.

These developments show that Chinese technology companies are not depending on a single Nvidia replacement.

Instead, they are building relationships with several domestic suppliers, each offering different hardware and software capabilities.

Huawei vs Cambricon vs Biren: Which Chinese AI Chip Is Best?

There is no single Chinese processor that can be described as the best Nvidia replacement for every workload.

Huawei has one of the most complete domestic computing platforms, combining Ascend accelerators with servers and the CANN software environment.

Cambricon offers specialized AI accelerators and has gained commercial traction among customers seeking alternatives to Nvidia.

Biren focuses on high-performance GPU computing, while Moore Threads and MetaX are developing their own GPU hardware and software ecosystems.

Baidu and Alibaba have additional advantages because they can use their own cloud infrastructure to support chip development.

For customers, the choice depends on several factors.

A company running an existing language model may prioritize inference speed, memory capacity and cost per generated token.

A research organization training a new model may place greater importance on high-speed communication between chips, memory bandwidth and the ability to operate thousands of processors together.

Software compatibility also matters because moving workloads between platforms can require substantial engineering work.

As a result, comparing theoretical chip performance alone does not provide a complete picture of real-world suitability.

ALSO READ: China’s AI Industry: Key Statistics and Data (2025–2026)

Chinese AI Chips Compared: Main Products and Strengths

The following comparison summarizes the major domestic suppliers and their positioning in China’s AI computing market.

CompanyMain AI chip familyMain focusKey challenge
HuaweiAscend 910C, 950PR, 950DTTraining, inference and large computing systemsManufacturing scale and performance gap
CambriconSiyuan 590, planned 690Dedicated AI acceleratorsMemory supply, pricing and software adoption
BirenBR-series GPUsHigh-performance parallel computingManufacturing restrictions and production scale
Moore ThreadsMTT seriesGPUs, AI computing and software compatibilitySoftware maturity and commercial deployment
MetaXC-series GPUsData center AI computingScaling adoption and ecosystem support
Baidu KunlunxinKunlun familyCloud AI training and inferenceBroader third-party adoption
Alibaba T-HeadZhenwu familyCloud computing and specialized AI workloadsExternal availability and ecosystem development
HygonDCU familyHigh-performance computing and AISoftware compatibility and efficiency
Iluvatar CoreXBI-series acceleratorsAI training and inferenceProduction capacity and customer expansion

These products are not all at the same stage of commercial availability. Some are established offerings, while others are newer designs, planned releases or products with limited public performance data.

Can Chinese AI Chips Match Nvidia’s Performance?

Chinese companies have made progress in developing AI accelerators, but Nvidia retains important advantages at the high end of the market.

A November 2026 report from the U.S.-China Economic and Security Review Commission compared Huawei’s Ascend 910C with Nvidia’s B200.

The report listed a total processing performance measure of 12,800 for the Ascend 910C and 40,000 for the Nvidia B200.

It also reported memory bandwidth of approximately 3.2 terabytes per second for Huawei’s chip, compared with 8 terabytes per second for Nvidia’s processor.

These figures suggest a substantial hardware performance gap, although specifications alone do not establish how the products perform in every application.

The difference can be particularly important for training large AI models, where processors must exchange large amounts of information.

However, Chinese companies can partly compensate for weaker individual processors by combining more chips into larger systems.

Huawei has pursued this strategy through its CloudMatrix computing architecture.

The approach can increase total computing capacity, but it also creates challenges involving networking, electricity consumption, cooling and system efficiency.

That means a large cluster of domestic processors may be capable of completing demanding tasks without necessarily matching Nvidia’s systems in cost or performance per watt.

Why Tencent Says Inference Is Safer Than AI Model Training

Tencent’s comments following the latest U.S. restrictions highlight an important distinction between two types of AI computing.

Training involves developing or improving a model by processing large amounts of data and adjusting its internal parameters.

For advanced language models, this can require thousands of accelerators working together over extended periods.

Inference is the process of using a trained model to generate responses, classify information or perform other tasks.

Inference workloads can often be distributed across different systems and optimized for particular hardware.

This makes it possible for companies to use a broader range of processors, especially when running models that have already been developed.

Tencent has indicated that its existing inference operations are relatively protected from the latest restrictions.

However, the company acknowledged that access constraints could delay training its next generation of models.

The distinction matters because China’s domestic semiconductor industry may be able to support many existing AI services before it can fully replace the advanced computing infrastructure needed to train the largest new models.

It also explains why companies may continue investing in domestic accelerators even when those products do not match Nvidia’s most advanced hardware.

China’s Biggest AI Chip Problem May Be Memory, Not Processing Power

One of the most serious challenges facing Chinese chipmakers is access to high-bandwidth memory, commonly known as HBM.

HBM allows AI accelerators to move large amounts of information between memory and processing units.

This is essential for many modern AI workloads, particularly large language models that require rapid access to model parameters and intermediate calculations.

Manufacturers including SK Hynix, Samsung and Micron have played major roles in the global HBM market.

Chinese accelerator developers face supply constraints and higher costs when sourcing advanced memory components.

In September, Reuters reported that Huawei and Cambricon had increased prices for several AI products as HBM costs rose.

Huawei’s Ascend 950DT accelerator card was reportedly quoted above 250,000 yuan, equivalent to approximately $37,255 at the exchange rate cited in the report.

The company had previously said the 950DT would become available in the fourth quarter of 2026.

Higher memory costs can make domestic accelerators more expensive, reducing the economic benefit of switching away from Nvidia.

They can also limit production volumes, making it difficult for Chinese suppliers to satisfy rapidly growing demand.

This shows why semiconductor independence requires more than developing a competitive processor design.

China must also strengthen its memory supply, advanced packaging, manufacturing equipment and supporting technologies.

Why Nvidia’s CUDA Software Is Difficult to Replace

Why Nvidia's CUDA Software Is Difficult to Replace

Nvidia’s advantage extends beyond the physical performance of its GPUs.

CUDA has become a widely used platform for AI development, supported by software libraries, frameworks and optimization tools.

Developers often build and test AI systems using software that is optimized for Nvidia hardware.

Moving those applications to another accelerator can require rewriting parts of the code or replacing specialized operations.

Huawei has developed CANN as its alternative computing platform.

Moore Threads uses MUSA, while other Chinese chipmakers maintain their own software tools and development environments.

These systems can support AI workloads, but they may not offer the same level of compatibility, documentation or optimization as Nvidia’s established ecosystem.

According to research from the Mercator Institute for China Studies, software compatibility has been one of the main obstacles facing China’s domestic AI accelerator industry.

The report identified challenges involving CUDA compatibility, reliability and the effort required to move workloads between platforms.

Chinese companies are working to reduce those barriers, including by improving support for widely used machine learning frameworks.

However, building a competitive software ecosystem can take years because developers need reliable tools, documentation and technical support.

Will China Completely Replace Nvidia by 2027?

China is making progress in reducing its dependence on Nvidia, but a complete replacement within a specific year cannot be established from current evidence.

Domestic accelerators are becoming more common in government-related procurement and among major technology companies.

Huawei, Cambricon and other suppliers are also expanding production plans to meet demand.

However, the industry still faces limitations involving advanced manufacturing, memory supply, software compatibility and the performance of large computing clusters.

Some Chinese companies may be able to operate most of their inference workloads using domestic hardware while continuing to face difficulties training their most advanced models.

Others may use a combination of domestic processors and existing foreign equipment for several years.

The pace of the transition will also depend on future export controls, semiconductor manufacturing progress and the availability of key components.

Rather than expecting Nvidia to disappear from China’s AI industry overnight, the more likely near-term development is a gradual shift toward a broader domestic supplier base.

The speed and extent of that shift remain uncertain.

What China’s Nvidia Alternatives Mean for the Global AI Industry

China’s investment in domestic AI accelerators could have consequences beyond its own technology market.

If Huawei, Cambricon and other companies successfully expand production, they could become more important suppliers to businesses seeking alternatives to American semiconductor technology.

Greater competition could also encourage new approaches to AI model optimization, chip design and computing infrastructure.

However, the development of separate hardware and software ecosystems may create compatibility challenges.

Companies operating across different markets could find themselves maintaining multiple versions of their AI applications to support different processors.

The restrictions could also influence the strategies of major Chinese technology companies.

Alibaba, Tencent, ByteDance and Baidu may invest more heavily in domestic chip suppliers, custom processors and software designed to reduce dependence on Nvidia.

At the same time, Nvidia remains a major force in the global AI computing market, supported by its hardware performance, manufacturing partnerships and established software ecosystem.

China’s domestic chip expansion therefore represents growing competition rather than proof that Nvidia’s global leadership has ended.

The Real Test for China’s AI Chips Is Training the Next Generation of Models

The latest U.S. restrictions have increased the urgency of China’s efforts to build an independent AI computing industry.

Huawei, Cambricon, Biren, Moore Threads and other domestic suppliers are gaining attention as Chinese technology companies look for more reliable access to processors.

Some of these products can already support important AI inference workloads, while newer systems aim to improve training performance and compatibility with existing software.

However, the challenges remain substantial.

Advanced AI development requires large amounts of computing power, high-bandwidth memory, efficient networking and mature development tools.

China has made progress in each area, but domestic suppliers have not yet demonstrated that they can replace Nvidia’s entire hardware and software ecosystem at the highest level.

Tencent’s warning about potential delays to next-generation model training illustrates the remaining gap.

China has alternatives to Nvidia, and those alternatives are becoming more important. The unanswered question is whether they can scale quickly enough to keep China’s most advanced AI developers competitive as access to foreign computing becomes more restricted.