Author：TechFlow
Author: Deep Tide TechFlow
Deep Tide Guide:According to IDC data, China's total shipments of AI accelerator cards will reach approximately 4 million units in 2025, with domestic manufacturers delivering a total of 1.65 million units, accounting for 41% of the market. Nvidia's market share has fallen from approximately 95% before the sanctions to 55%.
Huawei leads the domestic chip industry with 812,000 chips, and its newly released Atlas 350 accelerator card boasts inference performance 2.87 times that of NVIDIA H20.
Beijing's order last November for state-owned data centers to fully replace domestic products is accelerating the reshaping of the market landscape.
Three years ago, Nvidia virtually monopolized the Chinese AI chip market. Today, that landscape has changed drastically.
According to Reuters, citing data from market research firm IDC, total shipments of AI accelerator cards (dedicated computing chips for AI servers) in China are projected to reach approximately 4 million units in 2025. Nvidia remains the largest single supplier, shipping approximately 2.2 million units, accounting for 55% of the market share. However, this figure represents a significant decrease of nearly 40 percentage points compared to its pre-sanctions market share of approximately 95%. Meanwhile, domestic Chinese manufacturers will ship a total of approximately 1.65 million units, capturing 41% of the market. AMD will rank third with approximately 160,000 units shipped, accounting for 4%.
The rise of domestic manufacturers is both a passive product of US export controls and an active result of the "domestic substitution" policy.
Huawei leads the domestic camp, with the Atlas 350 competing with Nvidia's H2O.
Huawei is the biggest winner among domestically produced AI chips.
According to IDC data, Huawei shipped approximately 812,000 AI chips in 2025, accounting for about 20% of the entire market and nearly half of the total shipments from domestic manufacturers. Alibaba's chip design arm, T-Head, ranked second with approximately 265,000 chips, while Baidu's Kunlun Chip and Cambricon tied for third with approximately 116,000 chips each. Hygon, MetaX, and Iluvatar CoreX accounted for 5%, 4%, and 3% of the total shipments from domestic manufacturers, respectively.
At its China Partner Conference 2026 in Shenzhen last month, Huawei launched its new generation AI accelerator card, Atlas 350, powered by its self-developed Ascend 950PR chip. Zhang Dixuan, head of Huawei's Ascend computing business, stated at the launch event that the Atlas 350 achieves a computing power of 1.56 PFLOPS (petaflops per second) in FP4 low-precision computing, which is 2.87 times the performance of NVIDIA's China-specific H20. The card is equipped with 112GB of self-developed high-bandwidth memory HiBL 1.0, with a memory bandwidth of 1.4TB/s and a power consumption of 600W.
However, this performance comparison has a caliber issue. NVIDIA's Hopper architecture GPUs do not natively support FP4 precision, while the Atlas 350 is the first domestically produced accelerator card optimized for FP4. Therefore, the two cannot be directly compared at the same precision. Huawei's real competitive advantage lies in the inference side: the Atlas 350 is positioned for inference workloads during the AI model deployment phase, rather than large model training.
Seven Huawei partners have released complete server products based on the Atlas 350, and iFlytek has also announced that its next-generation Spark large model will be compatible with the Ascend 910/950 computing power base.
Driven by both export controls and domestic substitution orders
Nvidia's shrinking market share in China is the result of the dual pressure from escalating US export controls and Beijing's domestic substitution policies.
The timeline is roughly as follows: The US restricted AI chip exports to China starting in October 2022. Nvidia subsequently launched compliant, downgraded versions of the H20 and A800/H800 products. In April 2025, the Trump administration imposed a complete ban on all AI GPU exports to China; in July of the same year, export licenses for the H20 and AMD MI308 were reinstated; in October, Nvidia CEO Jensen Huang stated publicly that Nvidia's market share in China's advanced AI accelerator card market "fell from 95% to zero." In December, Trump allowed Nvidia to export the H200 to China, but Chinese companies were told to suspend orders for Nvidia chips.
The policy push on the other side is equally strong. According to a Reuters report in November 2025, Beijing issued guidelines to newly built data centers using state-owned assets, requiring them to use only domestically produced AI chips. Projects less than 30% complete were required to remove installed foreign chips or cancel their procurement plans.
According to Reuters statistics, since 2021, China's AI data center projects have received more than $100 billion in state-owned investment, and most of China's data centers have received some form of state-owned support during the construction phase, which means that this policy has a very wide coverage.
China Unicom's large-scale data center in Qinghai has been described by Reuters as a landmark example of this strategy: the project, valued at $390 million, is powered entirely by domestically produced AI chips such as Pingtouge.
The technological gap does exist, but the inference side has reached the "sufficient" threshold.
The rise in the market share of domestically produced chips does not mean that the technological gap has been eliminated.
Most industry analysts estimate that China's domestically produced AI chips still lag behind NVIDIA by 5 to 10 years in data center training. NVIDIA's high-end GPUs remain the preferred choice when training large language models (LLMs) with trillions of parameters. The 50,000 Hopper series GPUs used by DeepSeek to train its R1 model are a real-world example.
However, the situation is different on the inference side. Industry observers believe that for 90% of commercial application scenarios (including image recognition, chatbots, autonomous driving, etc.), domestically produced chips have reached the "good enough" threshold, making switching from Nvidia to domestic solutions a viable business decision. The anticipated further sanctions have accelerated this shift.
The real bottleneck lies in the software ecosystem. NVIDIA's CUDA platform, after more than a decade of development, has become the de facto standard for AI development. Domestic chip manufacturers are investing heavily in compatibility: Muxi announced that its C500 series will support CUDA compatibility, Huawei will fully open-source its CANN platform in 2025 to expand its developer ecosystem, and Cambricon and Moore's Threads have each built translation tools from CUDA to their own programming languages. The pace of ecosystem development will determine the ceiling of the domestic chip market share.
Domestic AI chip companies are rushing to the capital market.
The shift in market share is being realized simultaneously in the capital market.
Since the beginning of 2026, China's GPU industry has witnessed a wave of IPOs. Biren Technology and Muxi have already listed on the STAR Market, Tianshu Zhixin is listed on the Hong Kong Stock Exchange's main board, and Suiyuan Technology's STAR Market listing application has been accepted. Baidu announced its plan to spin off Kunlun Chip for an independent listing, and according to sources, Alibaba is also considering a similar spin-off of Pingtouge.
Huawei's R&D investment reached 192.3 billion yuan in 2025, accounting for 22% of its revenue, with a focus on chips, software, and manufacturing tools to further reduce its dependence on US technology. Huawei's rotating chairman, Xu Zhijun, stated at MWC 2026 that Huawei will become "an alternative that ensures uninterrupted global AI computing power supply." According to Reuters, Huawei's new-generation Ascend 950PR chip has attracted order interest from giants such as ByteDance and Alibaba, with a shipment target of approximately 750,000 units in 2026, and mass production scheduled to begin in the second half of the year.
For Nvidia, even with the H200 approved for export to China, the foundation of trust has been shaken. Beijing's policy of self-reliance is no longer just a vision, but a fait accompli embodied in every domestically produced chip running in data centers. Whether the 55% market share figure will rebound or continue to decline when the 2026 market share data is released will depend on whether Washington's export policy will shift again, and how quickly domestically produced chips catch up in the training sector.
