Author: Shen潮 TechFlow

Shen潮 Introduction: IDC data shows that by 2025, the total shipment volume of AI acceleration cards in China is expected to reach approximately 4 million units, with domestic manufacturers delivering a total of 1.65 million units, accounting for 41%, while Nvidia's share has dropped from about 95% before sanctions to 55%.

Huawei leads the domestic camp with 812,000 chips, while its newly released Atlas 350 acceleration card claims to achieve 2.87 times the inference performance of Nvidia's H20.

Last November, Beijing ordered state-owned data centers to fully replace foreign products with domestic alternatives, accelerating the reshaping of the market landscape.

Three years ago, Nvidia almost monopolized the Chinese AI chip market. Today, that landscape has changed dramatically.

According to Reuters citing data from market research agency IDC, the total shipment volume of AI acceleration cards (dedicated computing chips for AI servers) in China is expected to reach approximately 4 million units by 2025. Nvidia remains the largest single supplier, shipping about 2.2 million units, accounting for 55% of the market share. However, this figure has significantly shrunk by nearly 40 percentage points compared to the roughly 95% market share before sanctions. Meanwhile, domestic manufacturers collectively delivered about 1.65 million units, capturing 41% of the market. AMD ranks third with approximately 160,000 units shipped, accounting for 4%.

The rise of domestic manufacturers is both a passive result of U.S. export controls and an active outcome of the "domestic replacement" policy.

Huawei Leads the Domestic Camp, Atlas 350 Competes with Nvidia H20

Within the domestic AI chip camp, Huawei is the biggest winner.

IDC data shows that Huawei is expected to ship about 812,000 AI chips in 2025, accounting for about 20% of the entire market and nearly half of the shipments from domestic manufacturers. Alibaba's chip design arm, T-Head, ranks second with about 265,000 units, while Baidu's Kunlun and Cambricon each shipped about 116,000 units, tying for third. Additionally, Hygon, MetaX, and Iluvatar CoreX account for 5%, 4%, and 3% of the shipments from domestic manufacturers, respectively.

Last month, Huawei unveiled its new generation of AI acceleration card, the Atlas 350, at the 2026 China Partner Conference in Shenzhen, equipped with the self-developed Ascend 950PR chip. Zhang Dixuan, head of Huawei’s Ascend computing business, stated at the launch that the Atlas 350 achieves 1.56 PFLOPS (quadrillions of operations per second) in FP4 low-precision computing, with a performance that is 2.87 times that 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, there are discrepancies in this performance comparison. Nvidia's Hopper architecture GPUs do not natively support FP4 precision, while the Atlas 350 is the first domestically produced acceleration card optimized for FP4, meaning the two cannot be directly compared at the same precision. Huawei's true competitive edge lies in inference: the Atlas 350 is positioned for inference workloads during the AI model deployment phase, rather than large model training.

Seven of Huawei's partners have already released complete server products based on the Atlas 350, and iFlytek has also announced that its next-generation Spark large model will adapt to the Ascend 910/950 computing base.

Export Controls and Domestic Replacement as Dual Drivers

The collapse of Nvidia's share in China is the result of increasing U.S. export controls and the dual pressure of Beijing's domestic replacement policy.

The timeline is roughly as follows: Starting in October 2022, the U.S. imposed restrictions on AI chip exports to China, after which Nvidia launched compliant downscaled versions like the H20 and A800/H800. In April 2025, the Trump administration completely banned all exports of AI GPUs to China; in July of the same year, it restored export licenses for the H20 and AMD MI308; in October, Nvidia CEO Jensen Huang stated at a public event that Nvidia's share in the advanced AI acceleration card market in China "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 from the other side is equally strong. According to a report by Reuters in November 2025, Beijing issued guidelines to newly constructed state-owned data centers using state funds, requiring them to adopt only domestic AI chips. Projects with less than 30% completion were instructed to remove already installed foreign chips or cancel procurement plans.

Statistics from Reuters show that since 2021, Chinese AI data center projects have received over $100 billion in state funding, while most data centers in China have received some form of state support during their construction phase, implying that this policy has a wide coverage.

The large data center being built by China Unicom in Qinghai has been reported by Reuters as a landmark case of this strategy: the project, valued at $390 million, is powered entirely by domestic AI chips like T-Head.

Technical Gaps Still Exist, but Inference Side Has Reached "Good Enough" Threshold

The rise in market share for domestic chips does not mean that the technical gap has been eliminated.

Most industry analysts estimate that Chinese domestic AI chips are still 5 to 10 years behind Nvidia on the data center training side. For training large language models (LLMs) with trillions of parameters, Nvidia's high-end GPUs remain the preferred choice. For instance, a cluster of 50,000 Hopper series GPUs was used to train the R1 model by DeepSeek.

However, on the inference side, the situation is different. Industry observers believe that for 90% of commercial application scenarios (including image recognition, chatbots, autonomous driving, etc.), domestic chips have reached the "good enough" threshold, making the switch from Nvidia to domestic solutions a feasible business decision. Reinforced expectations for further sanctions have accelerated the motivation for this switch.

The real bottleneck lies in the software ecosystem. Nvidia's CUDA platform, built over more than a decade, has become the de facto standard for AI development. Domestic chip manufacturers have invested substantial resources in compatibility: MetaX announced that its C500 series will support CUDA compatibility, while Huawei plans to fully open source its CANN platform in 2025 to expand the developer ecosystem. Cambricon and Moore Threads have also developed their own tools to translate from CUDA to their programming languages. The pace of ecosystem catch-up will determine the ceiling for the market share of domestic chips.

Domestic AI Chip Enterprises Intensely Sprinting to Capital Markets

The shift in market share is reflected in the capital market concurrently.

Since the beginning of 2026, a wave of IPOs has surged in China's GPU sector. Wallan Technology and MetaX have listed on the Sci-Tech innovation board, Tian数智芯 has been listed on the main board of the Hong Kong Stock Exchange, and the listing application of Suiruan Technology on the Sci-Tech innovation board has also been accepted. Baidu has announced plans to spin off Kunlun芯 for independent listing, and sources reveal that Alibaba is also considering a similar spinoff for T-Head.

Huawei's R&D investment is expected to reach 192.3 billion yuan by 2025, accounting for 22% of its revenue, focusing on chips, software, and manufacturing tools to further reduce dependence on U.S. technology. Huawei's rotating chairman Xu Zhijun stated at MWC 2026 that Huawei will become "an alternative choice to ensure that global AI computing power supply is uninterrupted." According to Reuters, Huawei's new generation Ascend 950PR chips have attracted ordering interest from giants like ByteDance and Alibaba, with a shipment target of about 750,000 units for 2026, and mass production set to start in the second half of the year.

For Nvidia, even if the H200 has been approved for export to China, the foundation of trust has already been shaken. Beijing's self-controllable policy is no longer just a vision, but a reality constituted by each domestic chip operating in data centers. When the market share data for 2026 is released, whether the 55% figure rebounds or continues to decline will depend on whether Washington's export policies turn again, as well as the pace at which domestic chips catch up on the training side.