NVIDIA's market share in China drops below 60%, with domestic AI chip deliveries reaching 1.65 million units annually, capturing market share.

Beijing last November ordered state-owned enterprise data centers to fully replace domestically produced products, accelerating the reshaping of the market landscape.

Author: Deep Tide TechFlow

Deep Tide Intro: IDC data shows that in 2025, China’s total shipments of AI accelerator cards were about 4 million units. Domestic vendors collectively delivered 1.65 million units, accounting for 41%. Nvidia’s share fell from about 95% before sanctions to 55%.

Huawei leads the domestic camp with 812,000 chips. Its newly released Atlas 350 accelerator card claims inference performance at 2.87 times that of Nvidia’s H20.

Beijing last November ordered state-owned enterprise data centers to fully replace domestically produced products, accelerating the reshaping of the market landscape.

Three years ago, Nvidia nearly monopolized China’s AI chip market. Today, that landscape has been completely upended.

According to Reuters citing IDC data from market research firms, in 2025 China’s AI accelerator cards (dedicated computing chips for AI servers) had total shipments of about 4 million units. Nvidia remains the largest single supplier, shipping about 2.2 million units, for a 55% share. But compared with the roughly 95% market share before sanctions, this figure has shrunk by nearly 40 percentage points. Meanwhile, domestic vendors collectively shipped about 1.65 million units, taking 41% of the market. AMD shipped about 160,000 units to rank third, at 4%.

The rise of domestic vendors is both an indirect result of the U.S. export controls and a direct outcome of the “domestic substitution” policy.

Huawei leads the domestic camp, Atlas 350 benchmarks against Nvidia H20

In China’s domestic AI chip lineup, Huawei is the biggest winner.

IDC data shows that in 2025, Huawei shipped about 812,000 AI chips, accounting for about 20% of the entire market and nearly half of domestic vendors’ shipments. Alibaba’s chip design arm T-Head (PingTouGe) ranked second with about 265,000 units. Baidu’s Kunlun and Cambricon each shipped about 116,000 units to tie for third. In addition, Hygon, MetaX (Muxi), and TianShu Zhixin (Iluvatar CoreX) accounted for 5%, 4%, and 3% of domestic vendors’ shipments, respectively.

Last month, Huawei unveiled its new-generation AI accelerator card Atlas 350 at the 2026 China Partners Conference in Shenzhen. It is equipped with Huawei’s in-house Ascend 950PR chip. At the launch event, Zhang Dixun, head of Huawei Ascend computing business, said that under FP4 low-precision computing, the Atlas 350 delivers compute performance of 1.56 PFLOPS (quadrillionth operations per second; i.e., ten-thousand billions of operations per second). Its performance is 2.87 times that of Nvidia’s China-specific H20. The card comes with 112GB of in-house high-bandwidth memory, HiBL 1.0. Memory bandwidth is 1.4TB/s, and power consumption is 600W.

However, there is a problem with how this performance comparison is framed. Nvidia’s Hopper-architecture GPUs natively do not support FP4 precision. Atlas 350 is the first domestically produced accelerator card optimized specifically for FP4, so the two cannot be directly compared under the same precision. Huawei’s real competitiveness lies on the inference side: Atlas 350 is positioned for inference (inference) workloads in the AI model deployment stage, not for large-model training.

Huawei’s seven partners have already launched complete server products based on Atlas 350. iFlytek has also announced that its next-generation Spark large model will be adapted to the Ascend 910/950 compute base.

Export controls and domestic substitution drive from both sides

The collapse of Nvidia’s share in China is the result of a two-way squeeze: the U.S. export controls keep escalating, while Beijing’s domestic substitution policy also intensifies pressure.

The timeline is roughly as follows: Starting in October 2022, the U.S. restricted exports of AI chips to China. Nvidia subsequently rolled out compliant downgraded versions of H20 and products such as A800/H800. In April 2025, the Trump administration全面ly banned all AI GPU exports to China. In July of the same year, it resumed export licenses for H20 and AMD MI308. In October, Nvidia CEO Huang Renxun said at a public event that Nvidia’s share in China’s advanced AI accelerator card market “fell from 95% to zero.” In December, Trump allowed Nvidia to export H200 to China, but Chinese companies were told to pause orders for Nvidia chips.

The policy push on the other side was equally forceful. Reuters reported in November 2025 that Beijing issued guidance to new data centers using state-owned enterprise funds, requiring that all domestically produced AI chips be used. Projects with completion progress of less than 30% were required to remove already installed foreign chips or cancel procurement plans.

Reuters’ statistics show that since 2021, China’s AI data center projects have received more than $100 billion in state-owned enterprise funding. Most of China’s data centers are still under construction and have received some form of support from state-owned enterprise funds, which means this policy has extremely broad coverage.

A large data center in Qinghai being built by China Unicom was reported by Reuters as a landmark example of this strategy. The project is valued at $390 million and uses only domestically produced AI chips such as those from T-Head to power it.

The technological gap is real, but the inference side has reached the “good enough” threshold

The increasing market share of domestic chips does not mean the technology gap has been eliminated.

Most industry analysts estimate that China’s domestic AI chips still lag Nvidia by 5 to 10 years on the data-center training side. When training trillion-parameter-class large language models (LLMs), Nvidia’s high-end GPUs remain the first choice. DeepSeek’s training of the R1 model using a cluster of 50,000 Hopper-series GPUs is a realistic example.

But on the inference side, the situation is already different. Industry observers believe that for 90% of commercial application scenarios (including image recognition, chatbots, and autonomous driving), domestic chips have already reached the “good enough” threshold. This makes switching from Nvidia to domestic solutions a viable business decision. Expectations of further strengthened sanctions have further accelerated the momentum for this switch.

The real bottleneck lies in the software ecosystem. Nvidia’s CUDA platform, built up over more than a decade, has become a de facto standard for AI development. Domestic chip vendors have invested significant resources in compatibility: MetaX (Muxi) said its C500 series will support CUDA compatibility. Huawei, meanwhile, fully open-sourced its CANN platform in 2025 to expand the developer ecosystem. Cambricon and Moore Threads have also built translation tools that translate CUDA into their own programming languages. The pace of ecosystem catch-up will determine how high the ceiling for domestic chip market share can be.

Domestic AI chip companies rush into the capital markets

The shift in market share is being reflected in the capital markets at the same time.

Since early 2026, China’s GPU sector has seen a new wave of IPOs. Wallrun Technology and MetaX (Muxi) have listed on the STAR Market; TianShu Zhixin has been listed on the main board of the Hong Kong Stock Exchange; and the STAR Market listing application for Suiyuan Technology has also been accepted. Baidu announced a plan to spin off Kunlun for an independent listing. Insiders say Alibaba is also considering a similar split for T-Head.

In 2025, Huawei’s R&D spending reached RMB 192.3 billion, accounting for 22% of revenue. It focused on chips, software, and manufacturing tools to further reduce reliance on U.S. technology. At MWC 2026, Huawei’s rotating chairman Xu Zhijun said that Huawei will become an “alternative choice to ensure the uninterrupted supply of global AI compute power.” Reuters reported that Huawei’s new-generation Ascend 950PR chip has already attracted ordering interest from giants such as ByteDance and Alibaba. Huawei’s shipment target for 2026 is about 750,000 units, with large-scale mass production set to begin in the second half.

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—it is an established reality composed of every domestically produced chip running in data centers. When 2026 market share data is released, whether the 55% figure rebounds or continues to fall will depend on whether Washington’s export policy turns again, and on how quickly domestic chips catch up on the training side.