IDC Report: NVIDIA's China AI accelerator share falls to 34%, Huawei follows closely at 28%

Main body of the article

Recently, the latest report released by market research firm IDC shows that in the first half of 2026, the landscape of China's data center AI accelerator market underwent notable changes. Among them, NVIDIA's market share fell to 34%, while Huawei followed closely with a 28% share, with shipments reaching 572,000 units, nearly doubling year-on-year. At the same time, the market shares of domestic companies such as Alibaba's T-Head, Cambricon, and Baidu's Kunlunxin continued to rise, and domestic AI chip manufacturers are accelerating into the stage of large-scale delivery.

More noteworthy than the changes in market share is that the structure of domestic AI computing power demand is undergoing a shift. IDC data shows that in the first half of this year, China's total AI accelerator shipments were about 2.008 million units; among them, non-GPU accelerator shipments reached 1.016 million units, up nearly 80% year-on-year, exceeding GPU shipment scale for the first time.

Behind this change is the trend of China's artificial intelligence industry accelerating toward large-scale application. As applications such as large models and agents continue to expand, enterprises' demand for AI computing power infrastructure is growing day by day, especially in industries such as finance, telecommunications, the internet, and manufacturing, creating new conditions for domestic chip technology innovation and commercialization.

The IDC report points out that in the first half of 2026, Huawei's AI accelerator shipments reached 572,000 units, with a market share of 28%; T-Head held a 7% market share, while Cambricon, Kunlunxin, and Hygon held 5%, 4%, and 3% respectively. Among the 11 Chinese semiconductor manufacturers tracked by IDC, 9 achieved year-on-year shipment growth.

In this regard, Wang Qinglin, manager of Ruidaheng Research Institute, told a reporter from China Business Journal that this reflects that the domestic AI chip market is forming a competitive landscape led by leading manufacturers and jointly developed by multiple technology routes. On the one hand, companies represented by Huawei continue to expand in large intelligent computing centers, the internet, and industry computing markets by relying on chip architecture, software ecosystems, and system integration capabilities; on the other hand, companies such as T-Head, Cambricon, and Kunlunxin have also gradually formed differentiated product positioning and are expanding applications in fields such as cloud computing, large model training, and inference.

Luo Guozhao, director of the CHIP China laboratory, said that in the past few years, the demand for computing power from large models was mainly concentrated in links such as model pre-training and parameter optimization, but the market placed more emphasis on the high-performance computing capabilities of chips. As large models gradually enter the commercialization stage, the importance of inference computing continues to rise. In particular, as applications such as agents, multimodal interaction, and enterprise internal knowledge services become more widespread, demand for computing power oriented toward high concurrency, low latency, and low-cost operation continues to increase.

This change is also reflected in the shipment structure of AI chips. IDC data shows that in the first half of this year, China's GPU accelerator shipments were 991,000 units, down 26% year-on-year; non-GPU accelerator shipments reached 1.016 million units, up nearly 80% year-on-year. Among them, four companies—Huawei, T-Head, Kunlunxin, and Cambricon—together accounted for about 95% of the non-GPU accelerator market.

It is worth noting that the report points out that non-GPU accelerator shipments exceeding GPUs does not mean that the importance of GPU technology has declined, but rather reflects that the AI computing power system is moving toward diversification.

More importantly, domestic computing power infrastructure construction provides a continuously expanding market foundation for domestic chips.

The latest data released by the Ministry of Industry and Information Technology shows that as of the end of September 2026, China's intelligent computing power scale reached 2,592 EFLOPS, up 146% year-on-year. The rapid improvement in intelligent computing power supply capacity indicates that artificial intelligence infrastructure is still in an expansion stage, and also means that industrial demand for computing power chips, servers, network interconnection, and supporting software is continuing to be released.

Domestic manufacturers accelerate large-scale commercialization

Among them, Huawei Ascend is promoting the upgrade of domestic computing power from single-chip capability to large-scale cluster capability through supernode architecture and basic software ecosystem construction.

At the 2026 World Artificial Intelligence Conference held in July this year, Huawei publicly displayed the Ascend 950 supernode for the first time, deployed on-site with a 1,024-card cluster, targeting large model training and high-concurrency inference needs, and improving the collaborative efficiency of large-scale computing systems through technologies such as high bandwidth, low-latency interconnection, and unified memory addressing.

According to disclosures from Huawei, as of July this year, the Ascend 384 supernode has been commercially deployed more than 750 sets in total, covering many industries such as the internet, operators, finance, education, healthcare, and manufacturing; at the same time, Huawei has joined hands with more than 3,000 partners to develop more than 7,000 industry solutions, serving more than 2,000 core government and enterprise customers. At the software ecosystem level, Huawei is also continuing to promote the open-source opening of the CANN heterogeneous computing architecture and Mind series basic software, and strengthening adaptation with mainstream development frameworks and inference tools such as PyTorch and vLLM, in order to reduce the technical cost for developers to migrate models and applications between different computing power platforms.

Other domestic AI chip manufacturers are also expanding commercial markets through different paths. According to Alibaba, T-Head has formed a product matrix covering GPU, CPU, storage, and network chips. Among them, the Zhenwu series chips include the latest generation AI processor Zhenwu M890, which has provided commercial applications to more than 650 external customers in more than 20 industries such as autonomous driving, the internet, and financial services through Alibaba Cloud services, supporting different types of AI workloads such as training, fine-tuning, and inference.

At the same time, Cambricon has also achieved business growth driven by large-scale applications. Cambricon stated that, relying on continuous research and development of artificial intelligence chips, basic system software platforms, and cluster software toolchains, its products have achieved large-scale commercial deployment in core industries such as finance and the internet.

In addition, the Kunlunxin P800 has completed large-scale cluster practice verification and, relying on Baidu AI Cloud's BaiGe AI computing platform, achieved the lighting up and large-scale application of a 32,000-card cluster. Through collaborative optimization of chips, computing platforms, and inference frameworks, Baidu seeks to further improve the deployment efficiency of domestic computing power in large model training and inference scenarios.

Luo Guozhao said that competition among domestic AI chips is no longer limited to competition at the level of chip parameters, but has gradually formed an industrial system integrating research and development, software tools, industrial clusters, and applications. In particular, more and more domestic AI chips have entered enterprise business links, and through data feedback from use by a large number of enterprises, their stability and ease of use have improved, further reducing enterprise adaptation and migration costs. This means that the domestic AI chip industry is gradually moving from early product verification into a virtuous stage of large-scale delivery and business monetization.

Industrial competition shifts

As the scale of AI models expands and application loads continue to increase, the global AI chip industry is also undergoing new technological evolution. IDC points out that the focus of market competition is shifting from single-chip performance to overall system capability, including the collaborative operation performance of chips in large computing clusters.

Luo Guozhao believes that under the traditional computing power competition model, chip process technology, the number of computing units, and peak computing performance are often important factors determining product competitiveness. But in large-scale AI computing scenarios, data transmission between chips, memory scheduling, model parallelism efficiency, and software adaptation capabilities also all affect the actual performance of the system.

Especially in large model inference scenarios, enterprises pay more attention to how many tokens can be processed per second, whether response latency can meet business requirements, and the comprehensive computing cost per token. Higher chip peak performance does not mean that computing power utilization efficiency in actual business will be better.

This trend is also prompting domestic enterprises to adjust their technology development direction. The reporter noted that at Huawei Connect 2026 held in September this year, Huawei released the Ascend 960 supernode. Wang Tao, Huawei's deputy chairman and rotating chairman, said that Huawei insists on focusing on "supernode + cluster" to build competitiveness through system architecture innovation, while promoting the construction of an open-source computing power ecosystem.

At the same time, there is still large room for growth in demand in the AI chip market. IDC predicts that in 2026, China's overall AI accelerator shipments will grow 19% year-on-year. By 2029, the global market size of accelerated computing servers will exceed 1 trillion US dollars, reflecting that artificial intelligence infrastructure is still in a continuous expansion cycle.

Deng Zhonghan, an academician of the Chinese Academy of Engineering, previously publicly pointed out that domestic high-end computing power chips still face challenges in three aspects—technical adaptability, ecosystem, and computing power utilization efficiency—in large-scale applications in key scenarios such as data centers and computing power clusters. The adaptation of some domestic chips to the diverse computing needs of large models and computing power clusters is still insufficient, and links such as software-hardware collaboration, standard unification, and application verification also need to be improved.

At the same time, the software ecosystem is also an important factor affecting the commercialization of domestic computing power. Xuan Shanming, CTO of SenseTime's Large Device Business Group, previously said that at the current stage, the prominent bottlenecks in the implementation of domestic computing power are concentrated in aspects such as video memory capacity, inter-card interconnection bandwidth, and fused operator performance. In addition, because there are a large number of adaptation needs among different models, chips, and application scenarios, it is difficult for a single enterprise to complete all adaptation work, and all parties in the industrial chain need to strengthen collaboration.

(Source: China Business Network) Middle of the main text Embedded advertisement