China's homegrown AI accelerators to supply 90% of the country's domestic market, analysts suggest — Cambricon and Huawei expected to be the biggest winners in the shift away from Nvidia and AMD

Chinese AI accelerators are set to capture 90% of the country's domestic market as U.S. export controls and Beijing mandates push American-made hardware from AMD and Nvidia out of the market, according to a new report by TrendForce. Cambricon and Huawei are expected to be the biggest beneficiaries of the shift, according to DigiTimes. Yet, the big question is whether Chinese vendors can ship enough AI accelerators to satisfy demand.

China on track for AI accelerator self-sufficiency

Nvidia commanded 66% of China's AI accelerator market in 2024, but its share dropped to 40% in 2025 and was on track to drop to 8% in 2026, according to estimates made by Bernstein investment bank earlier this year. Considering the fact that Nvidia did not officially ship any new accelerators to Chinese clients in the first half of the year, Nvidia's chief executive Jensen Huang said in May that his company's market share in the PRC was 'zero.' Of course, some Nvidia GPUs make it to China 'unofficially' as local companies are too dependent on Nvidia's CUDA and high-end AI accelerators. Still, it is safe to say that the bulk of new deployments in the PRC are based on hardware designed and produced domestically.

China's total available market of AI accelerators topped 4 million units in 2025, according to numbers published by Guancha.cn. Last year, 2.2 million Nvidia AI GPUs made it to the Chinese market, and while Nvidia's market share shrank to 55%, it still significantly outperformed its closest rival, Huawei, which shipped 812,000 AI accelerators and commanded 20.3% of the market.

Shipments by other players were by far lower: Alibaba's T-Head produced 265,000 AI accelerators, followed by AMD with 160,000. Cambricon and Kunlunxin only supplied around 116,000 AI processors each, whereas others shipped fewer than 100,000 units.

"This year, the Chinese government has actively encouraged the adoption of domestic AI chips," the report from TrendForce reads. "This policy push will likely provide priority support to high-potential domestic players, allowing them to substantially expand their market share in China's high-end AI server market. At the same time, the domestic ecosystem is maturing in key areas such as advanced foundry nodes, advanced packaging, and thermal management."

The firm now expects shipments of high-end AI processors developed by Chinese companies to increase by more than 83% year-over-year in 2026 as domestic production capacity and deployments expand. As a result, its analysts project domestic AI accelerators to capture nearly 90% of sales (up from 45% last year), which means that foreign suppliers like AMD and Nvidia will be left with roughly 10%. This latest projection represents a major revision from the research firm's December 2025 outlook, which estimated that Chinese processors would account for approximately 50% of China’s high-end AI chip market in 2026.

Dual-track strategy

As AMD and Nvidia supplied some 2.36 million AI accelerators to the Chinese market last year, commanding a 59% unit share, replacing the majority of them will take a lot of effort, assuming that the TAM will remain at around 4 million units. TrendForce claims that China is set to adopt the so-called dual-track strategy, which involves AI accelerators from merchant suppliers like Huawei and Cambricon along with custom AI ASICs from Alibaba, Baidu, ByteDance, and Tencent.

"Together, these developments are moving China's AI infrastructure away from its heavy reliance on foreign GPUs, toward a dual-track model of 'domestic GPUs + proprietary ASICs,'" the firm claims.

Hyperscale cloud service providers (CSPs) are inclined to expand usage of their own silicon because it is cheaper compared to merchant accelerators and because it is optimized for their workloads and data formats. Meanwhile, developers of merchant AI hardware — such as Huawei, Biren, and Cambricon — will also gradually expand their output of accelerators as demand is very strong.

Bottlenecks

It remains to be seen whether the Chinese chipmaking industry can indeed replace 1.96 million high-end AI accelerators in just one year. To maintain the 4 million unit TAM, China's semiconductor industry will need to increase AI accelerator output by 2.2X in just one year.

SMIC — China's largest and most advanced foundry — this week announced that its Q2 2026 revenue increased to $3.005 billion, up from $2.505 billion in Q1 2026, and from $2.209 billion in Q2 2025. This suggests that the company is both increasing the output of chips and its prices. However, it remains to be seen whether SMIC's 36% YoY revenue increase is an indicator that it can increase output of high-end AI accelerators by over 2X compared to 2025.

Another major bottleneck for the Chinese industry is the lack of domestic production of high-bandwidth memory (HBM). Huawei has reportedly acquired plenty of HBM2-class memory from Samsung, but its stock is not endless, so its Ascend 950-series AI accelerators are set to rely on proprietary HiBL 1.0 and HiZQ 2.0 types of memory, not industry-standard HBM2 or HBM3. While China's DRAM champion CXMT is gearing up for HBM3 manufacturing in late 2026, it remains to be seen how quickly the company can ramp up production to decent levels.

Nvidia's CUDA software stack is the company's biggest advantage after the performance and versatility of its AI accelerators. But while performance can be matched with brute force, the software stack cannot be reproduced quickly. Last year, Huawei opened up its CANN software stack to accelerate its development, though we do not know if the company has achieved its targeted goals with this. Yet, without a doubt, China's AI software stack is getting more mature every year, so many new AI deployments may indeed rely on domestic stacks rather than on CUDA.

A shifting market

U.S. restrictions on exports of advanced AI accelerators, combined with China's own efforts to limit the use of American AI processors domestically, have largely pushed companies like AMD and Nvidia out of the Chinese market. Analysts now expect China-based independent hardware vendors to control 90% of the domestic AI accelerator market in 2026.

Chinese AI hardware has come a long way, and Huawei's solutions can outperform Nvidia's NVL72 GB200 rack-scale system, albeit while consuming more power. Therefore, if power is not a concern, Huawei can build AI data centers with performance that matches or exceeds those based on Nvidia GPUs.

However, replacing American GPUs almost completely while maintaining AI accelerator unit TAM at 4 million units will require China's industry to product 1.96 million AI accelerators in 2026, 2.2X more than in 2025. This seems impossible not only for TSMC, but also for local memory makers that still have to start making HBM memory.

To that end, while Chinese AI accelerators may indeed capture 90% of the domestic market, without hardware from American companies, that market can shrink dramatically in terms of units.