Analysts project Chinese domestic AI accelerator makers, led by Huawei and Cambricon, will capture 90% of China's local market by 2026, signaling rapid decoupling from Nvidia and AMD.

Chinese AI accelerators are poised to capture 90% of the country’s domestic market as U.S. export controls and Beijing’s procurement mandates systematically remove American-made hardware from AMD and Nvidia. According to a recent TrendForce report, Cambricon and Huawei are expected to be the primary beneficiaries of this transition, as noted by DigiTimes. However, the critical question remains whether Chinese vendors can manufacture and ship sufficient volumes to meet surging domestic demand.

Nvidia’s dominance in China has eroded sharply. The company commanded 66% of the Chinese AI accelerator market in 2024, but that share fell to 40% in 2025 and is projected to drop to just 8% in 2026, according to earlier estimates from Bernstein investment bank. With no official shipments of new accelerators to Chinese clients in the first half of the year, Nvidia CEO Jensen Huang stated in May that his company’s market share in the PRC was effectively “zero.” Nevertheless, some Nvidia GPUs continue to enter the market unofficially, as local firms remain heavily dependent on Nvidia’s CUDA ecosystem and high-end accelerators. Despite this leakage, the vast majority of new AI deployments in China now rely on domestically designed and produced hardware.

China’s total addressable market (TAM) for AI accelerators exceeded 4 million units in 2025, according to data published by Guancha.cn. During that period, 2.2 million Nvidia AI GPUs entered the Chinese market, securing a 55% unit share despite the broader decline. Nvidia still significantly outperformed its closest rival, Huawei, which shipped 812,000 AI accelerators and captured 20.3% of the market. Other manufacturers trailed considerably: Alibaba’s T-Head produced 265,000 units, followed by AMD with 160,000. Cambricon and Kunlunxin each supplied approximately 116,000 AI processors, while all other vendors combined shipped fewer than 100,000 units.

Government backing is accelerating the transition. China recently added homegrown AI chips to its “secure and reliable” procurement list for the first time, drafted a $295 billion plan to build a national AI data center grid running on 80% homemade silicon, and positioned Huawei to enter the South Korean market with new Atlas SuperPods, each cluster packing 8,192 Ascend 950 accelerators. “This year, the Chinese government has actively encouraged the adoption of domestic AI chips,” the TrendForce report states. “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.”

Building on this momentum, TrendForce now forecasts that shipments of high-end AI processors developed by Chinese companies will surge by more than 83% year-over-year in 2026 as domestic production capacity and deployment scale. Consequently, analysts project domestic AI accelerators will capture nearly 90% of sales—up from 45% last year—leaving foreign suppliers like AMD and Nvidia with roughly 10%. This represents a significant upward revision from the firm’s December 2025 outlook, which had estimated Chinese processors would account for only around 50% of the high-end AI chip market in 2026.

Replacing the 2.36 million AI accelerators previously supplied by AMD and Nvidia—which held a 59% unit share last year—will require immense industrial effort, assuming the TAM remains near 4 million units. To bridge this gap, TrendForce reports that China is adopting a “dual-track strategy.” This approach combines merchant accelerators from suppliers like Huawei and Cambricon with custom AI ASICs developed internally by hyperscalers including 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 notes. Hyperscale cloud providers are increasingly favoring their own silicon due to lower costs and workload-specific optimization, while merchant hardware developers such as Huawei, Biren, and Cambricon are simultaneously scaling output to meet robust demand.

Whether China’s semiconductor industry can realistically replace 1.96 million high-end AI accelerators within a single year remains uncertain. Sustaining a 4 million unit TAM would require a 2.2-fold increase in annual AI accelerator output. SMIC, China’s largest and most advanced foundry, recently reported Q2 2026 revenue of $3.005 billion, up from $2.505 billion in Q1 2026 and $2.209 billion in Q2 2025. While this suggests rising production volumes and pricing power, it is unclear whether the company’s 36% year-over-year revenue growth signals the capacity to scale high-end AI accelerator output by more than double compared to 2025. Compounding the challenge is a severe shortage of domestic high-bandwidth memory (HBM). Although Huawei has reportedly secured substantial stocks of Samsung-sourced HBM2-class memory, those supplies are finite. As a result, its Ascend 950-series accelerators will increasingly rely on proprietary HiBL 1.0 and HiZQ 2.0 memory architectures rather than industry-standard HBM2 or HBM3. Meanwhile, DRAM manufacturer CXMT is preparing for HBM3 production in late 2026, though the timeline for reaching meaningful volume remains opaque.

Beyond hardware, Nvidia’s CUDA software stack remains a formidable competitive moat. While raw performance can theoretically be matched through engineering effort, replicating a mature software ecosystem cannot be done quickly. Huawei opened its CANN software stack to external developers last year to accelerate iteration, though the extent to which it has met its internal targets remains unverified. Undoubtedly, China’s domestic AI software stack continues to mature annually, meaning a growing number of new deployments will likely operate on homegrown frameworks rather than CUDA. In terms of raw capability, Chinese hardware has progressed substantially. Huawei’s solutions can currently outperform Nvidia’s NVL72 GB200 rack-scale system, albeit with higher power consumption. Provided energy constraints are manageable, Huawei can deploy AI data centers that match or exceed the computational throughput of Nvidia-based equivalents.

Ultimately, replacing American GPUs entirely while maintaining a 4 million unit TAM would require China to produce 1.96 million additional AI accelerators in 2026 alone—a 2.2x jump that appears highly improbable given current constraints at both TSMC and nascent domestic memory producers. Consequently, while Chinese vendors may indeed secure 90% of the domestic market share, the absolute size of that market could contract significantly without continued access to American hardware.