Under sustained US export controls, China's AI chip industry is shifting from general-purpose GPUs to custom ASICs. Huawei leads the domestic market with a projected 62% share in 2026, followed by Cambricon at 14%, while Alibaba's T-Head unit pursues parallel processing. Despite hardware performance gains over Nvidia's export-compliant chips, Chinese firms face the critical challenge of breaking Nvidia's CUDA software lock-in by developing alternatives like Huawei's CANN.
US export controls impact
- ▪US export controls are pushing China's AI chip industry away from general-purpose GPUs and toward custom application-specific integrated circuits (ASICs).
- ▪Nvidia's share of the Chinese AI accelerator market has collapsed to zero, which Nvidia CEO Jensen Huang described as a "horrible outcome" for the United States.
ASIC versus GPU strategies
- ▪Moore Threads, founded in 2020 by Nvidia's former China executive Zhang Jianzhong, leads China's domestic general-purpose GPU effort with its MTT S5000 series.
- ▪Biren Technology, Enflame, and Iluvatar CoreX compete in the Chinese GPU space but have not achieved the scale of the ASIC leaders.
- ▪General-purpose GPUs are ideal for AI research, while ASICs sacrifice flexibility for raw efficiency, delivering faster performance at lower power consumption for specific AI tasks.
Huawei Ascend market dominance
- ▪Huawei is projected to capture 62% of China's domestic AI accelerator market in 2026, betting on neural processing units through its Ascend series.
- ▪Huawei's Ascend 950 cards can outperform Nvidia's H20 by 50% to 150% in tokens per second, according to Morgan Stanley data.
- ▪Huawei expects its AI chip revenue to reach approximately $12 billion in 2026, up from $7.5 billion in 2025.
Cambricon domain-specific chips
- ▪Cambricon is projected to capture 14% of China's domestic AI accelerator market in 2026, according to a Morgan Stanley report published on May 8, 2026.
- ▪Cambricon Technologies is building domain-specific architectures with its Siyuan 590 and 690 series.
- ▪Huawei's Ascend 950 cards can outperform Nvidia's H20 by 50% to 150% in tokens per second, according to Morgan Stanley data.
Alibaba T-Head processing units
- ▪Alibaba and Baidu are each expected to capture roughly 5% of China's domestic AI accelerator market in 2026, according to a Morgan Stanley report.
- ▪Alibaba's semiconductor unit T-Head launched the Zhenwu M890 parallel processing unit, claiming three times the performance of its predecessor.
CUDA software ecosystem challenge
- ▪DeepSeek spent months rewriting its core code to work with Huawei's CANN framework, moving away from the Nvidia CUDA ecosystem.
- ▪Huawei has developed CANN and Moore Threads has developed MUSA as software alternatives to Nvidia's CUDA platform.
- ▪China's chip industry faces the challenge of breaking the lock-in of Nvidia's CUDA platform, which millions of AI developers use to write code.
Story comments
Loading comments…