China AI Chip Industry 2025: Huawei Ascend, Biren, Moore Threads
China's AI chip industry has accelerated dramatically since US export controls restricted access to advanced NVIDIA GPUs. Led by Huawei's Ascend series, Biren Technology, and Moore Threads, domestic chipmakers are building alternatives for both AI training and inference workloads. While still behind NVIDIA in raw performance, Chinese AI chips are closing the gap in inference efficiency and are finding increasing adoption in data centers, smart city infrastructure, and edge computing applications.
TL;DR
China's domestic AI chip market reached approximately 80 billion RMB in 2025. Huawei's Ascend 910C became the most widely deployed domestic AI training chip, installed in over 200,000 servers across China's major cloud providers. Biren and Moore Threads focused on inference optimization. The gap with NVIDIA narrowed to approximately 2-3 years in inference performance, though training still lags by 4-5 years.
Key Insights
Domestic AI Chip Market Size
China's domestic AI chip market reached approximately 80 billion RMB, driven by government procurement mandates, cloud provider adoption, and enterprise AI deployment. The market is projected to exceed 150 billion RMB by 2028.
Huawei Ascend Deployments
Huawei's Ascend AI chips have been deployed in over 200,000 servers across China's major cloud platforms including Huawei Cloud, Alibaba Cloud, Tencent Cloud, and Baidu AI Cloud, making it the most widely used domestic AI accelerator.
Inference Performance Gap
The performance gap between Chinese AI chips and NVIDIA's latest offerings has narrowed to approximately 2-3 years for inference workloads (Huawei Ascend 910C vs A100 for inference), though the training performance gap remains larger at 4-5 years.
Government Procurement Mandate
Chinese government agencies and state-owned enterprises are required to source over 70 percent of their AI computing hardware from domestic suppliers, creating a guaranteed market for domestic chipmakers and accelerating adoption.
Side-by-Side Comparison
| Chip | Company | Process | Key Use Case | Performance Tier |
|---|---|---|---|---|
| Ascend 910C | Huawei | 7nm SMIC | Training + Inference | Near A100 level |
| Ascend 310P | Huawei | 12nm | Inference / Edge | Mid-tier inference |
| BR100 | Biren | 7nm | Training | Near A100 training |
| BR104 | Biren | 7nm | Inference | Cost-effective inference |
| MTT S4000 | Moore Threads | 12nm | Inference | Entry-level GPU |
| Cambricon MLU370 | Cambricon | 12nm | Inference | Cloud inference |
| Inferentia (generic) | Various | 7-14nm | Inference | ASIC inference |
Frequently Asked Questions
Chinese AI chips can increasingly replace NVIDIA for inference workloads, where Huawei Ascend 910C achieves approximately 80-90 percent of NVIDIA A100's inference performance. However, for large-scale training of frontier models, Chinese chips still lag significantly due to weaker interconnect technology (Huawei's HCCS vs NVIDIA's NVLink), software ecosystem maturity (CANN vs CUDA), and limited access to advanced manufacturing processes below 7nm. The gap is closing for inference but remains substantial for training.
US sanctions have both hindered and stimulated the industry. Restricted access to NVIDIA H100/H200 and AMD MI300X GPUs forced Chinese companies to develop domestic alternatives. SMIC's inability to access EUV lithography limits manufacturing to 7nm and above. However, sanctions created a captive domestic market, massive government funding (estimated 100+ billion RMB), and urgency that accelerated development. Chinese companies also found creative workarounds including chip stockpiling, third-party procurement through intermediaries, and architectural innovations to maximize performance within process constraints.
Huawei's CANN (Compute Architecture for Neural Networks) is the software framework for Ascend chips, analogous to NVIDIA's CUDA. In 2025, CANN matured significantly with support for PyTorch, TensorFlow, and MindSpore (Huawei's own framework). Major Chinese AI labs including Baidu's ERNIE, Alibaba's Qwen, and SenseTime have optimized their models for Ascend. However, CANN's developer community is still much smaller than CUDA's, documentation quality lags, and debugging tools are less mature. Huawei is addressing this through aggressive developer programs and university partnerships.