HiSilicon
Develops the Ascend series of AI processors and full-stack software/hardware AI solutions, serving as the absolute main force in domestic substitution for AI computing chips.
The only company in China with full-stack self-developed AI capabilities (chips + CANN operator library + MindSpore framework). Despite the most stringent US export controls, it has maintained product iteration and built the largest AI chip ecosystem in China.
Company Overview
HiSilicon is a semiconductor design subsidiary wholly owned by Huawei, founded in 2004 and headquartered in Shenzhen. Historically, HiSilicon was renowned for its Kirin series SoCs for mobile devices—the Kirin 9000 series at one point surpassed the flagship chips of Qualcomm in performance.
Since 2019, due to sanctions imposed under the U.S. Entity List, TSMC has been unable to manufacture advanced process chips for HiSilicon, forcing the discontinuation of the Kirin mobile SoC line. Following this, HiSilicon shifted its focus entirely to AI computing chips—the Ascend series—and built a comprehensive domestic AI computing ecosystem around Ascend. The Ascend series has now become the standard chip in domestic AI computing centers.
Core Business
Ascend AI Processors | Core Products
- Ascend 910B: Currently the most powerful AI training chip in China. According to industry evaluations, its performance on certain typical AI training tasks can reach 80%-90% of NVIDIA A100's level. It adopts the self-developed Da Vinci architecture, specifically optimized for AI matrix operations.
- Ascend 310: A chip designed for inference scenarios with low power consumption and high cost-effectiveness, widely used in intelligent edge and security scenarios.
- CANN (Heterogeneous Computing Architecture): The underlying operator library and programming interface benchmarked against NVIDIA CUDA. The maturity of CANN is a key bottleneck for China's domestic AI chip ecosystem. Huawei is investing substantial resources to promote CANN's adaptation to mainstream AI frameworks (PyTorch, TensorFlow).
MindSpore AI Framework
Huawei's self-developed deep learning framework, deeply integrated with Ascend chips. Although TensorFlow and PyTorch support Ascend through CANN adaptation, MindSpore achieves the highest runtime efficiency on Ascend. Through MindSpore, Huawei has built an end-to-end toolchain covering everything from model training to deployment.
Full Systems and Solutions
Huawei not only sells chips but also provides Ascend-based AI servers (Atlas series) and full-rack liquid cooling solutions (Atlas 900 training cluster), directly delivering projects for telecom operators and government AI computing centers.
Technical Moat
Full-Stack Autonomous Capability: Globally, only two companies can independently develop AI chips, underlying operator libraries, and deep learning frameworks simultaneously: Nvidia and HiSilicon. Although gaps exist at every level compared with Nvidia, under the U.S. export ban, Huawei is the only domestic enterprise capable of providing a complete "chip-software-framework-system" replacement solution.
Policy and Market Barriers: In China's domestic xinchuang (IT application innovation) market, especially in key industries such as government, telecommunications operators, and finance, Ascend is the preferred or even the only AI computing power solution. With its deep government-enterprise relationships and a nationwide marketing and service network, Huawei has established market barriers that competitors find difficult to overcome.
Market Landscape
| Dimension | Data |
|---|---|
| Domestic AI accelerator card share | First (accounting for approximately 60-70% of domestic intelligent computing center procurement volume) |
| Software ecosystem | The number of CANN operators has approached 80% of the core CUDA operator coverage rate |
| Main competitors | NVIDIA (global), Cambricon, Hygon Information Technology (domestic) |
| Core customers | The three major telecom operators, government intelligent computing centers across regions, iFlytek, and others |
Financials & Growth
| Indicator | Description |
|---|---|
| Revenue Estimate | HiSilicon's revenue is not disclosed separately; industry estimates place it at approximately 20-30 billion RMB |
| Core Growth Logic | Accelerated domestic AI computing center construction → Surge in Ascend procurement; CANN ecosystem maturation → Lowered adaptation threshold |
Risks and Summary
Key Risks:
- Advanced process node bottleneck: Due to the inability to use TSMC's 3nm/5nm processes, the single-card performance ceiling of Ascend chips is limited by domestic foundry capacity.
- Software ecosystem remains a weakness: Many AI developers are accustomed to the CUDA ecosystem, and migrating to CANN requires additional engineering costs.
- NVIDIA's China-specific chips (e.g., H20) are still sold in the Chinese market, directly exerting competitive pressure on Ascend.
Core Industrial Value: HiSilicon is China's last line of defense in AI computing power. It is not only a pioneer in domestic substitution but also a key variable in whether the domestic AI industry chain can achieve independence. The success or failure of the Ascend ecosystem will largely determine China's AI industry's position in global competition.