Skip to content
C
Core Compute & Comms Hardware · NPU/AI Compute Chips

Cambricon

📈 688256:SH🌍 China

Develops cloud and edge AI chips and supporting foundational system software platforms; it is a rare independent AI chip architecture design company in China.

⚔️ Core Moat

Possesses a self-developed AI instruction set architecture and independent underlying innovation capabilities in NPU chip design. It holds a place in the non-Huawei domestic AI chip market and has been deployed in multiple AI computing center projects.

Company Overview

Cambricon Technologies was founded in 2016 and co-founded by brothers Chen Tianshi and Chen Yunji, with its headquarters located in Beijing. The company originated from AI chip research achievements at the Institute of Computing Technology, Chinese Academy of Sciences. The name "Cambricon" is derived from the geological period "Cambrian Explosion," symbolizing an explosive growth in the AI chip industry.

Cambricon is the first unicorn company in China dedicated to AI chips and was listed on the STAR Market in 2020. In its early days, the company entered the market through terminal IP licensing models (which once licensed to HiSilicon for use in mobile SoCs), but later shifted to self-developed cloud and edge AI chips following sanctions imposed on Huawei. Despite facing competitive pressure from both NVIDIA and HiSilicon, Cambricon remains a scarce target in the domestic market for its independent NPU architecture design capabilities.

Core Business

Siyuan Series AI Accelerator Cards (Cloud Training and Inference)

  • Siyuan 590: Flagship cloud AI chip under development, targeting large model training scenarios, with performance goals benchmarked at NVIDIA A100 level.
  • Siyuan 370: Current main product, targeting AI inference scenarios, and has won bids in multiple smart computing center projects.
  • Siyuan 270: Previous-generation product, targeting edge computing scenarios.

Cambricon Neuware (Basic System Software)

A low-level software platform benchmarked against NVIDIA CUDA, providing AI framework adaptation, operator libraries, and model deployment toolchains. Although far behind CUDA in user base and ecosystem richness, Cambricon has gradually built its own software ecosystem through participation in domestic Xinchuang (information technology application innovation) projects and national smart computing center construction.

Technical Moat

Self-developed Instruction Set Architecture: Cambricon has an AI instruction set architecture with independent intellectual property rights, not relying on ARM or x86, and possesses independent innovation capabilities at the underlying level of chip design. This is rare among domestic AI chip companies.

Domestic Substitution Window: Against the backdrop of China-US technology decoupling, domestic governments and operators have strong localization demand when procuring AI computing power. Leveraging its brand recognition as a listed company and a relatively complete product line, Cambricon is highly competitive among non-Huawei domestic solutions.

Market Landscape

DimensionData
Cloud AI ChipsFirst tier among domestic independent vendors
Main CompetitorsHuawei HiSilicon, NVIDIA
Core CustomersGovernment intelligent computing centers, research institutions, telecom operators

Finance and Growth

MetricData
Revenue (2024)Approximately RMB 700 million
Net ProfitStill in a loss-making state
Core DriversAI computing center project deliveries + domestic substitution policy-driven growth

Risks and Summary

Key Risks:

  1. Small revenue scale with sustained losses; profitability remains questionable.
  2. Compared with HiSilicon (Huawei), there is a significant gap in the software ecosystem.
  3. Constrained by advanced process node foundry limitations, with a clear ceiling on chip performance.

Core Industry Value: Cambricon represents a possibility for China's independent AI chip design capability. Although there is still a considerable gap compared with NVIDIA and HiSilicon, it has demonstrated that domestic companies can build the underlying architecture of AI chips from scratch without relying on external IP.