Cambricon

Full briefing

Cambricon Technologies (寒武纪科技; SH: 688256) was founded in 2016 by Chen Tianshi and Chen Yunji, brothers who were researchers at the Institute of Computing Technology at the Chinese Academy of Sciences. The founding team had previously developed the DianNao neural network accelerator, one of the first academic AI chip architectures, which gave Cambricon a strong IP foundation and credibility with Chinese government stakeholders. Cambricon IPO'd on China's STAR Market (Shanghai Stock Exchange) in June 2020, raising approximately 2.5 billion RMB ($350 million) — becoming one of the first AI chip companies to go public in China. Cambricon's MLU (Machine Learning Unit) product family is its core offering. The MLU270, MLU290, and MLU370 series are data center AI accelerators designed for both training and inference workloads. The MLU370-X8 (launched in 2022) delivers approximately 256 TOPS of INT8 performance on a PCIe card and is compatible with a custom SDK called BANG C that allows developers to write GPU-like kernels for Cambricon's architecture. The company has also developed the MLU590, its highest-end training-focused chip, claiming performance in the range of competitive with low-end NVIDIA data center products. All current Cambricon chips are fabricated by SMIC, primarily on SMIC's 7nm-class processes, though SMIC's process limitations constrain Cambricon's density and performance relative to TSMC-fabbed competitors. The December 2020 Entity Listing has been the defining constraint on Cambricon's development. The BIS placed Cambricon on the Entity List in December 2020, citing national security concerns related to China's military modernization. The practical consequences include: inability to access EDA tools from Cadence and Synopsys except under individual license; inability to procure IP blocks with US-origin content; and inability to use TSMC, GlobalFoundries, or other US-regulated fabs. Since SMIC's most advanced nodes use equipment sourced largely from ASML and other suppliers operating under US technology jurisdiction, even SMIC's access to cutting-edge equipment is partially circumscribed, indirectly limiting Cambricon's process options. Cambricon's domestic market position depends on Chinese cloud and enterprise buyers choosing domestically-sourced AI chips for reasons of supply chain security, government procurement mandates, or cost competitiveness. Cambricon's chips have been deployed by Lenovo, Inspur (浪潮), and Chinese university research labs. Chinese government policy — including the "dual circulation" economic strategy and "Made in China 2025" — actively promotes procurement of domestic chips, which provides Cambricon with a policy tailwind. However, the software ecosystem disadvantage remains significant: Cambricon's BANG C programming model lacks the breadth of CUDA's ecosystem, and most open-source AI frameworks (PyTorch, TensorFlow) have optimized CUDA backends but only experimental Cambricon support. Cambricon has also pursued edge AI chip products through its MLU series, targeting smart city surveillance, autonomous driving, and IoT inference applications where Chinese domestic procurement is strong. The company's Sichuan datacenter chip production partnership and its research collaborations with Chinese military and academic institutions have been cited as factors in the Entity Listing and continue to complicate its relationship with Western technology suppliers. Cambricon's financial performance has been loss-making, reflecting the high R&D costs of chip development versus the relatively small domestic market for its products compared to the global addressable market NVIDIA serves.