What Cambricon Technologies does

Cambricon Technologies (寒武纪), formally 中科寒武纪科技股份有限公司, is a Beijing-headquartered Chinese AI chip designer founded in 2016 by Chen Tianshi and listed on the Shanghai STAR Market (ticker 688256) since 2020. The company designs AI processor chips (ASICs) across three deployment tiers: terminal processor IPs (Cambricon-1A, 1H, 1M), edge chips (Siyuan 220 / MLU220 in M.2 and SOM form factors), and cloud accelerators (Siyuan 270, 290, 370, 590 series plus the XuanSi 1000 4-chip module). The Siyuan 370 was China's first cloud AI chip to use chiplet technology at 7nm, integrating two AI chiplets per die and supporting MLU-Link multi-chip interconnect at 200GB/s. On the software side, Cambricon operates a unified development environment — Cambricon NeuWare for training and inference, the MagicMind MLIR-based inference acceleration engine, and the proprietary BANG programming language with associated CNCC/CNAS/CNLINKER/CN-GDB toolchain — enabling model migration across its full hardware portfolio.

The company's revenue model is primarily B2B direct enterprise sales of AI accelerator cards to server OEMs and system integrators (Inspur, Lenovo, Foxconn, Supermicro, AMAX, Nettrix, Powerleader, WUZHOU, Supercloud), with secondary revenue from terminal processor IP licensing (Cambricon-1A was deployed in tens of millions of smartphones). Customers include hyperscale internet companies (ByteDance, Alibaba, Tencent), foundation-model developers (DeepSeek), telecom operators (China Telecom), cloud and edge platform providers (Jinshan Cloud), and government, manufacturing, transportation, energy and retail enterprises. Sales are quote-based under multi-year enterprise contracts; no public list pricing is published. Customer wins accelerated sharply in 2025: ByteDance drove H1 2025 revenue to surge 43x to $404M, and Q1 2026 revenue reached $423M (+160% YoY) with net profit up 185% to CNY 1B.

The addressable market is structurally supported by China's Xinchuang (Information Technology Innovation) procurement framework, which has approved only Cambricon and Huawei as government-approved AI hardware suppliers — effectively excluding Nvidia and AMD from state-funded data center procurement — and by Beijing electricity subsidies (up to 50% off) for data centers deploying domestic chips. Cambricon ships across terminal, edge, and cloud tiers, has offices in eight Chinese cities, and was ranked China's most valuable AI company at a $88B valuation in the Hurun 2025 list (+165% YoY). The firm employs 1,001–5,000 staff, achieved its first annual profit in 2024 (CNY 2.1B / ~$300M, reversing a 2023 net loss of CNY 455.8M), and declared a maiden dividend plus CNY 20M in share buybacks. Strategic partnerships in the 2025–2026 period include DeepSeek (V4 co-optimization), SenseTime (chip+model ecosystem), StepFun (AI-Chip Ecosystem Alliance), and Baidu PaddlePaddle framework integration.

Name

Cambricon Technologies

Legal name

中科寒武纪科技股份有限公司

Website

https://cambricon.com

Company type

Public

Founded year

2016

Operating status

Operating

Headcount range

1,001–5,000 employees

Short description

Cambricon Technologies is a Beijing-based AI chip designer producing proprietary processors and accelerator cards across terminal, edge, and cloud tiers, sold via server OEMs (Inspur, Lenovo, Foxconn, Supermicro) and direct to hyperscalers (ByteDance, Alibaba, Tencent) and Chinese government buyers under the Xinchuang procurement framework.

Ownership category

akta.pro rank

Product category

AI Accelerator Chips

NAICS

Computer Terminal and Other Computer Peripheral Equipment Manufacturing (334118)

SIC

Computer Peripheral Equipment, Nec (3577)

akta.pro primary industry

AI Hardware IP & Chiplet/Packaging for Accelerators (EDA/IP, advanced packaging) (HDAAAAAL)

akta.pro secondary industry

Edge AI Hardware & NPUs (SoCs, modules, accelerators) (HDAAAJAC)

Keywords

• AI accelerator chips
• Cloud AI processors
• Edge computing chips
• AI inference engines
• Processor IP licensing

Headquarters

HQ city

Beijing

HQ country

China

HQ region

Asia

Offices7 records

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Personnel, Operations, Supply Chain, Marketing or Sales, Infrastructure

Revenue model

1. AI Accelerator Cards and Chips Sales: Cambricon generates revenue primarily from selling AI accelerator cards (MLU100, MLU270, MLU290, MLU370 series) and chips to enterprise customers including server OEMs, system integrators, and end-user enterprises. Products are sold directly through sales teams. The company achieved dramatic revenue growth from ~CNY 1 billion in 2023 to CNY 6.5 billion in 2024 (5x+ increase), driven by surging domestic AI demand and China's semiconductor self-sufficiency push.
2. Terminal Processor IP Licensing: Cambricon licenses its terminal AI processor IP (Cambricon-1A, 1H, 1M series) to smartphone and device manufacturers. Cambricon-1A was deployed in tens of millions of smartphones. IP licensing provides recurring revenue from device volume royalties.

Pricing tiers

Go-to-market motion1 record

Distribution channels3 records

Marketing channels5 records

Core offering

Cambricon Technologies designs and sells AI accelerator chips (ASICs), processor IPs, and accelerator cards for cloud, edge, and terminal deployment, supported by a unified NeuWare software platform and MagicMind inference engine. The company sells these products directly to server OEMs, system integrators, and large enterprise customers across China, spanning internet/cloud, government, telecom, finance, manufacturing, transportation, and energy sectors. The company also licenses terminal processor IPs (Cambricon-1A/1H/1M) to device manufacturers.

Product overview

Cambricon Technologies, founded in 2016, provides a comprehensive AI chip portfolio spanning terminals, edge, and cloud with unified software platforms. The product lineup includes: Cambricon-1M and Cambricon-1H terminal processor IPs for edge devices; Siyuan 220 (MLU220) edge AI chips with M.2 accelerator and SOM module form factors; Siyuan 270 and Siyuan 370 cloud AI chips with chiplet technology; MLU370-X8/X4/S8 training and inference accelerator cards; and Siyuan 290 training chips with XuanSi 1000 accelerator. The software ecosystem comprises NeuWare unified development platform supporting TensorFlow, PyTorch, Caffe, MXNet frameworks, and MagicMind MLIR-based inference engine for cross-framework model deployment. Cambricon serves applications in smart manufacturing, autonomous driving, smart finance, healthcare, energy, and transportation through partnerships with leading Chinese technology companies and system integrators.

Differentiator

Problem solved

Functional benefit

Brands

• 思元 (Siyuan): Cloud AI chip series including MLU370, MLU270 products
• Cambricon NeuWare
• MagicMind
• BANG Language

Products and services

• Siyuan 370 (MLU370) Cloud AI Chip Third-generation cloud AI chip using 7nm process and chiplet technology, integrating 39 billion transistors. Delivers 256TOPS INT8 inference and supports LPDDR5 memory with 3x bandwidth. Features MLUarch03 architecture and MLU-Link multi-chip interconnect.
• MLU370-X8 AI Training Accelerator Card High-end training accelerator card featuring dual Siyuan 370 chips with MLU-Link interconnect. Provides 250W max power, 256TOPS INT8 inference, 24TFLOPS FP32 training performance, and 200GB/s MLU-Link bandwidth (3.1x PCIe 4.0). In 8-card configuration achieves 155% of 350W RTX GPU performance on YOLOv3, Transformer, and BERT.
• MLU370-X4 AI Accelerator Card

Quantifiable outcome

• MLU370-X8 8-card parallel achieves average performance of 155% of mainstream 350W RTX GPU on YOLOv3, Transformer, and BERT training tasks
• +4 more outcomes

Named customers5 records

Segments7 records

Ideal customer profiles4 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration6 records

AI capability8 records

Feature7 records

Partnerships

Eleven partnerships are on record, tiered core and minor.

• DeepSeekcoreStrategic or Co-development Partner · 4 April 2026DeepSeek collaborated directly with Cambricon to adapt V4 model's underlying code for Cambricon chips, enabling the V4 model to run on Cambricon hardware. Cambricon's Siyuan series chips achieved compatibility with DeepSeek-V3.2, enabling frontier-level AI inference on domestic chips.
• SenseTimecoreStrategic or Co-development Partner · 17 October 2025Cambricon and SenseTime formed a strategic partnership to strengthen China's AI supply chain. The collaboration focuses on joint AI chip deployment and ecosystem development, combining SenseTime's AI model capabilities with Cambricon's hardware.
• StepFun (阶跃星辰)coreStrategic or Co-development Partner · 29 July 2025Cambricon joined Huawei Ascend, Biren, and Moore Threads in the 'AI-Chip Ecosystem Alliance' with domestic AI model startup StepFun. The alliance aims to optimize StepFun's foundational model Step 3 for domestic AI chips to counter Nvidia's renewed presence in China following US export control adjustments.
• Baidu PaddlePaddlecoreTechnology or Integration · 11 May 2020Cambricon and Baidu completed adaptation of Baidu's Paddle Lite inference framework with Cambricon chips, enabling faster AI deployment across edge and cloud scenarios. The collaboration began with cloud inference and expanded to edge inference markets, with joint work on cloud training planned.
• China TelecomcoreStrategic or Co-development Partner · 19 November 2019China Telecom and its DICT Application Capability Center signed strategic cooperation agreements with Cambricon to jointly build 5G, AI, and edge computing ecosystems. Cambricon provides cloud, edge, and terminal AI chips while China Telecom offers its telecom infrastructure and market reach.
• IDC (International Data Corporation)minorOthersIDC provides market research data on China's AI chip market that Cambricon uses for industry benchmarking. IDC's data shows Cambricon shipped approximately 116,000 AI chips in 2025.
• Inspur (浪潮信息)coreImplementation/ SI/ Consulting PartnerInspur is a key hardware partner integrating Cambricon MLU accelerator cards into its servers. Cambricon's MLU270-X8 and MLU370 series have completed certification with Inspur's server platforms (NF5468M5, NF5280M5, etc.). Inspur's AI product line副总经理 personally praised the MLU370 series.
• LenovocoreImplementation/ SI/ Consulting PartnerLenovo integrates Cambricon MLU270 accelerator cards into its servers (including SA5212M5 platform). Cambricon's Siyuan270 series is officially supported on Lenovo server products.
• Foxconn Industrial InternetcoreImplementation/ SI/ Consulting PartnerFoxconn Industrial Internet integrates Cambricon AI accelerator cards into its server and data center solutions, serving enterprise customers across manufacturing and cloud infrastructure.
• SupermicrocoreImplementation/ SI/ Consulting PartnerSupermicro integrates Cambricon MLU accelerator cards into its server products (SYS-4029GP-TRT2, S120M3 models), enabling GPU and AI accelerator server solutions for enterprise customers.
• AMAX, Nettrix, Powerleader, WUZHOU, SupercloudminorImplementation/ SI/ Consulting PartnerMultiple additional hardware partners integrate Cambricon MLU accelerator cards into their server products, expanding Cambricon's reach across China's server OEM ecosystem.

Scale indicators11 records

Recent moves7 records

Expansion highlights6 records

Direct peers

• Huawei (HiSilicon / Ascend): Huawei's HiSilicon division designs Ascend AI accelerators (910B, 910C) targeting Chinese data center training and inference. Shipped 812K units in 2025 per IDC, making it the largest domestic competitor; the only other company on the Xinchuang government-approved supplier list alongside Cambricon.
• Moore Threads: Chinese fabless AI GPU/chip designer founded in 2020, competing directly in domestic AI training/inference accelerators. Part of the same AI-Chip Ecosystem Alliance with StepFun, Huawei, and Biren, optimizing domestic models for non-Nvidia hardware.
• Biren Technology: Chinese AI chip designer developing training and inference accelerators (BR100 series) for the domestic market. Direct competitor in the same Xinchuang-aligned segment and fellow member of the StepFun AI-Chip Ecosystem Alliance.
• MetaX: Chinese AI chip startup developing GPU/accelerator products (C500 series) for domestic data center and edge AI workloads. Direct competitor targeting the same Chinese AI compute market as Cambricon with similar fabless positioning.
• Iluvatar CoreX: Chinese AI accelerator designer (BI/CL series) competing for cloud training and inference workloads in domestic data centers. Comparable fabless model targeting similar hyperscaler and government customers as Cambricon.

Broad incumbents

• Nvidia: Global AI accelerator leader (H100, H200, B100, GB200) and the benchmark Cambricon is positioned against. Currently restricted from Chinese government procurement under Xinchuang but remains the dominant force Cambricon is measured against technologically and commercially.
• AMD: Global AI accelerator competitor (MI300X, MI325X) excluded from China's Xinchuang government procurement. AMD's China-market trajectory is structurally linked to Cambricon's domestic-substitution narrative and US export policy.
• Intel (Habana Labs): Intel's Habana division (Gaudi 2, Gaudi 3) competes in cloud AI training/inference. Like Nvidia and AMD, currently constrained in China by Xinchuang policies that effectively reserve government AI workloads for Cambricon and Huawei.

Emerging players

• Alibaba T-Head (Pingtouge): Alibaba's in-house semiconductor subsidiary designing AI and cloud chips for hyperscale workloads. Shipped 265K AI chips in 2025 per IDC; a key in-house alternative for Alibaba that competes with Cambricon for the same hyperscaler procurement budgets.

Regional players

• Graphcore: UK-based AI processor (IPU) designer (now part of SoftBank) comparable in targeting AI training/inference with proprietary architectures. Different geographic footprint but similar ASIC-first approach and software-stack integration philosophy to Cambricon.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks7 records

Key highlights7 records

Customer concentration

• GitHubgithub.com/Cambricon

Revenue estimate

Valuation estimate

Number of profiles

Profiles4 records

Funding overview

Funding rounds5 records

Investors21 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments1 record

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Cambricon Technologies