Cambricon: Siyuan Series and Cloud AI Chip Commercialization

1. Introduction

1.1. Vendor Positioning

Cambricon (688256.SH) is the "first AI chip stock on the A-share market": it listed on the STAR Market in 2020-07, achieved its first annual profit since listing in 2025, and had its special "U" designation removed starting 2026-03-16. Within this group of domestic vendors, Cambricon's distinguishing feature is its cloud specialization along the general-purpose GPU/ASIC path — rather than pursuing Huawei-style full-stack ecosystem, it serves leading customers through a three-layer structure of "chip + basic software platform + cluster software toolchain," leaving the upper-layer ecosystem to be built by its customers.

This is the page with the most information gaps in the entire group: official disclosures of customer lists, order amounts, and chip specifications are extremely scarce, and strict search discipline was followed while writing it (see Section 6.3 and the Information Gap Statement).

1.2. Basic Information Card

1.3. 2025 Annual Performance (per Annual Report)

The 2025 annual report is the most important A/B-grade data source for this page (compiled by China Securities Journal and China Economic Net from the annual report); key figures are as follows:

Quarterly net profit (per industry articles relaying the financials, B-grade): Q1 net profit of ¥355 million (first profitable quarter), Q2 ¥683 million (stage high), Q3 ¥567 million, Q4 ¥454 million (two consecutive QoQ declines).

1.4. Position within the AI Harness Framework

Cambricon's support for the six Harness layers is concentrated in L2 (inference execution) and L5 (part of the cluster operations components); the remaining layers are largely filled in by the customer side. Its industry-analytical value lies precisely in revealing another Harness division-of-labor model: when the customer is a leading internet company with complete in-house R&D capability, the chip vendor's Harness role is reduced to "bare compute + basic operator library" (industry analysis view, B/C-grade).

2. Glossary

3. Functions and Product Lines

3.1. Chip Product Lines

Compiled per the Cambricon official website (A-grade) and financial report accounts:

It must be made clear: the official spec sheets for Siyuan 590/690 (compute, memory, interconnect bandwidth) are not publicly listed; no specific parameters are cited anywhere in this document (passing off third-party self-media parameters as official specs is strictly prohibited; see the Information Gap Statement).

3.2. Neuware Software Platform

1. MagicMind inference acceleration engine: quantization, operator fusion, cross-framework support (adapting PyTorch / TensorFlow), positioned as a "cross-chip, cross-framework, unified edge-cloud inference engine" (B/C-grade industry-article account);
2. Cluster operations components: the Neuware platform includes cluster-operations-related components, but public information on the specific monitoring/scheduling components is extremely scarce (gap);
3. Ecosystem division of labor: industry analysis view (B/C-grade) holds that big-tech customers (Alibaba, ByteDance) build their own upper-layer ecosystems, and Cambricon only provides underlying compute — the ecosystem is "sealed off" at the basic software layer.

3.3. Business Structure

Business structure per the 2025 annual report: cloud product line 99.69% / edge 0.05% / IP licensing and software 0.04% / other 0.23%. This is both a success of commercial focus (fully supplying the cloud during the compute-demand surge) and a structural risk (single business, single customer base).

4. Platform Architecture

4.1. Product and Software Architecture

Figure 4-1｜Cambricon Product and Business Structure (per 2025 Annual Report)

Data sources: Cambricon 2025 annual report (compiled by China Securities Journal and China Economic Net); the schematic is drawn based on this document's analysis.

4.2. Niche: Underlying Compute Supplier under Big-Tech Self-Built Upper-Layer Ecosystems

Cambricon's architectural choice forms a sharp contrast with Huawei:

1. Huawei path: full-stack in-house operation of chip + framework + inference engine + industry solutions, with deep ecosystem reach and heavy investment;
2. Cambricon path: primarily chip + Neuware basic software platform, with the upper layer built by customers — efficient at serving leading internet customers, but making its technology narrative hard to communicate outward (Neuware public documentation is scarce).

This niche determines that its visible capabilities across the six Harness layers are concentrated at the bottom two to three layers.

5. Harness Design

5.1. Six-Layer Capability Overview

Methodology note: The "Weak (opaque)" ratings above reflect the absence of publicly verifiable mechanisms, not the absence of actual product capability — the orchestration and context capabilities actually running on the customer side (big tech) do not pass through public channels and fall within the Harness-invisible zone.

5.2. L1 / L2 Context and Tool Layers

L2 is the layer where Cambricon's visible capability is most concentrated: MagicMind provides quantization, fusion, and cross-framework (PyTorch/TensorFlow) inference support, serving the execution efficiency of customer-side inference services. L1 (KV Cache management, context compression) has no disclosed public mechanism.

5.3. L3 / L4 Orchestration and State Layers

No public mechanism. Industry analysis view (B/C-grade) notes that big-tech customers build their own scheduling and orchestration systems — meaning L3/L4 capability on Cambricon's stack exists in the form of "customer-proprietary implementations" and does not constitute an assessable public product capability.

5.4. L5 Evaluation and Observability Layer

The Neuware platform includes cluster operations components (per industry-article accounts), but the specific form and indicator system of the monitoring and scheduling components have no public documentation (gap). By contrast, Moores Threads in the same group has publicly disclosed its KUAE Platform/ModelStudio dual-platform layering (see article 05); Cambricon is clearly more conservative in its public disclosure.

5.5. L6 Governance and Security Layer

Domestic substitution compliance is one of the core drivers of Cambricon orders (in the operator, financial, and government-enterprise centralized-procurement context). It should be noted: references to specific policy document names were not verified against the original text, and this document does not cite specific policy provisions (see the Information Gap Statement).

6. Actual Cases

6.1. Operator Centralized Procurement (per Industry-Article Relays)

The order information below comes from industry media relays (B/C-grade) and is not per the company's official announcements; it must be flagged when cited:

1. China Mobile single purchase: 7,994 Siyuan 590 accelerator cards, amount ¥430 million (2025-04, per industry-article relay);
2. China Mobile AI server centralized procurement: won approximately ¥2.4 billion in share (42%) in 2025-08 (per media account);
3. Harbin project: 18,000 Siyuan 590 cards integrated by Huawei into Atlas 800/900 servers (Cambricon as the card supplier, per industry-article account).

6.2. Customer Concentration and Customer Structure

Per annual report (A/B-grade, directly citable): sales to the top five customers totaled ¥5.76 billion, or 88.66%; the company's products are deployed at scale across multiple key industries such as operators, financial institutions, and internet (qualitative annual report wording).

Per media accounts (not confirmed by the company, weak wording): According to public reports, in 2025 core customer ByteDance accounted for over 90% of purchases, and purchased approximately 40–50 thousand Siyuan 590 chips in 2025 Q3; there are reports of ICBC purchasing 50,000 Siyuan 590 chips under a ¥3.25 billion contract — all of the above are media estimates; Cambricon has never confirmed any specific customer list or amounts in its announcements.

Risk dimension (per annual report and industry-article accounts): big customers developing in-house chips (ByteDance in-house development rumors, Alibaba iterating) may impact orders; R&D spending in absolute terms is low (only surpassing ¥1 billion in 2024, ¥1.169 billion in 2025); mismatch between capacity ramp-up and shipment recognition pace (analysts estimate shipments at 3–4 times the company-disclosed figure, B/C-grade, to be treated only as a risk clue).

6.3. Handling Discipline for Media-Source Information

This page follows the three disciplines below, for readers to identify information quality:

1. Revenue, profit, production-sales-inventory, customer concentration: all use the 2025 annual report basis (A/B-grade);
2. Customer lists and order amounts: with no official company disclosure, all are downgraded to "per public reports (not confirmed by the company)" and marked [To be verified];
3. Performance benchmarking (e.g., "Siyuan 590 benchmarks against A100/H100, overall performance about 80% of contemporary international products, price about 1/3"): exists only in self-media and AI-generated Q&A (D-grade), and the two third-party accounts are mutually inconsistent (the other account: FP16 256 TOPS, ResNet-50 single-card throughput about 70% of A100, cost 60%) — this page does not adopt these figures; it only states here that they exist and are unusable.

7. Summary

Strengths:

1. Financial validation complete: 2025 revenue of ¥6.497 billion (+453%), first annual profit and dividend, the first complete report card for domestic AI chip commercialization;
2. Cloud focus: 99.69% of revenue came from the cloud product line, delivering supply capability during the compute-demand surge (sales of 117,400 sets);
3. A-share benchmark status: after the "U" designation was removed, it became one of the first batch of "layer-down" enterprises on the STAR Market growth tier, with an edge in financing and talent attraction.

Weaknesses:

1. Insufficient public disclosure: chip spec sheets, software documentation, and customer lists all have no official disclosure, making external verification difficult;
2. High customer concentration: top five customers at 88.66%, and core customers carry potential substitution risk from in-house chips;
3. Inventory risk: ¥4.944 billion in inventory (+178.67%) is sensitive to demand fluctuations;
4. Thin visible Harness capability: of the six layers, only L2 has public mechanisms, creating a gap versus Huawei's full-stack path.

Applicable boundaries: leading internet customers with in-house AI infrastructure capability that need only underlying compute; operator and government-enterprise domestic-substitution centralized-procurement scenarios; not suitable for customers needing complete software stacks and industry-solution delivery capabilities (such needs should turn to article 03 on Huawei Ascend).

Selection recommendations: use the annual report's production-sales-inventory figures and operator centralized-procurement announcements as objective anchors for supply capability; require the vendor to provide test reports of MagicMind with the target frameworks (PyTorch/TensorFlow) on target models, as a substitute for the missing public benchmarks; incorporate customer-concentration risk into supply-chain continuity assessment.

Information Gap Statement

1. Cambricon has never officially disclosed its customer list or order amounts; ByteDance, ICBC, and others mentioned in the text are all per public reports (not confirmed by the company) and have been uniformly downgraded;
2. Official spec sheets for Siyuan 590/690 are missing (not listed on the official website), and this document cites no third-party self-media parameters;
3. Public documentation and version history of the Neuware software stack are missing;
4. Original texts of policy documents on domestic-substitution procurement ratios were not verified, and this document cites no specific policy provisions;
5. For the original 2025 annual report (via CNINFO), a secondary verification is recommended before citation: search for "Cambricon" (寒武纪) at http://www.cninfo.com.cn.

8. References

1. Cambricon official website (product lines) — Cambricon Technologies, 2025. https://cambricon.com
2. Key points of Cambricon 2025 annual report (revenue ¥6.497 billion, net profit ¥2.059 billion, revenue structure by product) — China Securities Journal · CS Smart Finance, 2026-03-13. <https://newzzbcx.cs.com.cn/cxnews.html?name=new20260313185352mwlpcfpa&random=dp7iRuOp>
3. Cambricon 2025 revenue up 453%, with inventory surge and customer concentration concerns — China Economic Net, 2026-03. https://www.ce.cn/xwzx/gnsz/gdxw/202603/t20260313_2825102.shtml
4. Cambricon 2025 net profit of ¥2.059 billion turns around for the first time, plans a ¥632 million dividend — Huanqiu.com Finance (republished via Toutiao), 2026-03. https://www.toutiao.com/article/7616576194573648426/
5. Industry｜Cambricon, Muxi, and Moore report earnings together; the profits/losses and divergence of domestic AI chips — Tencent SOSO industry article (B/C-grade), 2026. <https://so.html5.qq.com/page/real/search_news?docid=70000021_65269b2b38e61052&faker=1>
6. "Hanwang" is no longer "sexy" (customer structure and ecosystem analysis, C-grade) — CNYES Hao, 2026. https://hao.cnyes.com/post/234380
7. Analysis of the reasons behind Cambricon's rally (Lantern Finance AI-generated Q&A, D-grade, for accounts comparison only) — Lantern Finance, 2025. https://www.lanjinger.com/answer/44846
8. CNINFO (search entry for the original Cambricon 2025 annual report) — Shenzhen Securities Information Co. http://www.cninfo.com.cn