The Great Wall of Computing: Exploring the Sustainable Moat of the Edge AI Chip Industry in A-Share Market

The Great Wall of Computing: Exploring the Sustainable Moat of the Edge AI Chip Industry in A-Share Market

Disclaimer: This article does not constitute any investment advice.

Core Summary

The Chinese edge artificial intelligence (Edge-AI) inference ASIC accelerator card industry is experiencing a golden development period driven by both policy and real demand. The “domestic substitution” strategy, spurred by geopolitical factors, has created a protected market space for local companies, while the diverse demands from smart vehicles, industrial IoT, and smart cities provide growth opportunities for companies with varying technological paths. However, policy protection is not a long-term solution; true sustainable competitiveness must be built on technological and ecological advantages. This report believes that the most defensible market positions stem from two types of companies: platform companies focused on building high-switching-cost software ecosystems, and “shovel sellers” that dominate in high-speed interconnect technology. In the next 3-5 years, the industry will differentiate based on ecological building capabilities and foundational technology control.

Industry Framework and Market Landscape

Market Size and Structure

The Chinese edge AI market is in the early stages of rapid expansion, expected to grow from $1.629 billion in 2024 to $6.217 billion in 2030, with a compound annual growth rate (CAGR) of 25.3%, significantly higher than the global market’s 17.6%. Hardware is absolutely dominant, accounting for 66.18% of market revenue in 2024, highlighting the critical position of ASIC accelerator cards. Market demand is highly fragmented, with consumer electronics pursuing ultra-low power consumption of 1-3W, while industrial automation and autonomous driving require more powerful computing capabilities exceeding 10W, creating niche market opportunities for companies with specific domain advantages.

Core Driving Factors

·Domestic Substitution Policy: U.S. export controls have caused NVIDIA’s share of high-end AI chips in China to plummet from over 90% to about 50%, creating a significant market vacuum. National projects such as the “East Data West Computing” initiative (with over $28 billion in investment) systematically create a huge demand for domestic AI hardware.

·Technological Upgrade Demand: The scale of AI models is growing exponentially, shifting the computing bottleneck from internal chip computation to data transmission efficiency, increasing the demand for high-speed interconnect technologies such as PCIe Retimer and CXL.

·Application Scenario Explosion: Smart vehicles (Level 5 autonomous driving requires 32 sensors), industrial IoT (predictive maintenance, visual inspection), smart cities (real-time video analysis), and consumer electronics (edge AI) collectively drive the demand for edge computing power.

Core Competitors Deep Dive

Cambricon (688256.SH) – Ecological Builder

Company Overview and Strategy: Originating from the Institute of Computing Technology, Chinese Academy of Sciences, a leading domestic AI chip company, providing integrated “cloud-edge-end” full-stack solutions, with core products being the “Siyuan” series chips and accelerator cards, strategically shifting focus to the cloud market and fully building its own software ecosystem “NeuWare”.

Financial Health and Performance:

·Revenue: In 2024, cloud product revenue is expected to increase by 1187% year-on-year, with total revenue in Q1 2025 increasing by over 42 times year-on-year, showing explosive growth.

·Profitability: Long-term high R&D investment (157.53% R&D expense ratio in 2023) has led to continuous losses, but Q4 2024 and Q1 2025 are expected to achieve quarterly profitability for the first time, indicating operational leverage. Gross margin increased to 69% in 2023, maintaining at 56% in Q1 2025.

·Balance Sheet: Ample cash flow post-IPO, but inventory reached 2.755 billion yuan and prepayments 973 million yuan (as of Q1 2025), revealing dependence on large project orders.

Moat Depth Assessment:

·Intangible Assets (Narrow): Self-developed chip architecture MLUarch03 and the under-construction NeuWare ecosystem, with developer scale and application richness far from CUDA.

·Switching Costs (Narrow): The largest customer in 2024 contributed 79% of revenue, forming deep binding, but switching costs for a broad customer base have not yet been established.

·Network Effects (None): Small developer community, no self-reinforcing cycle formed.

·Cost Advantage (None): Fabless model, using TSMC’s 7nm process, no cost advantage.

·Effective Scale (None): Intense market competition, not yet reaching a limited scale state.

Growth Prospects and Risks: High growth potential, benefiting from domestic substitution high-end computing demand. Core risks include extremely high customer concentration, large revenue volatility, reliance on non-sustainable government projects, and geopolitical risks in obtaining advanced processes.

Hygon Information (688041.SH) – Successor of the X86 Ecosystem

Company Overview and Strategy: High-end processor designer, with two major product lines: x86 architecture CPUs and GPGPU-based DCUs (Deep Computing Units). Core advantages stem from technical cooperation with AMD, with DCUs adopting a “CUDA-like” ecosystem, lowering market entry barriers.

Financial Health and Performance:

·Revenue: 2023 revenue of 6.012 billion yuan (+17.3%), with Q1 2024 increasing by 37.09%, showing steady growth.

·Profitability: Continuous profitability, with 2023 net profit attributable to shareholders of 1.263 billion yuan (+57.17%), R&D expense ratio of 33.14%, well-managed. Forecasted ROE could reach 17.7% in the next three years.

·Balance Sheet: Healthy, with good cash flow and controllable liabilities.

Moat Depth Assessment:

·Intangible Assets (Wide): x86 architecture licensing is a rare intangible asset, creating high barriers in the server CPU market; DCU’s “CUDA-like” compatibility solves the ecological cold start problem.

·Switching Costs (Narrow): x86 ecosystem binds customers, with low migration costs for DCUs but increasing with usage deepening.

·Network Effects (Narrow): Benefiting from the existing large software networks of x86 and CUDA, with first-mover advantages.

·Cost Advantage (None): Fabless model with no cost advantage.

·Effective Scale (None): Broad market space with intense competition.

Growth Prospects and Risks: High growth potential, with DCUs as the growth engine and CPUs providing a stable base. Core risks include geopolitical risks of technology licensing agreements and customer concentration (the top five customers accounted for 91.23% before 2021).

Lanqi Technology (688008.SH) – “Shovel Seller” of High-Speed Interconnect

Company Overview and Strategy: A global oligarch in memory interface chips (RCD, DB), strategically positioned as a “toll booth” for high-speed data transmission paths, expanding into AI-related interconnect technologies such as PCIe Retimer and CXL, collaborating with Intel on the Jinda® CPU.

Financial Health and Performance:

·Revenue: In 2023, revenue decreased by 37.76% due to cyclical impacts, but increased by 75.74% in Q1 2024, with rapid growth in PCIe Retimer orders.

·Profitability: Historically high profitability and gross margin, with Q1 2024 net profit attributable to shareholders increasing by over 1000%, showing rapid recovery.

·Balance Sheet: Strong, with abundant cash reserves and very low liabilities.

Moat Depth Assessment:

·Intangible Assets (Wide): One of the standard setters for JEDEC, with patents and self-developed SerDes IP, creating high technical barriers.

·Switching Costs (Wide): Products embedded in CPU/server platforms, switching suppliers requires re-validation, with extremely high costs.

·Network Effects (Narrow): As a standard setter, ecological partners design around its standards, solidifying its core position.

·Cost Advantage (None): Fabless model.

·Effective Scale (Narrow): The memory interface chip market is oligopolistic, with high barriers preventing new entrants.

Growth Prospects and Risks: High growth potential, driven by the upgrade from DDR4 to DDR5, AI server demand (each requiring 8-16 PCIe Retimers), and new CXL technology. Risks are related to performance and the cyclical nature of the server/memory market, with the Jinda® CPU business dependent on Intel collaboration.

Other Participants Overview

·SoC Suppliers (Allwinner Technology, Rockchip) : SoC chips integrated with NPU primarily target consumer-grade edge devices, with moats built on specific customer relationships, scene understanding, and cost performance, not direct competitors to high-performance accelerator cards, but important roles in the low-power market.

·Upstream Enablers (SMIC, Huahong, Northern Huachuang, Zhongwei) : The cornerstone of “domestic substitution”, with moats derived from high capital investment and complex processes creating “effective scale”, their success is a prerequisite for the development of downstream fabless companies.

Moat and Financial Comparison Scorecard

Company Name	Stock Code	Moat Rating	Growth Potential	Core Competitive Advantage	Financial Health	Risk Level
Cambricon	688256.SH	Narrow (Widening)	High	Only company dedicated to building a unified “chip + software” ecosystem (NeuWare)	Revenue explosion but high volatility, just achieved quarterly profitability	Very High (Customer Concentration, Process Acquisition)
Hygon Information	688041.SH	Narrow (Stable)	High	x86 licensing + “CUDA-like” ecosystem, dual-driven	Continuous profitability, steady growth	High (Technology Licensing Risk)
Lanqi Technology	688008.SH	Wide	High	Global oligarch in memory interfaces, leader in high-speed interconnect technology	High profitability, abundant cash flow	Medium (Cyclical)
Jingjia Micro	300474.SZ	Narrow	Medium	Domestic GPU leader, first-mover advantage in high-reliability markets	–	–
Xilinx	688521.SH	Narrow	Medium	Leading semiconductor IP and chip customization services, widely used NPU IP	–	–
Allwinner Technology/Rockchip	300458.SZ/603893.SH	None	Medium	Smart SoC leaders, cost advantages in specific terminal markets	–	–

Risk Analysis

1.Ecological Building Risk: The long-term value of Cambricon and Hygon Information depends on the effectiveness of software ecosystem (NeuWare/”CUDA-like”) construction. Key observation indicators include developer community activity, mainstream framework adaptation progress, and actual performance.

2.Customer Concentration Risk: Cambricon (with the largest customer accounting for 79% in 2024) and Hygon Information (with the top five customers accounting for 91.23% before 2021) have performance highly dependent on a few large customers (mostly related to government projects), requiring identification of revenue sustainability (one-time projects vs. sustainable contracts).

3.Supply Chain and Geopolitical Risks: The fabless model heavily relies on foundries like SMIC and Huahong. U.S. sanctions restrict access to advanced lithography equipment, making local wafer fabs’ capacity and technology a bottleneck for the entire domestic AI chip ecosystem. Technology licensing (like Hygon) also carries uncertainties.

4.Technology Iteration Risk: Rapid iteration of AI algorithms (such as the new DeepSeek model) may render hardware optimized for older models obsolete, requiring companies to possess architectural flexibility and rapid adaptability.

Competitive Landscape Summary and Competitiveness Ranking

Based on moat width, ecological niche advantages, technological barriers, and financial stability, core companies are ranked for competitiveness:

·Leaders:

oLanqi Technology (688008.SH): Outstanding competitiveness. Wide moat that is hard to disrupt, its “shovel seller” model benefits directly and defensively from the explosion of AI computing power, financially healthy, with high growth certainty.

·Challengers:

oCambricon (688256.SH): Huge potential but risks coexist. The ecological building strategy is high-risk, high-reward, with recent financial turning points showing strong execution, but the moat is not yet solidified, highly dependent on policies and projects, requiring continuous observation of ecological construction effectiveness.

oHygon Information (688041.SH): Clear advantages. The x86 ecosystem inheritance and “CUDA-like” strategy are pragmatic and effective, with dual business lines providing stable cash flow and growth engines, but technology licensing is its lifeline, with geopolitical risks.

·Niche Players:

oJingjia Micro (300474.SZ), Xilinx (688521.SH): Have advantages in specific technology fields (GPU/IP licensing), benefiting from the overall prosperity of the industry, but the moat is relatively narrow, with moderate growth potential.

oAllwinner Technology (300458.SZ), Rockchip (603893.SH): Solid position in the consumer-grade edge SoC market, but facing intense competition, with shallow moats, growth more reliant on industry beta.

A truly sustainable moat will ultimately stem from technological depth and ecological breadth, rather than temporary policy protection. The future differentiation is already in sight.