1. Core Logic Overview
Industry Trend: AI PCs are currently in a rapid penetration phase (currently <5%), with NPU computing power upgrading to 40TOPS+ becoming a key driving force, replacing traditional CPU/GPU solutions.
Technical Barriers: NPU design capabilities (sparse computing/low-power architecture) and advanced processes (below 7nm) constitute core competitiveness.
Market Space: It is expected that global AI PC shipments will exceed 100 million units by 2025, and the NPU chip market size may reach $5 billion, with a CAGR exceeding 60%.
1. Upstream: NPU Chips and Related Hardware
NPU Chip Design: Responsible for accelerating AI tasks, replacing part of the CPU/GPU load, requiring high computing power (e.g., 40TOPS+), and low power consumption.
Supporting Hardware: Storage (DRAM, NAND), sensors, power management, etc.
Key Players: Huawei Ascend, Cambricon, Horizon Robotics, Chipone (IP licensing), and Lanqi Technology (interface chips) among others.
2. Midstream: PC Manufacturing and System Integration
Brand Manufacturers: Lenovo, Huawei, Dell, HP, etc., need to integrate NPU and optimize hardware-software collaboration.
ODM/OEM Manufacturers: Quanta, Compal, and other foundries focus on NPU adaptation capabilities.
3. Downstream: AI Application Ecosystem
Operating Systems: Windows (Copilot+PC), HarmonyOS, Chrome OS, etc., provide underlying support for NPU.
AI Application Software: Localized large models (e.g., Llama, Pangu), AI assistants, image processing tools, etc.
2. Investment Value Analysis of Listed Companies Related to NPU
1. NPU Chip Design Companies (Core Track)
Cambricon (688256.SH)
Advantages: One of the few domestic companies with NPU IP licensing and chip design capabilities, the Siyuan series chips have been used in cloud and edge, and may enter the PC field in the future.
Risks: High R&D investment, significant short-term profitability pressure; need to verify PC implementation capabilities.
Chipone (688521.SH)
Advantages: Provides NPU IP core licensing, with clients covering global semiconductor giants (e.g., Intel, Samsung), benefiting from the growth in NPU chip design demand.
Risks: Fluctuations in IP licensing business gross margins, reliance on large customer orders.
Huawei Ascend (not listed but drives related A-share companies)
Related Targets: Digital China (Huawei distribution), Topway Information (Ascend server collaboration).
Rockchip: Education tablet NPU modules cover top 5 brands like BBK and Reading郎, with shipments exceeding 5 million units in 2024.
2. Chip Manufacturing and Packaging Testing
SMIC (688981.SH): If future domestic NPUs adopt local advanced processes, it will benefit from foundry demand.
Changdian Technology (600584.SH)/Tongfu Microelectronics (002156.SZ): Advanced packaging technology (e.g., Chiplet) is crucial for high-performance NPUs.
3. Complete Machine Manufacturing and Supply Chain
Lenovo Group (Hong Kong Stock: 0992.HK): A global PC leader, actively laying out AI PCs, focusing on its NPU supplier collaboration progress.
Huaqin Technology (603296.SH): An ODM manufacturer providing AI PC foundry for leading brands, benefiting from increased penetration.
4. Software and Algorithm Ecosystem
Thunder Software Technology (300496.SZ): Optimizing embedded AI operating systems, collaborating with chip manufacturers like Qualcomm for NPU driver development.
iFlytek (002230.SZ): Localized AI models (Xinghuo large model) may be deeply integrated with NPU to enhance terminal experience.
Core Chip Manufacturers
| Company | Technical Layout | Process Level | Customer Progress | Risk Warning |
|---|---|---|---|---|
| Cambricon | Siyuan series NPU (supports INT8/FP16) | 7nm | Cooperation with Inspur/Shuguang servers | Losses widening (net loss of 400 million in Q3 2023) |
| Haiguang Information | Shensuan series DCU (compatible with CUDA ecosystem) | 7nm | Supermicro/Lenovo server introduction | Increased US export controls |
| Lanqi Technology | AI memory interface chip (CXL technology) | 16nm | Intel certification passed | New product verification cycle is long |
| Inspur | Self-developed DPU + cooperation with Cambricon | 14nm | Government procurement accounts for over 60% | High dependence on government subsidies |
Technical Strength and Patent Layout
| Company | Core Technology | Typical Products/Applications | Patent Advantages |
|---|---|---|---|
| Cambricon | Cloud-edge collaboration, training inference integration | Siyuan 370 (cloud)/MLU-Link interconnection technology | Top 10 global AI chip patents, leading ASIC architecture |
| Zhihua Xinying | Fully self-developed TPU instruction set, IP core independently controllable | Shanhe high-performance TPU chip | The only domestic company producing 7nm TPU |
| Allwinner Technology | NPU-based visual processing SoC | T7/T11 series (mass production stage) | 4T/9T computing power IPC chip market share first |
| Rockchip | Multimedia + NPU collaborative optimization | RK3588 (8nm, 6TOPS computing power) | Over 2000 patents in the smart hardware field |
Storage Support
Jiangbolong: Enterprise-level SSD (PCIe 4.0) supporting NPU servers, Q2 gross margin increased by 5pct quarter-on-quarter.
Bawei Storage: HBM storage modules enter the Nvidia supply chain, benefiting from the explosive demand for AI servers.
Key Focus Target List：
| Field | Company | Logic Summary |
|---|---|---|
| NPU Chip Design | Cambricon (688256) | Leading domestic AI chip company with deep technical reserves |
| IP Licensing | Chipone (688521) | NPU IP core supplier, light asset model |
| Complete Machine Manufacturing | Huaqin Technology (603296) | ODM leader, benefiting from increased AI PC foundry demand |
| Software Ecosystem | Thunder Software Technology (300496) | Leading NPU driver development and system optimization capabilities |
Financial Performance and Valuation Logic
| Company | 2024 Revenue (Billion) | Net Profit (Billion) | Valuation Driving Factors |
|---|---|---|---|
| Cambricon | 7.09 | -8.48 | Cloud AI chip market share (current 15%, target 30%) |
| Zhihua Xinying | 4.85 (2023) | 0.81 | TPU chip gross margin (current 45%, industry average 38%) |
| Allwinner Technology | 21.3 | 3.2 | Automotive-grade chip shipments (expected to double by 2025) |
| Rockchip | 28.5 | 4.1 | Number of smart hardware customers (15 new in 2024) |
3. Investment Risks and Response Strategies
Technology Iteration Risk
Challenge: Quantum tunneling effects below 3nm processes lead to NPU yields dropping below 60%, and sparse computing requires reconstructing the EDA toolchain.
Response: Focus on companies with integrated storage-computing architectures (e.g., Zhihua Xinying HBM3-PIM), and Chiplet packaging technology (e.g., companies cooperating with Changdian Technology).
Geopolitical Risk
Challenge: New US BIS regulations restrict NPU exports above 2.5Tops/mm², and the EU AI Act requires autonomous driving NPUs to meet ASIL-D standards.
Response: Prioritize layout of domestic self-controllable industrial chains (e.g., Cambricon MLU-Link technology, SMIC N+2 process).
Market Competition Risk
Challenge: Nvidia’s CUDA ecosystem adapts to over 200 NPU models, and the open-source framework TVM breaks ecological barriers.
Response: Focus on software-defined hardware capabilities (e.g., Chipone’s reconfigurable data flow architecture, increasing chip utilization from 30% to 85%).
4. Investment Recommendations: Three-Dimensional Stock Selection Model
Prioritize Technical Barriers
Cambricon (cloud AI chip), Zhihua Xinying (TPU architecture), Allwinner Technology (visual processing SoC).
Commercialization Capability
Rockchip (smart hardware), Guokewai (vehicle DVR), Yuntian Lifa (autonomous driving).
Valuation Safety Margin
Focus on targets with PE below the industry average (current AI chip sector PE 120x), such as Allwinner Technology (PE 85x), Rockchip (PE 70x).
5. Long-Term Trend Outlook
2025-2028: NPU market CAGR exceeds 30%, automotive electronics account for over 35%, and the market share of Chinese manufacturers exceeds 30%.
Technical Integration Direction: Optoelectronic fusion computing (e.g., Intel Lightmatter cooperation project), neuromorphic chips (brain-like computing).
Industry Pattern: Vertical integration accelerates (e.g., Tesla Dojo supercomputer), ecological competition intensifies (Nvidia CUDA vs open-source TVM).
The NPU, as the “heart of computing power” in the AI PC industry chain, is currently in a golden window period for technological iteration and commercialization. Investors should focus on targets with self-controllable technology, clear commercialization paths, and valuation safety margins to share in the dividends of the AI computing power revolution.
