In-Depth Analysis of the Full Industry Chain of Edge Computing Power (2025 Edition)

Conclusion in one sentence: “Computing power is not only in the data center but also in your pocket and on your desk.” The explosion of edge SoCs (including NPUs) comes from three forces: Platform Barriers (Apple/Google/Microsoft embedding functional tickets into chip specifications), User Demand (privacy/low latency/offline), and Cost Optimization (cloud inference is too expensive). This is followed by the multi-terminal penetration of AI smartphones/AI PCs/AI wearables, and the A-share industry chain from upstream packaging and testing to SoC design is currently in a structural opportunity period.

Quick Overview of Conclusions

• Technical Route: Edge SoCs adopt a heterogeneous integration of CPU (mostly Arm/RISC-V) + GPU/DSP + NPU + ISP + connectivity, combined with ONNX Runtime / Core ML / Android GenAI APIs (Gemini Nano) and other software stacks; the industry focuses more on effective speed/power consumption/memory bandwidth, rather than just looking at TOPS.
• Demand Driven:
• Platform Barriers: Microsoft’s Copilot+ PC requires ≥40 TOPS NPU; Apple’s Apple Intelligence is only available on A17 Pro / M series and above; Android accelerates Gemini Nano and GenAI APIs for “local intelligence.” These three directly determine the shipment structure and installation volume.
• User Side: privacy/low latency/offline robustness; Vendor Side: decreasing cloud costs and differentiated selling points.
• Market Size and Penetration:
• AI PCs: Expected shipments in 2025 will exceed 100 million units, accounting for about 40%; in 2028, 205 million units.
• AI Smartphones: In 2025, the proportion of “AI capabilities” is expected to exceed 30%+; the total smartphone market in 2025 is approximately 1.24 billion units, with steady quarterly growth.
• A-share Value Chain
• SoC Design: Rockchip/AmLogic/Guokai Micro/Fuhang Micro/Guoxin Technology/Hengxuan/Leixun/Zhongke Lanxun/Juxin/Ankai Micro/Xingchen/Broadcom Integration/XinHai Technology, etc. (varied segments, AI multimedia/low-power wearables/connectivity, etc.).
• Manufacturing/Testing: SMIC (wafer), Changdian Technology/Tongfu Microelectronics/Huatian Technology (advanced packaging, AI-related substrates/FCBGA/SiP).
• Competitive Barriers: IP/instruction set and ecosystem (Arm/RISC-V), edge AI toolchains (Core ML/ONNX/TFLite/vendor SDKs), customer certification and power management capabilities.
• Policies and Standards: The State Council’s “AI +” initiative emphasizes smart terminals (phones/computers/wearables), setting popularization goals for 2027 and 2030; MLCommons MLPerf (Edge/Tiny) is gradually becoming a universal benchmark for energy efficiency/performance.
• 12–24 Month ScenarioBenchmark: AI PCs and AI smartphones will jointly drive the penetration of edge NPUs, resonating with upstream advanced packaging and domestic SoC design; Optimistic: AI glasses/headphones and other new hardware iterations accelerate; Conservative: changes in platform API barriers and macro consumption slowdown.

1. Definition and Technical Route

SoC is “putting a small computer into a chip”: CPU for general computing, GPU/DSP for multimedia and signal processing, NPU specifically for running neural networks, plus ISP (image), baseband/Bluetooth/Wi-Fi, and other modules. The core of edge AI is to make space in this SoC, connect to faster NPUs, and more efficient memory/bandwidth.

• NPU Evolution: From convolution-dominated to supporting Transformer/attention, Arm Ethos-U85 is clearly aimed at edge Transformer; wearable SoCs (like Hengxuan BES2800) have integrated NPU + Wi-Fi/BLE.
• Toolchain: Commonly used on mobile ONNX Runtime (including Mobile/NNAPI EP), Apple Core ML, Android GenAI APIs + Gemini Nano; they serve as the “highway” for “model → device”.
• Platform Barriers (Very Critical):
• Windows Copilot+ PC: New features generally require ≥40 TOPS NPU → directly boosting the demand for the next generation of PC SoCs in 2024/25 (e.g., Snapdragon X series 45TOPS).
• Apple Intelligence: A large number of edge functions are only available on A17Pro / M1+ and above devices.
• Android: Promotes Gemini Nano (local large models), developers access through GenAI API.

2. Demand and Applications: Why Go “Edge”?

• Experience: Low latency (camera/subtitles/call summaries/screen understanding), offline availability (travel/weak networks), privacy (data does not leave the device). Both Apple and Google’s official pages emphasize “local” and “privacy” as core selling points. Cost: Cloud inference is costly per instance, and peak concurrency is expensive; moving common inference to the edge is a clear way to reduce costs and increase efficiency (Microsoft/Google’s localization routes essentially do this).
• “Platform Ticket” Logic: Without 40+ TOPS NPU (PC)/without a new generation of neural engines (iPhone/Mac)/without Nano capabilities (Android), many new features cannot be introduced → shipment structure is passively upgraded.

3. Policies and Standards

• China: Deeply Implementing“AI +” Initiative (2025.8)

• Clearly proposesAI smartphones and computers, smart wearables as “new generation smart terminals,” and sets 2027/2030 popularization rate targets, providing policy backing for the edge ecosystem.
• Industry Benchmark:MLCommons / MLPerf (Edge/Tiny) is gradually becoming a public measure of edge performance and energy efficiency, with new results released in 2025 for version 5.0.
• Platform Specifications: Microsoft has announced the NPU capability thresholds and development guidelines for Copilot+; Android is gradually migrating NNAPI to the new GenAI route; Apple uses Private Cloud Compute + Core ML to define boundaries.

4. Market Size and Penetration (2025E–2028E)

• AI PCs: In 2025, shipments of AI-capable PCs are expected to exceed 100 million units (about 40% market share), and in 2028, 205 million units (24–28 years 44% CAGR).
• Smartphones: In 2025, global shipments are expected to be ~1.24 billion, with low growth; the proportion of “GenAI/AI-capable” smartphones in 2025 is expected to be 30%+, which is a consensus direction among mainstream institutions.
• AI Wearables/New Terminals: Smart glasses are experiencing rapid growth in shipments, with domestic and international brands intensively releasing products; A-share wearable SoC/connectivity chip manufacturers are benefiting (institutional/IDC data metrics).

5. Value Chain and Representative Companies (Including A-shares)

5.1 Global Ecosystem

• IP/Instruction Set: Arm (Ethos series NPU, v9 architecture), RISC-V (gradually in wearables/MCU/camera SoCs).
• Platform Providers:Microsoft Copilot+ PC (≥40 TOPS), Apple Intelligence, Android Gemini Nano.
• SoC/PC: Qualcomm Snapdragon X (45 TOPS), AMD/Intel’s next-generation NPU PCs.

5.2 Key Examples in A-shares (Not Investment Advice)

Company	Main Track/Product	Advantages and Progress (Related to Edge AI)
Rockchip	Multimedia SoC (e.g., RK3588/NPU about 6TOPS)	Strong video encoding/decoding/ISP, rich NPU adaptation solutions; clients in education tablets, robots, etc. are iterating edge AI.
AmLogic	Set-top boxes/multimedia/AIoT SoC	One of the leading high-performance multimedia main control manufacturers, expanding AIoT scenarios (institutions/media have repeatedly included it as a core company in A-share SoCs).
Guokai Micro/Fuhang Micro/Xingchen	Security/image ISP/SoC	Focusing on ISP/video SoCs, adapting edge algorithms, benefiting from “local video understanding.”
Guoxin Technology	General processing/specialized SoC	Domestic route + security/industrial control application penetration.
Hengxuan Technology	Wearable/audio/glasses SoC	BES2800 (6nm, integrating NPU+Wi-Fi/BLE) has been mass-produced, accelerating the landing of AI headphones/watches/glasses.
Leixun Technology	Wi-Fi/BLE SoC (ESP32 series)	ESP32-S3 includes vector instructions/AI acceleration libraries, highly adaptable for edge inference (lightweight models).
Zhongke Lanxun/Juxin/Ankai Micro	Low-power audio/wearable SoC	Upgrades in TWS/watches/toys/learning machines drive demand for “low power + inference.”
SMIC	Wafer foundry	Domestic leader, advanced/mature process collaboration to expand edge SoC production.
Changdian Technology/Tongfu Microelectronics/Huatian Technology	Advanced packaging/testing	Platforms such as SiP, FCBGA, Chiplet/2.5D/3D are improving, benefiting AI terminal SoC/substrate packaging.

6. Competitive Landscape and Barriers

• Ecological Binding: Who adapts faster to the edge AI routes of Apple/Android/Windows (Core ML / GenAI APIs / ONNX Runtime + NPU EP) will find it easier to secure customer mass production.
• Effective Value of Computing Power: Focus on latency/power consumption/bandwidth/memory and quantization (INT8/INT4), rather than just comparing TOPS. The public results of MLPerf Edge/Tiny are becoming a reference for customers.
• IP/Compliance Risk Management: The licensing game between Arm and OEM chip manufacturers remains one of the industry variables (recent litigation results between Arm and Qualcomm remind us that legal and ecological certainty = valuation anchor).
• Supply Chain:Advanced packaging/substrates/yield and Wi-Fi/Bluetooth/storage supply elasticity determine the cost and delivery of edge products.

7. Key Risks

Main Risks

1)Platform strategy changes (e.g., API migration, functional threshold adjustments); 2)Macroeconomic and consumer recovery falling short of expectations; 3) Supply-side (foundry/testing/substrate) and geopolitical/compliance; 4) Misreading TOPS metrics leading to product capability mismatches.

8. Future 12–24 Month Scenarios and Configuration Suggestions (Not Investment Advice)

Benchmark Scenario (Highest Probability)

• AI PCs will continue to grow (in 2025, >100 million units, further increase in 2026), driving ≥40 TOPS NPU to become standard; both Android and Apple will iterate edge capabilities quarterly.

Configuration: ① High-performance multimedia SoCs (mid-to-high-end edge inference and video capabilities); ② Advanced packaging/SiP/FCBGA (with the complexity of AI terminals increasing); ③ Low-power wearables/connectivity SoCs (TWS/watches/glasses).

Optimistic Scenario

• AI glasses/headphones become a true “second growth curve,” with the domestic “AI +” initiative creating real demand in consumer electronics; A-share wearable SoCs and connectivity chips see both volume and price increases. Increase investment in wearable SoCs and connectivity chips (Bluetooth/Wi-Fi).

Conservative Scenario

• Weak macro consumption + changes in platform thresholds/API + supply disruptions; AI functionality usage rates fall below expectations, leading to insufficient upgrade drives. Defensive: Leaders in advanced packaging/multi-client multi-category companies, those with more stable cash flow and R&D intensity.

Stock Selection Key Points

• Technical Strength: NPU architecture/bandwidth/power management, model toolchain adaptation (ONNX/TFLite/Core ML/own SDK).
• Customer Structure: Whether to penetrate AI smartphones/PCs/wearables of first-line brands; multi-regional revenue; new product introduction rhythm.
• Manufacturing and Testing Collaboration: Process nodes/SiP and substrate assurance, yield and delivery.
• Policies & Compliance: Alignment with “AI +” landing scenarios, overseas compliance capabilities.

9. Chip Side “Ticket” and Examples

• A-share wearable SoC Example:Hengxuan BES2800 (6nm, integrating multi-core CPU/GPU/NPU + Wi-Fi/BLE) has been mass-produced, first launched in Samsung Buds3 Pro. Leixun ESP32-S3 provides vector instructions/ESP-NN libraries supporting lightweight inference.

Appendix: SoC Edge Computing Power Sector Risk/Opportunity Comparison Table (2025 Edition)

Company	Core Advantages	Main Application Scenarios	Recent Market Performance
Rockchip	A-share AIoT SoC leader; NPU computing power 0.2–6 TOPs, supporting edge large models; high revenue growth	Smart home, AI glasses, tablets, educational electronics	2024 revenue 3.1 billion, net profit year-on-year +300%; stock price increase of 187% in the past year
AmLogic	Audio and video SoC market share leader; launched 6nm series products; supported by international orders	Smart TVs, set-top boxes, AI video terminals	2024 revenue 5.92 billion, net profit year-on-year +60%; in 2025, sales are expected to exceed 10 million units
Hengxuan Technology	Ultra-low power wireless AI SoC, deeply involved in wearables and Bluetooth audio; clients include Huawei, Xiaomi	Smart watches, headphones, wearable devices	Benefiting from the wearable market explosion, stock price increase of over 100% in the past year
Leixun Technology	Wi-Fi SoC leader; ESP32-S3 is the first Wi-Fi MCU supporting edge AI	Smart home, IoT terminals, smart hardware	A-share Wi-Fi SoC market share is the highest; stock price increase of over 120% in the past year
Zhongke Lanxun	Bluetooth audio SoC leader; partnered with ByteDance Volcano Engine to enter AI glasses	Bluetooth headphones, AR/AI glasses, smart speakers	Stock price increase of over 100% in the past year; frequent new product releases attract funding
Fuhang Micro	Visual processing SoC, leading in security monitoring, with ISP+NPU architecture	Smart cameras, security monitoring, automotive imaging	Security AI upgrades drive performance; stock price steadily rises

Time is the best compound interest!