China Securities Co., Ltd.: On-device AI is expected to become the next wave of AI, gradually penetrating fragmented scenarios.

China Securities Co., Ltd. released a research report stating that computing power is currently mainly deployed on the cloud side. The firm believes that the device side can not only serve as the entry point to the cloud side, but also provide differentiated value in privacy protection, instant response, low cost, and proactive response, thereby becoming the next wave of AI. From a scenario perspective, on-device AI can not only be applied to large consumer terminals such as smartphones, PCs, and glasses, but its value lies more in penetration across multiple fragmented scenarios. This is reflected in how on-device AI will empower the upgrading of traditional mature products, while also driving the birth of entirely new terminals. Finally, from a commercialization perspective, the upgrading and iteration of mature products can be implemented faster, and some products have already begun to ramp up. New AI terminals, however, require longer development cycles and need deep integration of models, scenarios, and computing power, focusing on core user pain points to build system-level capabilities. The last mile for on-device AI from "general capabilities" to "industry value" still requires collaboration between model vendors and terminal makers, and will also be gradually implemented in the future. The firm believes that the gradual penetration of these fragmented scenarios is one of the core reasons why related leading domestic companies can sustain rapid earnings growth. The main views of China Securities Co., Ltd. are as follows: The value of on-device AI: entry point, proactive response, privacy, low latency, and cost Computing power is currently mainly deployed on the cloud side. The firm believes that the device side can not only serve as the entry point to the cloud side, but also provide differentiated value in privacy protection, instant response, low cost, and personalized proactive interaction. First, on-device AI can serve as the nerve endings for AI entering the physical world. Locally, the device side is mainly responsible for perception, wake-up, and first-layer judgment, and after desensitizing tasks that require deep thinking, hands them over to cloud large models. Second, on-device AI deeply integrates multi-dimensional real-time data, broadening the coverage boundary and experience depth of personalized services, and driving the service logic from "passive response" to "proactive prediction." Cloud AI cannot obtain and analyze comprehensive local user data, so on-device AI has unique advantages in realizing "proactive AI" and deeply understanding user needs. In addition, for scenarios such as industrial control and autonomous driving, latency in cloud computing power directly affects interaction smoothness, task continuity, and even safety boundaries, while the device side can respond instantly through localized deployment. At the same time, for scenarios such as financial risk control, medical assistance, and enterprise office work, on-device processing can significantly reduce the number of times sensitive data leaves the device or organizational boundary, providing strong privacy protection value. Finally, compared with the cloud, on-device AI has a cost advantage in high-frequency task scenarios with high data processing demand or large inference request volumes. Especially in long context, multi-turn dialogue, multi-tool invocation, and multi-agent collaboration, the token consumption of a single task may multiply, while on-device AI adopts a hardware cost model, making long-term costs basically determined. Scenarios for on-device AI: AI empowers all industries, AIoT is colorful and diverse On-device AI can not only be applied to large consumer terminals such as smartphones, PCs, and glasses, but its value lies more in penetration across multiple fragmented scenarios. On the one hand, on-device AI will empower the upgrading of traditional mature products, such as smart home devices, cleaning Siasun Robot&Automation, and large screens. These mature products will transition and upgrade from past automation and digitalization to intelligence and AI. The implementation of on-device AI in these scenarios will also be faster, and some products have already begun to ramp up. On the other hand, on-device AI will also drive the birth of entirely new terminals, such as various AIBOX, service Siasun Robot&Automation, AINAS, and industrial quality inspection. New AI terminals require longer development cycles and need deep integration of models, scenarios, and computing power, focusing on core user pain points to build system-level capabilities, and will likewise be gradually implemented in the future. Finally, with the development and maturity of Personal Agent and the launch of open-source hardware development kits such as Muse Gadgets, on-device AI is expected to become the physical carrier of cloud AI, and lower-cost peripheral hardware is expected to grow rapidly. The firm believes that the gradual penetration of these fragmented scenarios is one of the core reasons why related leading domestic companies can sustain rapid earnings growth. Implementation of on-device AI: models, scenarios, and computing power require deep integration; system-level capabilities are key to large-scale implementation Unlike cloud large models, which mainly revolve around computing clusters and model services, on-device AI imposes higher requirements on software-hardware adaptation, power consumption control, response latency, system integration, and scenario understanding, requiring collaboration among models, terminals, and chips. Its large-scale implementation cannot be completed independently by a single link. Looking back at the development of cloud AI, improvements in model capabilities played a key role. The firm believes the same is true for the device side. With the improvement of capability density and lightweight development of on-device models, the capabilities of general on-device models continue to break through. For example, the intelligence level of small models at the 30B parameter level is rapidly approaching that of hundred-billion-level large models, and 3-7B small models can also complete medium-difficulty tasks such as Q&A, document processing, and visual speech recognition locally. However, for industry scenarios, the last mile for on-device AI from "general capabilities" to "industry value" still requires collaboration between model vendors and terminals. In addition, computing power is also an important part of the on-device experience. For a usable on-device computing chip, the nominal computing power is only the starting point. In actual use, toolchains, memory bandwidth, and real-time power consumption are also key factors affecting inference experience. To meet the differentiated deployment requirements of on-device AI across terminals, leading domestic SoC companies are all accelerating product iteration while continuously improving toolchains and ecosystems. For example, Rockchip Electronics launched the RK182X coprocessor, adopting an innovative 3D stacking architecture to break through memory bandwidth limitations. The firm believes that in the AIoT field, leading domestic SoCs are globally competitive, while leading overseas manufacturers focus more on cloud AI and chips for core consumer terminals such as smartphones and PCs. With accelerated product iteration, these leading domestic SoCs will fully benefit from the on-device AI wave and maintain rapid growth. Investment recommendation The firm is optimistic about the gradual implementation of on-device AI across all industries in the future, focuses on the SoC industry chain, and highlights the following recommended targets: 1) Rockchip Electronics: As a leader in AIoT SoC, the company has built full-stack capabilities of "chip + algorithm + industry solution." Its software tools and ecosystem accumulated across industries over the past decade are very deep. In particular, its AI coprocessor breaks through memory bandwidth limitations through a 3D stacking architecture. At present, hundreds of customers and dozens of industries are deepening R&D, and the first batch of customers in cleaning/companion Siasun Robot&Automation, NAS, and smart home have entered the product launch and mass production stage. 2) Espressif Systems: Centering on the product strategy of "processing + connectivity," the company continues to iterate AIoT chips and other products. The ramp-up of new products such as S31 and P4 is expected to drive accelerated revenue growth. Emerging downstream businesses such as industrial control, smart agriculture, and the new developer market are growing rapidly. At the same time, the company attaches great importance to product ecosystem development and, through ESP-Claw, ESP-Vision, and support for cutting-edge AI development technologies such as MCP, further lowers the threshold for combining AI with physical devices, thereby attracting more developers and expanding product application areas. 3) Amlogic: Based on TV SoC and set-top box SoC, the company is actively building a general-purpose AIoT computing power platform and has launched the A9 high-computing-power 6nm on-device AI chip. Starting from its accumulated major TV screen customers, it has advanced into fields such as Siasun Robot&Automation, edge computing, and smart cockpits. New AIoT products have reached cooperation with leading major customers and are expected to enter mass production in the second half of 2026, which is expected to become a long-term growth engine. 4) Bestechnic (Shanghai) Co., Ltd.: The company is a globally leading designer of low-power wireless computing SoC chips. In the first half of 2026, affected by rising memory prices and the phasing out of national subsidy policies, downstream terminal market sentiment declined, and the company's operations came under pressure. The firm believes that the company's inflection point is imminent. The company's new-generation flagship BES 6100 completed tape-out in the first half of 2026 and is expected to enter the testing and customer sampling stage in the second half of the year, which is expected to help the company start a new growth cycle. Risk analysis Risk of AI terminal demand falling short of expectations; risk of on-device model development falling short of expectations; risk of technological iteration and route changes.