Intelligent Finance and Economics APP (智通财经APP) learned that Huafu Securities released a research report stating that running AI models at the edge has advantages such as cost, availability, and privacy and security, and the cloud can make up for the insufficient computing power of edge devices. From a historical perspective, the previous AI tasks were also shifted to the terminal after the improvement of edge computing power. Due to the easier exploration of high-value user scenarios, mobile phones and PCs are expected to become the first carriers of edge AI landing, and there is great potential for AI to be applied in XR/earphones/smart automobiles/robots and other fields in the follow-up.

Huafu Securities' main points are as follows:

The industry is thriving, and the entry of giants is leading the ecological development.

Brand end: vigorously promote edge AI, a single device loads a large number of models through the MoE architecture to complete complex tasks. For example, Microsoft's Surface Pro is deeply integrated with more than 40 local large models and GPT-4o and other cloud large models to construct system-level AI capabilities. The current mainstream parameter amount of mobile phones is 7B, which will continue to grow and continuously expand the types of tasks that can be processed.

System manufacturers: Microsoft/Google/Apple have great industry appeal and are key players in the AI ecosystem. They have updated a series of AI functions in their respective systems for developers to use.

Chip manufacturers: The AI performance of newly released products from various manufacturers has been greatly improved, and the layout is aggressive. For example, after Apple's M3 was released, it launched the M4 chip half a year later, and the NPU computing power soared from 18 TOPS to 38 TOPS. Intel's next-generation chip, Lunar Lake, will have a total computing power of more than 100 TOPS, which is three times higher than Meteor Lake. Among them, the NPU computing power will exceed 40 TOPS, and the GPU will have a computing power of more than 60 TOPS.

AI processing capability has become a new track, and hardware and software innovation is evolving.

Independent NPU is the general trend: XPU collaborates to process different types of AI tasks, among which NPU is a low-power, long-lasting, and efficient computing power base, which is the best processor for executing ubiquitous AI workloads. At the same time, NPU will be applied to more end consumer electronics, and 1TOPS may be a watershed for whether to add NPU units.

ARM stirs up the PC processor market: ARM architecture has obvious power consumption advantages, especially suitable for AI scenarios. After Apple launched the M-series chips, Apple's PC market share skyrocketed from about 7.1% in 2018/19 to 9.5% in 22. After obtaining the process advantages, the ARM architecture became popular with the help of the transitional ecological method. Microsoft quickly followed suit. Microsoft supported Qualcomm's entry into the PC processor market. Not only did the hardware Copilot+PC first use the Qualcomm X-series chip, but it also launched a new translation tool. The application running speed on the same ARM hardware will be improved by 10%-20% after translation.

Machine design or change: The growing demand for memory/bandwidth calls for near-memory computing, which can reduce power consumption and delay through higher degree of integration, or further accelerate ARM PC penetration and heat dissipation demand improvement. The PC motherboard design is expected to usher in a large-scale reform.

Edge AI leads to a rapid increase in storage demand: AI consumes a huge amount of storage for a single device. 7B parameter models need to occupy 4GB of memory and also put forward higher requirements for memory bandwidth. Model loading and running will require a lot of hard disk space. The proportion of downstream DRAM and NAND flash in mobile phones and PCs both exceeds 50%, and edge AI will profoundly change the medium- to long-term demand growth rate and supply-demand pattern of the storage market.

Investment advice

Hardware AI has the potential to trigger a replacement cycle and is expected to land first on mobile phones and PCs. It is recommended to pay attention to the following industries: Mobile phone industry chain-Luxshare Precision Industry (002475.SZ), Suzhou Dongshan Precision Manufacturing (002384.SZ), Avary Holding (002938.SZ), Shenzhen Everwin Precision Technology (300115.SZ), Lingyi Itech (002600.SZ), Lens Technology (300433.SZ), etc.; AI PC industry chain-Huaqin Technology (603296.SH), CECEP CEC Environmental Protection New Material (301387.SZ), Suzhou Chunqiu Electronic Technology (603890.SH), Longyang Electronics (301389.SZ), Jones Tech Plc (300684.SZ), SiQuan New Material (301489.SZ), etc.

Other hardware carriers such as XR/earphones/speakers: AI brings about changes in human-machine interaction and has great potential in new hardware forms. It is recommended to pay attention to the following: Smart speakers-Guoguang Electric (002045.SZ), Edifier Technology (002351.SZ), Bestechnic (Shanghai) Co., Ltd. (688608.SH), Juxin Technology (688049.SH), etc.; XR-Goertek Inc. (002241.SZ), Zhejiang Crystal-Optech (002273.SZ), Shenzhen Zhaowei Machinery & Electronics (003021.SZ), Tianjian Stock Co., Ltd. (301383.SZ), etc.

Storage: AI large models will require a lot of storage for both deployment and operation, and the packaging of storage is expected to undergo changes. It is recommended to pay attention to Montage Technology (688008.SH), Giantec Semiconductor Corporation (688123.SH), Gigadevice Semiconductor Inc. (603986.SH), and JangBoo-long (301308.SZ), etc.

Cooling: AI workloads and near-storage computing increase the heat dissipation challenge and raise the replacement cycle. It is recommended to pay attention to Jones Tech plc (300684.SZ), SuZhou Anjie Technology (002635.SZ), Shenzhen FRD Science & Technology (300602.SZ), and SiQugn New Materials (301489.SZ), etc.

Risk Warning: unexpected technical development, slow application landing, market competition.