The Zhitong Finance App learned that Guohai Securities released a research report saying that compared to PCs, mobile phones are terminal devices closer to users' lives, so users will naturally generate this demand, but mobile phones at this stage are not intelligent enough, so they are equipped with large models and develop more functions or become a trend; the early popularity of PCs is due to its positioning as a productivity tool, and AIGC as a generative AI will naturally match users to improve production efficiency, so AI PCs are also on the rise. AI technology continues to evolve, and large models continue to iterate and show a downward trend towards the end. The bank is optimistic about investment opportunities in AI PCs and AI phones, and maintains the industry's “recommended” rating.

Guohai Securities's views are as follows:

Judging from the development history of mobile phones and PCs, the empowerment of AIGC may bring both to a new stage of development

As far as mobile phones are concerned, the bank summarizes its development path: emphasis on communication functions → smartphones → personal assistants. Compared to PCs, mobile phones are terminal devices that are closer to users' lives, so users naturally have this kind of demand, but mobile phones are not intelligent at this stage, so they are equipped with large models and developed more functions or become a trend; it is because of its positioning as a productivity tool in the early stages that AIGC is a generative AI that naturally matches productivity tools to help users improve production efficiency, so AI PCs are also on the rise.

From a hardware perspective, AI PC/Phone needs to solve three problems: First, what can be used to speed up inference? Second, how do you load a large, high-spec model into the device? Third, reasoning is a high-power task, how should the cooling system of the equipment cooperate? Derived from this, the bank's discussion on PC and mobile phone processors, memory, and cooling systems.

As for question 1, all current AI PC/phone manufacturers have given an answer. In the past, deep learning or machine learning on the end side usually used GPUs as computing units, but with the implementation of AIGC on the end side, special processors for artificial intelligence tasks have also followed the trend. Although GPUs have stronger parallel computing capabilities than CPUs, their processing neural networks are still inefficient and consume high power. Compared to GPUs, the optimization of NPU mainly focuses on interaction with the CPU. NPU does tensor calculation. If the GPU sees the matrix as a “calculation unit,” then NPU sees the neural network as a “calculation unit”. As a result, NPU has a better energy consumption ratio. The bank believes that the future may be an era where CPUs, GPUs, and NPUs coexist. Using the history of GPU development as a reference, in the future, NPU may gradually become one of the essential hardware for PCs and mobile phones. Currently, NPU is still integrated with the CPU, and may also develop a form similar to a unique display with independent memory in the future, and become an important standard for PC product positioning (for example, today's high-performance books and game books are usually unique displays, while lightweight and commercial books are generally core displays). Currently, the development of NPU is still in its early stages. The bank believes that its focus is mainly on volume, that is, popularizing this hardware on PCs; in the future, as NPU gradually becomes accepted by users and places higher demands on performance, reference GPUs may bring more premium space. The bank believes that this revenue is mainly distributed among upstream hardware vendors (NPU suppliers), large model suppliers, and terminal equipment manufacturers.

Regarding question 2, at present, PCs and mobile phones have relatively small memory, and the carrying capacity of large models is limited, and intelligence can only emerge when large models reach a certain level of parameters. Therefore, in order for large models to bring users a better experience, on the one hand, it is necessary to cut and quantify the model, and on the other hand, it is also necessary to increase the memory configuration of the device to make it more balanced in terms of storage, transmission, and calculation. Currently, the large models that have been implemented in terminal devices are still mainly in the order of several billion parameters. There is still room for improvement in performance. In the future, commercialization of high-spec large models will inevitably place higher demands on the memory of terminal devices.

Regarding question 3, as an important part of the electronic equipment field, the cooling problem of PCs and mobile phones has always attracted much attention. NPU, which is an AI chip with high computing power, will place higher demands on the cooling system of devices, so as the specifications of models mounted on AI PCs and AI phones continue to improve, NPU performance release may be more aggressive. The bank believes that this will bring about the upgrading of cooling solutions for electronic products and related upstream and downstream investment opportunities.

Large models land on the end side, and the main beneficiaries include hardware (such as NPU, etc.) design and manufacturers, large model providers, and terminal equipment manufacturers.

Although AI PCs and AI phones are still in the early stages of commercialization, the bank believes that terminal equipment manufacturers may be the core for the following reasons: terminal equipment manufacturers are the only direct customer players; in addition, they must also pre-load large models into terminal equipment products based on the underlying ecosystem (such as cooperation between Intel and Microsoft) built by software and hardware manufacturers collaboratively. The differentiated competition in this production process mainly focuses on cooling module design, hardware assembly capabilities, and model distillation levels. In terms of revenue, on the one hand, terminal equipment manufacturers can raise the price of products based on the new hardware called NPU, and on the other hand, charge users service fees for large models and related applications; judging from the cost, on the one hand, terminal equipment manufacturers must purchase basic components from hardware design manufacturers, and on the other hand, they must pay model transfer fees to large model providers. In summary, terminal equipment manufacturers act as intermediaries to connect models, hardware, and C-side users to earn the difference in price while also earning service fees (for example, for enterprise users, selling devices while also providing model deployment services). As a result, terminal device manufacturers may end the era of a single business model where large models used large model manufacturers as the core when providing services through API interfaces or ChatGPT-like applications in the cloud, and become a new force competing for traffic entry.

Investment advice:

In terms of hardware, it is recommended to focus on chip and memory manufacturers, such as Haiguang Information (688041.SH), Tongfu Microelectronics (002156.SZ), and Juchen (688123.SH);

In terms of models, it is recommended to focus on Microsoft (MSTF.US), Google (GOOGL.US), etc.;

In terms of terminal equipment, it is recommended to focus on PC/mobile phone providers and foundries with a relatively high market share, such as Lenovo Group (00992), Xiaomi Group-W (01810), and Everbright Tongchuang (301387.SZ).

Risk warning: AI PC/mobile phone commercialization falls short of expectations; technology leapfrogging and substitution, disrupting existing technology paths; industry competition increases risk; industry valuation correction risk; companies are not fully comparable; relevant data and data on targets are for reference only.