CITIC Construction Investment Electronics Industry Investment Strategy 2025: AI End-side Applications Emerge, High-End Chips Urgently Need Localization

The Zhitong Finance App learned that CITIC Construction Investment released a research report saying that since the launch of ChatGPT in November 2022, the big model has been rapidly iterated, and hundreds of companies are in dispute. OpenAI launched GPT-4O in the first half of 2024, marking a new era of expansion from single text processing to multi-modal understanding and generation. At the same time, the commercialization of end-side AI applications has accelerated, AI phones and AI PCs have been released successively, and extended to wearables, smart cars, XR, etc., with a focus on end-side AI. Rapid iteration of AI has brought rapid growth in computing power demand, demand for advanced manufacturing processes and advanced packaging is rising, and related manufacturers are actively expanding production. The localization rate of traditional semiconductors in China is high, but the self-sufficiency of high-end chips is limited, and there is an urgent need for localization, focusing on the manufacture of domestic high-end chips, core equipment materials, EDA software, etc.

CITIC Construction Investment's main views are as follows:

Looking ahead to 2025, we should focus on two main lines: (1) AI: cloud infrastructure (computing power, storage, communication, etc.), end-side AI (mobile phones, PCs, wearables, smart cars, etc.); (2) domestic alternatives: advanced manufacturing, advanced packaging, equipment materials, etc.

1. Computing power hardware remains booming, and AI end-side applications are on the rise

Hardware: Rubin and HBM4 can be expected, and AI hardware will maintain a strong boom in 2025.

Nvidia released the Blackwell architecture computing power chip, which greatly improved product performance. At the same time, it also launched a new rack AI server GB200 to further drive the demand for hardware related to computing power, storage, and network transmission. The next Rubin architecture is expected to be released in 2025, and HBM4/4e will also be launched at the same time. Demand for computing power, represented by GPUs, CoOS/SoIC, HBM, and high-speed PCBs, continues to expand, and suppliers are vigorously expanding production. It is expected that the AI hardware industry will maintain a high boom in 2025.

Application: Small models empower the end side, accelerate commercialization of AI applications.

The training and inference of the big cloud model requires a large amount of computing power and supporting hardware. The cloud has received a large amount of investment in the past two years, and various big models have been introduced, and the next big model represented by GPT-5 has attracted much attention. At the same time, the end-side model has penetrated various scenarios, such as smartphones, smart glasses, PCs, etc. As end side computing power increases, end side models will play an important role in more fields, especially in application scenarios requiring real-time processing and high privacy requirements. The commercialization of AI applications around the world is expected to accelerate, and the utility of AI applications, such as saving labor costs and personal assistants, is the focus of market attention.

Terminal: Hybrid AI is expected to become a trend, and the value of end-side AI is expected to be monetized.

End-side AI brings advantages in cost, energy consumption, reliability, privacy, security, and individuality. It already has a practical foundation, and terminal devices are expected to usher in a new innovation cycle under the catalyst of AI. Looking at terminals, the first to land and become large-scale terminals will be mobile phones and PCs. In 2024, Apple and Android both released AI flagship models. In 2025, AI phones will innovate and upgrade, and prices will sink. In addition, XR, smart glasses, headsets, smart homes, etc. are also being incorporated into AI.

2. AI leads the semiconductor cycle, and domestic high-end chips urgently need breakthroughs

In this semiconductor cycle, the core requirement is AI.

In 2023, AI demand was mainly in the cloud. The iterative evolution of large models drove rapid growth in demand for computing power chips and infrastructure. At the same time, hardware technology was rapidly iterated. GPUs and HBM were iterated almost a year, and supporting network cards, optical modules, heat dissipation, copper cables/PCBs, etc. were iterated at a similar speed. The learning curve of AI was in a very steep position in the early days. In 2024, AI will enter the end side. As the largest AI carrier, the smartphone AI penetration rate is expected to reach 15%, and industries such as wearables, pharmaceuticals, and high-end manufacturing are also introducing AI. The first beneficiary of this industry trend is AI's upstream hardware industry chain: GPU, storage, PCB, OEM, equipment materials, etc.

AI computing power is limited, and high-end chips are in urgent need of localization.

Hardware infrastructure is the cornerstone for the development of large AI models, but overseas supply of high-end GPUs and HBMs to China is limited, making it difficult to obtain production capacity for advanced manufacturing and advanced packaging. The domestic AI computing power industry chain is still in its infancy. It can be compared to the eve of the outbreak of overseas AI before 2023, but the key link is in a state of “demand, no supply.” Although the domestic semiconductor localization rate has continued to increase over the past few years, the localization rate is still low, such as high-end chip manufacturing, research and development of advanced packaging technology, research on key equipment materials, and development of EDA software. With the localization of traditional semiconductors already having a certain foundation, there is still plenty of room for improvement in the localization of high-end chips, advanced storage, advanced packaging, core equipment materials, and EDA software.

Risk warning:

1. Trade frictions between China and the US may intensify further in the future, and there is a risk that the US government will continue to impose tariffs, set import restrictions or other trade barriers;

2. The upstream AI infrastructure has invested a large amount of capital in R&D and construction. There is no killer application or rigid demand on the end side, and there is a risk that AI applications will fall short of expectations;

3. Adverse factors in the macro environment may slow down the growth rate of the global economy, and residents' income, purchasing power and willingness to spend will be affected, and there is a risk that downstream demand will fall short of expectations;

4. Commodity prices have not stabilized, and the possibility of continuing to rise is not ruled out. There is a risk that raw material costs will rise; 5. The complex global political situation, intensifying disputes among major economies, and increasing uncertainty in the international trade environment may slow down the growth rate of the global economy, thereby affecting the structure of market demand, and there is a risk of the international political and economic situation