Major manufacturers have also launched AI terminals, and cloud integration has become the next trend in industry development. In the supply chain transformation, the fully upgraded computing power industry chain is also worth attention. In terms of product structure, the operating income of 10-30 billion yuan products is 401/1288/60 million yuan respectively.
Zhongtong Finance APP learned that, according to the research report released by Minsheng Securities, the biggest change this year is the rise of edge-side models. Compared with cloud-side models, edge-side models have advantages in terms of delay, privacy, and personalization. Major manufacturers have also launched AI terminals, and cloud integration has become the next trend in industry development. In the supply chain transformation, the fully upgraded computing power industry chain is also worth attention. On the one hand, natural language operations and graphic processing require a large amount of memory to assist in calculation and storage, and AI servers have higher requirements for data transmission rates, which has led to the growth of storage and PCB value. On the other hand, in order to meet the strong demand for AI servers, upstream packaging and manufacturing process technology innovation and capacity expansion are also particularly important.
The main points of the Minsheng Securities report are as follows:
Trend of model change and cloud integration
In recent years, generative AI has entered a period of rapid development, with open source and vertical large models blossoming. However, the biggest change this year is the rise of edge-side models. Compared with cloud-side models, edge-side models have advantages in terms of delay, privacy, and personalization. Major manufacturers have also launched AI terminals, and cloud integration has become the next trend in industry development.
Cloud transformation, GB200 solves the bottleneck of interconnection
The current core contradiction in the development of AI industry is the growing demand for applications+terminals and the bottleneck of computing power growth itself. The global competition for AI chips is becoming increasingly fierce. Cloud operators have launched their own AI accelerator card plans in order to reduce capital expenditures. The launch of GB200 NVL72 has significantly widened the gap between Nvidia and its competitors. Through architectural innovation, Nvidia has solved the problem of interconnection bandwidth between GPUs, achieving bidirectional interconnection bandwidth of up to 576 cards and 1.8TB/s, leading other competitors significantly. The Blackwell platform has brought value to upstream optoelectronic modules and PCBs of servers, while the main value increment of the Rack architecture of GB200 lies in copper interconnection, power supply and other links, especially the from zero to one of copper interconnection, which is worth paying attention to.
Supply chain transformation: fully upgraded computing power industry chain
Natural language operations and graphic processing require a large amount of memory to assist in calculation and storage, and AI servers have higher requirements for data transmission rates, which has led to the growth of storage and PCB value. In terms of storage, HBM is one of the important technologies currently used to break the "memory wall". In addition, the penetration rate of DDR5, CXL, MCR and other content technologies is also the focus of industry development. In terms of PCB, server PCBs require high-layer, high-density, and high-transmission-rate PCBs. The OAM and UBB boards of AI servers use up to M7 materials and can reach up to 20-30 layers of high-end HDI. The value of a single AI server PCB can reach 10,000 yuan. In addition, switches and power supply PCBs also have corresponding value increments.
In order to meet the strong demand for AI servers, upstream packaging and manufacturing process technology innovation and capacity expansion are also particularly important. In terms of packaging, the mainstream packaging scheme for computing chips is the COWOS chip + HBM video memory. In terms of equipment, the process capability of 3D stacking and testing equipment is considered, while in terms of materials, thinning, high integration and thermal conductivity are emphasized. In terms of manufacturing, TSMC predicts that the semiconductor market will reach $1 trillion by 2030, with high-performance computing (HPC) accounting for 40%, becoming the largest sub-segment market.
Edge-side change: the computing terminal around us
Under the trend of AI cloud integration, major manufacturers are making efforts in AI terminals to embrace the next wave of growth. Minsheng Securities pointed out that the definition of AI terminals can be refined into three elements: computing power, model, and ecology. Among them, computing power and model are necessary conditions for AI terminals. Computing power is the hardware foundation, which executes neural network operations through NPU to achieve terminal inference; while the pruning and distillation of the model can effectively compress the model parameters, reduce inference time, and reduce edge-side load. The above two elements determine the hard power of the terminal. Compared with the former, ecology requires a long time to polish, and requires industry chain manufacturers to work together to develop applications and explore vertical fields. Ecological construction represents the soft power of terminals.
Minsheng Securities believes that the construction and improvement of ecology will determine the success or failure of AI terminals. And the brand manufacturers are at the core of this process. Apple benefits from its years of accumulation and the appeal of the consumer market, and has a greater opportunity to build a complete ecology. In addition, the PCB, heat dissipation, and optical links of the edge-side supply chain are also worthy of attention.
Investment advice: Regarding the symbols, we suggest paying attention to: 1) Cloud: Foxconn Industrial Internet (601138.SH), Cambricon Technology (688256.SH), WUS Printed Circuit (002463.SZ), Victory Giant Technology (300476.SZ), Montage Technology (688008.SH), Shenzhen Woer Heat-Shrinkable Material (002130.SZ); 2) Supply Chain: Semiconductor Manufacturing International Corporation (688981.SH), Hua Hong Semiconductor Ltd. (688347.SH), ASMPT (00522), Hangzhou Changchuan Technology (300604.SZ), Shenzhen Fastprint Circuit Tech (002436.SZ); 3) Terminal: Lenovo Group (00992), Xiaomi Corporation (01810), Luxshare Precision Industry (002475.SZ), Avary Holding (002938.SZ), Lingyi iTech (002600.SZ), Lens Technology (300433.SZ), Huaqin Technology (603296.SH), Zhejiang Crystal-Optech (002273.SZ).
Risk warning: Downstream demand is lower than expected; development of large models is slower than expected; wafer fab expansion is slower than expected; progress in new product development is slower than expected.