Open Source Securities: PCs are expected to become the first station for the implementation of AI edge models due to their high-performance hardware and productivity attributes
KaiYuan Securities released a research report stating that with the local deployment of models like DeepSeek on PCs, users will experience more efficient and personalized AI services. PCs, with their powerful hardware performance and productivity attributes, have become the preferred platform for the implementation of AI models. The DeepSeek R1 model significantly reduces development costs through innovative architecture and algorithms, supports the local deployment of multi-level distilled models, enhances model performance, and meets consumer demands for personalization and complex production
According to Zhitong Finance APP, Kaiyuan Securities released a research report stating that the DeepSeek R1 model has recently received widespread attention. It significantly reduces development costs through architectural and algorithmic innovations, achieving relatively ideal model performance while supporting local deployment of different levels of distilled models (currently supporting 1.5/7/14/32/70/671B parameters), retaining its capabilities as much as possible during the process of "shrinking" the large model. With the advancement of local deployment of models like DeepSeek on PCs, it is expected to drive users to experience more efficient and personalized AI services. In this context, it emphasizes that model capabilities determine product experience, and a qualitative change in PC product experience is likely to trigger a wave of device upgrades.
The main points of Kaiyuan Securities are as follows:
Why the first stop for local deployment is PC: Strong hardware capabilities and productivity attributes
(1) PC: The terminal with the most abundant hardware performance. PCs have strong scalability, capable of configuring higher-performance GPUs, NPUs, etc., to efficiently handle complex calculations, while also having more space to configure additional storage and memory to ensure smooth model operation, whereas devices like smartphones and smart glasses are relatively limited.
(2) PC: AI + productivity directly addresses consumer pain points. PCs encompass two major attributes: productivity tools and personal assistants. AIPC can quickly process documents and analyze data during work; plan schedules and manage affairs in daily life; and provide inspiration and efficiently complete designs during creation, such as the Doubao video model to be released in September 2024. Devices like smartphones are relatively limited in the aforementioned scenarios.
Why consumers need AI PCs: The demand for personalization and complex production
The core of AIPC lies in running large models locally to solve personalization issues in a more customized, efficient, and secure manner. Previously, AIPC was able to complete functions such as text generation, image generation, automatic report generation, and AI local knowledge bases, but due to model capability limitations, it usually adopted a cloud + end hybrid model solution. Currently, the performance of models distilled from DeepSeek R1 has further improved, for example, DeepSeek-R1-Distill-Qwen-32B achieved scores of 72.6% and 94.3% on AIME2024 and MATH-500 respectively, significantly surpassing OpenAI-o1-mini on most benchmarks. At the same time, brands like Lenovo and Huawei have integrated DeepSeek into their systems to provide users with a more intelligent and convenient AI interaction experience.
Speculations on the impact on the industry chain:
Complete machines: Complete machine manufacturers play a decisive role in the landing process of AIPC, completing technological convergence and industrial integration, delivering a complete AIPC experience. They are expected to benefit from the device upgrade demand brought about by AIPC penetration. Storage: AIPC penetration is expected to drive growth in storage demand and performance upgrades. According to Micron, the Llama2 model with 70B parameters and 4-bit precision requires approximately 42GB of memory to achieve an output of 1.4 Token/s. Models with fewer than 10B parameters also require about 2GB of incremental memory. Observing the Lenovo Yuanqi series AIPC, 32GB of memory + 1TB of flash storage may be the standard configuration. Cooling: The increase in AIPC computing power and power consumption drives the demand for cooling solutions, with high-value options like heat spreaders/graphite heat sinks expected PCB: Low loss + HDI may be the two major trends for PCBs. Structural components: The trend towards lightweight and high-end is obvious, in addition, AI is also promoting PC definitions, such as Huawei's expected launch of a foldable screen PC in March, which will also drive related demand.
Beneficiary Targets
(1) Brand Whole Machines and ODM: Lenovo Group (00992), Huaqin Technology (603296.SH), Luxshare Precision (002475.SZ), Yidao Information (001314.SZ), etc.
(2) Heat Dissipation: Feirongda (300602.SZ), Zhongshi Technology (300684.SZ), Siquan New Materials (301489.SZ), Longyang Electronics (301389.SZ), etc.; PCB: Aoshikang (002913.SZ), Shenghong Technology (300476.SZ), etc.
(3) Structural Components: Chunqiu Electronics (603890.SH), Lingyi Technology (002600.SZ), Lens Technology (300433.SZ), Guangda Tongchuang (301387.SZ), etc.
(4) Components: Laibo Gaoke (002106.SZ), Hanbo High-tech (301321.SZ), Huichuangda (300909.SZ), etc.
(5) Chips: Haiguang Information (688041.SH), Longxin Zhongke (688047.SH), Xinhai Technology (688595.SH), Dongxin Co., Ltd. (688110.SH), Juchen Co., Ltd. (688123.SH), Lanqi Technology (688008.SH).
Risk Warning
AIPC penetration is less than expected, AIPC product experience is less than expected, downstream demand is less than expected
