电子行业行业点评报告：DeepSeek降低制程依赖，带动国产供应链

Investment Rating - The report maintains a "Recommended" investment rating for the electronic industry [1]. Core Insights - DeepSeek has achieved cost reduction and efficiency improvement through engineering innovation, utilizing only 2048 NVIDIA H800 GPUs at a cost of $5.576 million to train models equivalent to GPT-4 and Claude-3.5-Sonnet [3]. - The introduction of DeepSeek's new open-source inference model marks a significant milestone in making AI more accessible, reducing reliance on high-end GPUs and adapting to domestic chip architectures [3]. - The emergence of DeepSeek's model is expected to stimulate domestic computing power demand and enhance the semiconductor manufacturing supply chain in China [3]. - The report suggests a positive outlook on the explosion of domestic computing power demand and the elevation of the semiconductor manufacturing industry's ceiling, recommending attention to specific companies in the semiconductor sector [3]. Summary by Sections Industry Overview - The report highlights the adaptation of 16 domestic AI chip companies to DeepSeek's model services, indicating a growing ecosystem around domestic AI capabilities [3]. - Traditional AI training has heavily relied on high-performance chips, but DeepSeek's innovations allow for compatibility with domestic chip architectures, paving the way for broader AI adoption [3]. Technology Innovations - DeepSeek's technological advancements include architectural innovations, efficient pre-training frameworks, and knowledge distillation techniques that enhance model performance and reduce training costs [3]. Investment Recommendations - The report recommends focusing on domestic computing power chip companies such as Cambricon and Haiguang Information, semiconductor manufacturing companies like SMIC, and semiconductor equipment and materials companies [3].