DeepSeek惊艳世界，算力与应用将迎来结构性变化

Investment Rating - The report maintains a "Positive" investment rating for the computer industry [5] Core Insights - AI applications are entering a practical phase, presenting numerous investment opportunities, particularly for domestic computing and inference service companies benefiting from the AI application boom [3][12] - DeepSeek's recent models, DeepSeek-V3 and R1, have significantly reduced training and inference costs, impacting the AI application and computing industry profoundly [12][29] Summary by Sections Introduction - DeepSeek has released the DeepSeek-V3 and R1 models, which rank among the best in performance compared to both open-source and closed-source models [12][13] Model Performance - DeepSeek-V3 outperforms other models like Qwen2.5-72B and Llama-3.1-405B, matching the performance of top closed-source models like GPT-4o [13][14] - The training of DeepSeek-V3 required only 278.8 million GPU hours, significantly less than competitors, showcasing its efficiency [29][30] Technical Innovations - DeepSeek-V3 employs several innovative techniques, including FP8 precision training and DualPipe architecture, to lower training costs and enhance performance [35][40] - The R1 model utilizes reinforcement learning to improve inference capabilities, achieving performance comparable to OpenAI's o1 model [19][23] Market Impact - The introduction of DeepSeek's models has led to significant stock price drops for major players in the computing industry, indicating a shift in market dynamics [52] - The report suggests that the demand for inference computing will grow as costs decrease, potentially leading to a structural change in the AI computing market [55][57] Investment Recommendations - The report recommends focusing on companies with strong B-end enterprise applications and those with advantages in vertical industry applications, as well as domestic computing and service firms [3][12]