计算机行业深度分析：三大要素齐发力，AI应用步入全面加速期

Investment Rating - The report upgrades the investment rating for the computer industry to "Outperform the Market" [1] Core Insights - The report highlights that the AI application is entering a phase of comprehensive acceleration, driven by three key factors: enhanced logical reasoning capabilities, reduced inference costs, and the performance surpassing of open-source models over closed-source models [4][11] Summary by Sections 1. Enhanced Logical Reasoning Capabilities - OpenAI's o1 has opened new pathways for improving logical reasoning capabilities in large models [7] - DeepSeek has publicly validated the effectiveness of reinforcement learning in enhancing the reasoning capabilities of large models, achieving a significant increase in AIME scores from 15.6% to 71.0% [19] 2. Reduced Inference Costs - DeepSeek has established a low-cost training path, with the DeepSeek-R1 model achieving performance comparable to OpenAI's o1 at only 1/30 of the API calling cost [27][30] - The architecture of DeepSeek models incorporates innovations that significantly reduce computational costs, including the Multi-Head Attention mechanism and the MoE architecture, leading to efficiency improvements of 2-4 times and over 4 times, respectively [31][32] 3. Performance Surpassing of Open-Source Models - The report notes that the open-source model ecosystem has for the first time surpassed the closed-source model ecosystem in terms of performance, with DeepSeek's R1 and V3 models achieving capabilities on par with the most advanced closed-source models while maintaining cost advantages [66] - The open-source approach adopted by DeepSeek and other companies like Meta and Alibaba has contributed significantly to the advancement of AI technology, allowing for broader application and deployment [65][66]