DeepSeek重构算力基建长期价值的认知

Investment Rating - The report rates the industry as "Buy" [1] Core Insights - The market has underestimated the amplifying effect of the DeepSeek ecosystem on computing power demand, with an expected near million PFLOPS of demand generated solely from its inference end [3] - Domestic AI chip manufacturers, particularly those like Huawei Ascend, are poised to benefit significantly from the reduction in entry barriers for large model training, expanding the overall market size [12] - The emergence of the DeepSeek ecosystem presents unprecedented opportunities for domestic AI chips, with Huawei Ascend's performance nearing international standards [12] Summary by Sections Investment Recommendations - DeepSeek's technological breakthroughs, while raising short-term concerns about high-end AI chip demand, have expanded the overall market size by lowering the entry barriers for large model training. Domestic chip manufacturers, especially Huawei Ascend, are expected to gain market share due to their cost-performance advantages in enterprise deployment [12] - Recommended stocks include Unisplendour, Inspur Information, and iFlytek, with beneficiaries including CloudWalk Technology, Topwise Information, Digital China, and Zhongke Shuguang [12] DeepSeek - DeepSeek-V3 has set a new economic benchmark for large language model training costs at $557.6 million, utilizing only 2.788 million GPU hours to complete full training, which has led to a reevaluation of AI computing cost [12] - The technology innovations from DeepSeek have not diminished the demand for high-performance AI chips but have instead expanded the market size by lowering entry barriers and generating massive inference demand [12] Training Innovations - DeepSeek V3 and R1 have significantly reduced large model training costs through innovations such as MLA mechanisms, FP8 mixed precision training, and DualPipe parallel frameworks [14] - The Multi-Token Prediction (MTP) mechanism in DeepSeek-V3 allows for more efficient data utilization and dense training signals, enhancing the model's long-term dependency capabilities [19] Inference Optimization - DeepSeek V3 employs a dual-stage inference architecture to balance service quality and throughput, optimizing the deployment costs for large-scale applications [35] - The R1 series utilizes model distillation techniques to achieve smaller model deployments, significantly lowering inference costs [41] Market Dynamics - The low-cost breakthroughs from DeepSeek have prompted a reassessment of AI development paths, with a notable market reaction reflected in Nvidia's stock price drop [42] - Despite the reduction in per-call costs, the rapid user growth of DeepSeek has led to a surge in overall computing demand, highlighting the ongoing need for high-performance computing infrastructure [44] Scaling Law and Future Trends - The report emphasizes that AI development continues to follow Scaling Law, with increasing model, data, and computing scales driving demand [52] - The trend towards multi-agent and multi-modal AI systems is expected to further increase computing power requirements, as these systems necessitate complex reasoning and real-time adjustments [59][63]