国君计算机|效率革命剑指“暴力计算法则”——Deepseek重塑AI时代大模型研发范式

Investment Rating - The report suggests a positive outlook for cloud service providers due to Deepseek's reduction in hardware computing power demand, indicating a new growth momentum in the short term for local deployment in large enterprises and specific industries [1] Core Insights - Deepseek aims to enhance unit computing efficiency by tenfold through algorithm optimization, significantly lowering model training and inference costs. The training cost for the 671B DeepSeek V3 is $5.576 million (approximately 40.7 million RMB), which is only 7% of Llama 3's cost, while OpenAI's ChatGPT-4o training costs range from $78 million to $100 million, requiring thousands of NVIDIA H100 chips. In contrast, DeepSeek V3 utilizes the NVIDIA H800, a specialized AI chip with reduced performance [1] - The technological revolution of Deepseek introduces a new paradigm in large model development, employing innovative architectures like MoE and MLA for efficient inference and cost-effective training. The dynamic sparse expert network design allows the model to utilize less than 4% of neural network parameters during inference, and the FP8 low-precision training framework reduces energy consumption by 80% while maintaining model stability [2] - Open-source models like DeepSeek are expected to play a crucial role in the AI era, similar to how Android transformed the mobile internet. This will reshape the industry ecosystem, accelerating the development of upper-layer applications and unifying lower-layer systems, thereby enhancing collaboration across software, hardware, and supply chains [3] Summary by Sections - Investment Outlook: The report highlights the potential for local domestic inference computing to explode, alongside the expansion of new foundational software like vector databases [1] - Technological Innovations: Deepseek's introduction of reinforcement learning-driven paradigms and self-evolving training mechanisms significantly reduces the data annotation requirements for efficient training, showcasing a systematic disruption of the "computing power arms race" [2] - Market Impact: By reducing reliance on high-end imported chips, Deepseek provides a viable technological path for domestic enterprises, boosting confidence in the development of self-researched computing power chips [3]