DS普及和训推一体的趋势下-AI算力需求变化的测算

Summary of Conference Call Records Industry Overview - The records focus on the AI computing power demand changes driven by the acceleration of AI applications, impacting cloud vendors and enterprise users [2][7]. Key Points and Arguments AI Computing Power Demand - The demand for AI computing power is primarily driven by the rapid deployment of AI applications, leading to increased investments in AI software and hardware infrastructure by cloud vendors [2]. - Major public cloud platforms such as Alibaba Cloud, Baidu Smart Cloud, Tencent Cloud, Amazon AWS, and Microsoft Azure are integrating and providing DeepSeek cloud services [2]. - DeepSeek's web version saw a significant increase in traffic, with 190 million visits from January 27 to February 2, 2025, and the iPhone app had a weekly download of 1.78 million [2]. Cost Reduction and Market Impact - DeepSeek's model utilizes technologies like the Mixture of Experts (MOE) and FP8 mixed precision, significantly lowering inference costs, which benefits small and medium enterprises and developers [1][3]. - The API service pricing for DeepSeek is set at 1 to 4 RMB per million tokens for input, and approximately 16 RMB per million tokens for output, making it 56 to 110 times cheaper than previous versions [4]. Hardware Demand Forecast - Due to increased user interactions, an estimated 77,000 to 108,000 additional accelerator cards will be needed, corresponding to 19,000 to 27,000 new servers [5]. - For API calls, an additional 14,000 to 21,000 accelerator cards will be required, leading to the need for 3,376 to 5,144 new servers [6]. Local Deployment and Security - The reduction in inference costs is facilitating the local deployment of AI models, particularly in sectors with high data security requirements, such as state-owned enterprises and financial institutions [8]. - Companies like CloudWalk Technology and Yuntian Lifeng are launching integrated machines pre-installed with DeepSeek models, lowering deployment barriers [8]. AI Chip Market Dynamics - The shift in demand towards inference capabilities is expected to benefit domestic AI chip manufacturers like Cambricon and Huawei Ascend, as inference scenarios have lower performance requirements [7]. - DeepSeek's customized PTX platform may replace some functions of NVIDIA, potentially impacting NVIDIA's ecosystem expansion [7]. Software Market Development - The privatization of DeepSeek is expected to accelerate the development of AI software markets, with platforms like StarRing Technology's TKH Knowledge Platform and Fourth Paradigm's Prophet AI platform poised for commercial deployment [9]. Networking Equipment Demand - The demand for switches in AI computing clusters is expected to rise, with companies like H3C benefiting from their technological leadership and market position [11]. Risks for Domestic AI Products - Domestic AI computing products face risks related to supply chain stability and performance discrepancies, which could affect procurement volumes [13]. - The homogeneity of models developed by various manufacturers, primarily relying on Huawei's components, may lead to competitive challenges [13]. Additional Insights - The overall investment in AI-related hardware and software infrastructure is expected to grow, benefiting various stakeholders, including IDC companies and software developers [12]. - The trend towards AI applications is likely to drive cloud vendors to increase their infrastructure investments, indicating a positive outlook for domestic AI chips, servers, switches, and IDC manufacturers [13].