Workflow
互联网行业:以DeepSeek视角,解读DeepSeek逆袭
浦银国际证券·2025-02-06 14:33

Investment Rating - The report suggests a positive outlook for the AI industry, particularly highlighting DeepSeek's potential to disrupt the market and challenge existing players like OpenAI [2]. Core Insights - DeepSeek has achieved significant cost innovation, with overall costs being only 5%-10% of OpenAI's, including training, inference, and commercialization pricing [2][12]. - The shift in the AI industry is moving from a "compute race" to "algorithm optimization," emphasizing the importance of algorithm density and ecosystem collaboration over mere resource monopolization [2][27]. Summary by Sections Cost Innovation - DeepSeek's training cost is 5.58million,significantlylowerthanOpenAIs5.58 million, significantly lower than OpenAI's 1 billion, achieved through an efficient MoE architecture and sparse activation strategy [7][8]. - Inference costs for DeepSeek are 0.14permilliontokensforinputand0.14 per million tokens for input and 0.28 for output, compared to OpenAI's 2.5and2.5 and 10 respectively, showcasing a cost advantage of 1/10 to 1/20 [9][10]. - Commercialization pricing for DeepSeek is 0.48permilliontokens,farbelowOpenAIs0.48 per million tokens, far below OpenAI's 18, making it more accessible for budget-sensitive enterprises [10][11]. Impact on AI Industry Chain - The low-cost advantage of DeepSeek is reshaping the AI industry chain, leading to a transformation in cloud computing, large models, and application layers [22][27]. - In cloud computing, the demand for inference is expected to exceed 70% of total GenAI demand by 2026, prompting hardware manufacturers to innovate [22]. - The rise of edge computing and hybrid cloud solutions is driven by DeepSeek's support for consumer-grade hardware, indicating a shift from centralized cloud to edge deployment [23]. Large Models and Open Source Ecosystem - DeepSeek's approach is breaking the "compute monopoly" and promoting technological democratization, with its open-source model achieving 1.1 million downloads within six days [24][25]. - The focus is shifting from pre-training to reinforcement learning and inference optimization, fostering the development of vertical models in sectors like finance and healthcare [25][26]. Application Layer Transformation - DeepSeek's low-cost API enables small developers to create AI applications affordably, unlocking potential in underserved markets [26]. - Traditional cloud providers are transitioning to high-margin "AI + industry solutions" models, creating new growth opportunities [26][27].