Sou Hu Cai Jing·2025-05-07 03:45Core Insights - The multi-modal large model industry focuses on deep learning models capable of processing, understanding, and generating various types of data, including text, images, audio, and video, enabling complex and intelligent tasks [1] - The industry has a wide application potential across various sectors such as natural language processing, image recognition, speech recognition, intelligent driving, and medical imaging diagnosis [1] Industry Overview - The multi-modal large model industry chain is complex, encompassing hardware facilities, software development, and various model types, including CLIP, BLIP, and LLaMA, among others [1] - The industry is divided into three layers: the foundational layer (hardware and basic software), the model layer (various types of multi-modal large models), and the application layer (industry-specific applications) [1] Cost Structure - The training costs for mainstream domestic large models range from tens of millions to hundreds of millions of dollars, with major companies like Baidu, Alibaba, and Tencent investing over $200 million [3][5] - Startups like Kimi and DeepSeek have managed to reduce training costs to between $30 million and $60 million through technological optimizations [3] - Cloud hosting costs are significantly influenced by model scale, with major companies leveraging their own cloud platforms to reduce costs [3] Development History - The global large model industry has evolved through several phases: early exploration (1956-2005), rapid growth (2006-2019), the rise of large models (2020-2022), and the current phase of widespread application starting in 2023 [6] Computational Demand - The demand for computational power in AI is increasing, with larger models requiring exponentially more computational resources; for instance, the GPT-3 model requires 3640 PF-days of computation and at least 10,000 GPUs [9] - As model parameters increase, the computational investment needed grows significantly, influenced by model architecture, optimization efficiency, and hardware capabilities [9]