狂飙的算力基建，如何实现「价值闭环」？丨GAIR 2025

Core Viewpoint - The key to achieving a commercial closed loop in the computing power industry is to provide "convenient, easy to use, and inexpensive" computing power [3][12]. Group 1: Current State of Computing Power Infrastructure - The average utilization rate of computing power in intelligent computing centers is below 40%, indicating a significant issue with computing power consumption [4]. - The demand for reasoning has shifted as large model training has declined, leading to fragmented reasoning scenarios that need to be addressed [4][25]. - The industry is transitioning from a focus on construction to a focus on usability and cost-effectiveness, emphasizing the need for clear user scenarios before building [9][12]. Group 2: Commercial Closed Loop in Computing Power - The commercial closed loop is defined as the ability for AI solutions to be implemented in business scenarios and generate profit [12][14]. - Key conditions for achieving this closed loop include the ease of use and low cost of computing power, which allows creators and developers to fully leverage their capabilities [12][14]. - The MaaS (Model as a Service) model has emerged as a solution to enhance the usability and cost-effectiveness of computing power [12][18]. Group 3: Future Trends and Opportunities - The AI reasoning market is on the verge of a significant explosion, with predictions of a 10-fold growth in the coming year [5][25]. - The integration of multi-modal applications is expected to drive the next wave of growth in computing power demand, with advancements in image and video generation technologies [25][27]. - The widespread adoption of AI glasses and other hardware products could lead to a dramatic increase in token consumption, potentially reaching hundreds of billions [35][36]. Group 4: Key Milestones and Industry Developments - The rise of DeepSeek has reshaped public and industry perceptions of AI, highlighting the importance of AI infrastructure software [31][32]. - Domestic companies are making strides in the super-node architecture, which could lead to breakthroughs in computing power capabilities [33][34]. - The introduction of AI glasses is expected to accelerate data collection and model training processes, marking a significant milestone in the data dimension [34][35].