中国信通院发布《算力互联网体系架构研究报告（2025年）》（附下载）

Core Insights - The report by the China Academy of Information and Communications Technology (CAICT) outlines the development background, architecture, networking models, business ecosystem, technical systems, construction progress, and development recommendations for the computing power internet, aiming to provide a comprehensive framework to promote consensus in the industry and empower the development of artificial intelligence, scientific computing, and industrial computing [12][13][14]. Group 1: Development Background - The rapid development of the artificial intelligence industry has significantly expanded the demand for computing power, with global AI spending projected to reach nearly $1.5 trillion by 2025, growing at a compound annual growth rate (CAGR) of 27.5% [20][21]. - Major countries are implementing policies to establish interconnected computing power networks, with the U.S. and EU taking significant steps to enhance interoperability and resource sharing [22][23]. - China is exploring paths for computing power interconnection, with local governments initiating the construction of scheduling platforms to meet regional computing needs [24][25]. Group 2: Overview of Computing Power Internet - The computing power internet (Internet of Compute) is defined as an upgraded system that enhances capabilities for computing tasks and scheduling needs through unified identification and protocol interfaces [26]. - The flow mechanism of the computing power internet involves the movement of computing tasks across different resources, characterized by large data uploads and time-sensitive computations [27][28]. Group 3: Architecture of Computing Power Internet - The architecture consists of three layers: computing network facilities, interconnection resource layer, and application service layer, designed to support efficient interconnection and service provision [41][42]. - The interconnection resource layer serves as the core node, providing essential functions such as resource identification, scheduling parameters, and measurement capabilities [48][50]. Group 4: Networking Models - The networking model includes various configurations for connecting computing centers, emphasizing high-speed and reliable data transmission through both public internet and dedicated enterprise networks [55][56]. - The interconnection resource layer utilizes a "1+M+N" node architecture to facilitate flexible scheduling and resource integration across regions and industries [58][60]. Group 5: Business and Ecosystem - The computing power internet encompasses five main business areas: resource provision, scheduling access, application services, network transmission, and interconnection services [66][67]. - The ecosystem consists of diverse participants, including resource providers, network operators, and application service providers, all collaborating to enhance resource integration and service evolution [72][73].