Abstract: Generative large models, exemplified by ChatGPT, have achieved unparalleled general-purpose capabilities. They further drive the shift of the scientific research paradigm toward AI for Science, accelerating the convergence of traditional high-performance computing and AI computing. Intelligent supercomputing systems have thus emerged as a pivotal piece of infrastructure underpinning the future development of large models and scientific discoveries.This article systematically analyzes the significant opportunities presented to intelligent supercomputing systems at this historic crossroads, while delving into the critical challenges facing the field in aspects such as computing chips, system architecture, software ecosystems, reliability, and energy efficiency.Addressing these challenges necessitates hardware-software co-design and close collaboration throughout the industrial chain, thereby laying the foundation for building a next-generation intelligent supercomputing system that is efficient, inclusive, and sustainable.