Abstract: With the increasing complexity and intelligence of cyber-physical systems (CPS), traditional centralized computing architectures struggle to meet the demands for low latency and high computational power, making cloud-edge collaboration a promising alternative. However, the current “programming-deployment-debugging” paradigm relies heavily on manual expertise and cannot adapt to dynamic environments. This article proposes an “AI-native cloud-edge-device collaborative computing” framework, which leverages artificial intelligence (e.g., large language models) to achieve autonomous instruction generation, autonomous scheduling and deployment, and autonomous computation evolution, reducing human intervention and enhancing system autonomy. The article focuses on the three challenges of AI-native computing, i.e., controllable instruction generation under software-hardware constraints, scalable computing behavior evolution, and the modeling of complicated heterogeneous resources. This article also surveyed the existing related works and outlines future research directions on AI-native computing.