Abstract
Low Earth Orbit (LEO) satellite constellations are rapidly evolving, providing pervasive coverage that enables new opportunities for AI-driven services. Delivering Deep Neural Network (DNN)-based applications on these satellites faces unique challenges due to limited onboard computational resources, energy constraints, and stringent latency requirements. In this article, we review the state-of-the-art techniques for onboard DNN inference and task offloading, highlighting the key bottlenecks and resource management challenges. We further discuss emerging solutions, including reinforcement learning-driven offloading strategies and lightweight AI models, that can enhance resource efficiency while maintaining service quality. Finally, we outline future research directions and potential application scenarios for realizing high-quality AI services in LEO satellite networks.