The increasing complexity and diversity of ArtificialIntelligence (AI) workloads, from generative models to realtimedecision-making systems, require a transition from traditionalcentralized architectures to a seamless, cognitive cloudedge-IoT continuum. This paper presents a comprehensive andunified reference architecture for AI-driven orchestration andinfrastructure management across heterogeneous computingenvironments. We introduce novel mechanisms to optimize AItraining and inference workflows across distributed edge, cloud,and high-performance computing (HPC) infrastructures, whileensuring trust, privacy, and energy efficiency. Our approachintegrates virtualization, federated learning, data compression,and conditional computing techniques to support scalable,secure, and context-aware AI applications. Finally, we analyzeand classify the main architectural and operational challengesinvolved in enabling AI-