The recent evolution of artificial intelligence (AI), from early rule-based systems to deep learning-powered foundation models like GPT, has primarily focused on virtual environments. However, the emergence of Physical AI marks a pivotal shift—enabling machines to operate autonomously in dynamic, unpredictable real-world settings. This talk explores the computing challenges and opportunities in deploying Physical AI systems, particularly embodied AI and robotic foundation models, where perception, action, and decision-making must occur in real time with strict energy and latency constraints. It presents key insights into NVIDIA’s ecosystem—including Omniverse, Isaac, Jetson, and Project GR00T—that supports the simulation, training, and deployment of physically intelligent agents. The presentation also discusses the limitations of conventional compute architectures in meeting Physical AI demands, pointing to domain-specific accelerators (DSAs), Processing-In-Memory (PIM), and neuromorphic/analog computing as promising alternatives. Emphasizing the need for holistic cross-stack innovations, the talk highlights that hardware-software co-design is essential to overcoming scalability and energy efficiency barriers in the path toward general-purpose physical intelligence.