Jayson Bethurem

Adaptable Compute Architectures for Physical AI Systems

Abstract

Physical AI is transforming industries ranging from autonomous transportation and robotics to aerospace, industrial automation and smart infrastructure. Unlike cloud AI, Physical AI must interact with the real world in real time — sensing, deciding and acting with deterministic latency, reliability and safety. However, one critical challenge is often overlooked: these systems are being deployed into environments where AI models, sensor pipelines, communication standards, security requirements and safety regulations will continuously evolve over operational lifecycles that can span decades.

This presentation explores why static hardware architectures fundamentally limit the long-term viability of Physical AI systems and why adaptability must become a core architectural requirement. We will examine how evolving workloads, sensor fusion complexity, emerging security threats and changing AI algorithms create unavoidable obsolescence risks for fixed-function hardware approaches.

The session will also discuss how adaptable and reconfigurable hardware architectures enable long-lifecycle Physical AI platforms by supporting post-deployment evolution, deterministic edge processing, hardware/software co-design and real-time optimization. Through real-world examples across robotics, aerospace & defense, autonomous systems and edge AI, we will demonstrate why adaptable compute is rapidly becoming foundational infrastructure for the next generation of intelligent physical systems.