“For physical AI applications like robotics and self-driving cars, on-device processing is necessary to achieve the required fault-tolerant decision-making latency of under 50 milliseconds.”