COMPUTEX 2026 CEO Keynote: NXP — Physical AI and the Neural Axis Architecture

Author: Rafael Sotoayo, President and CEO, NXP Semiconductors Collection: Computex 2026

Summary

NXP Semiconductors CEO Rafael Sotoayo delivered a 71-minute keynote at Computex 2026 Taipei framing the company's strategic position in physical AI — the deployment of AI into real-world autonomous systems including drones, software-defined vehicles (SDVs), and humanoid robots. The central thesis is that physical AI requires solving Moravec's paradox: the observation that tasks easy for humans (balance, reflexes, gripping) are the hardest for machines, while tasks hard for humans (chess, theorem proving) are easiest for machines. Sotoayo argues that edge computing — fit-for-purpose semiconductors with real-time performance, ultra-low latency, low power, and high safety/security — is the essential substrate for physical AI, and that NXP is uniquely positioned to provide it having built products for automotive, industrial, and infrastructure markets for decades.

The keynote introduces NXP's "neural axis architecture" as the organizing framework, drawing a direct analogy to human neurobiology: the cerebrum (reasoning, 80% of brain mass, ~17 billion neurons, 300+ ms response time) maps to the AI reasoning layer; the cerebellum (coordination, 10% of brain mass, ~69 billion neurons, <100 ms) maps to the coordination layer; and the spinal cord (reflexive, 2.5% of brain mass, ~200 million neurons, <40 ms event-driven response) maps to the reflex layer. This three-layer architecture is shown applied consistently across drones, autonomous mobile robots (AMRs), SDVs, and humanoid robots, with NXP's S32N family (5 nm process) anchoring the coordination layer in vehicles and its S32K family anchoring the reflexive/zone-control layer.

A second conceptual pillar is the distinction between robot motion and robot understanding, addressed through Vision-Language-Action (VLA) models fed by world models encoding physics (gravity, liquid flow, friction, inertia, human safety awareness). The keynote introduces NXP's eIQ toolkit as the software bridge that takes cloud-trained VLA models and quantizes, prunes, and compiles them for constrained edge hardware. The final pillar is trust, defined through four primitives: Contain (isolation, redundancy, no single point of failure), Protect (hardware-injected security, post-quantum cryptography readiness), Verify (SafeAssure program, SIL/ASIL certifications, PSA/SESIP/FIPS 140-3/IEC security certifications), and Adapt (long-horizon secure update capability for decade-deployed edge devices). Real-world deployment evidence includes 40% productivity gains over traditional automation in factory settings and 610% increased sales of medical laboratory robots and diagnostics in 2025, with named partnerships including Boston Dynamics and GE Healthcare.