From Prototype to Production: Securely Accelerating Physical AI with Vision-Language-Action Models