AI adoption across the region is accelerating faster than the security models surrounding it. Most AI security discussion today stops at the model and the application: prompt injection, data governance, responsible use. Far less attention is paid to the physical and firmware layer the models actually run on — GPUs, data processing units, high-speed fabric, and the out-of-band management controllers beside them, many of which are reachable independently of the operating system. This session maps the AI infrastructure attack surface end to end and walks the GPU server lifecycle: onboarding, multi-tenant sharing, redeployment between customers, and decommissioning after three to five years of service. It is written to be equally useful to organisations building their own AI capacity and to those consuming it from a cloud or specialist provider, because renting the compute does not transfer the risk. Attendees leave with a due-diligence checklist for interrogating both their own AI infrastructure and their providers'.