Most AI security conversations begin at the model: prompt injection, data poisoning, agent permissions, model governance. That focus is necessary, but it rests on an assumption almost nobody questions. It assumes the compute underneath is trustworthy.
Before an AI workload runs, it passes through a long physical chain. Silicon is fabricated, boards are assembled, firmware is flashed, servers are tested, shipped, received, racked, and commissioned, often across several countries and dozens of handoffs. Each handoff is a trust decision. Security teams rarely see most of them, because they sit with operations, procurement, and logistics, not with security.
Drawing on two decades across semiconductor process engineering, automated acceptance testing in server manufacturing, and cloud infrastructure program management, this session moves the AI threat model upstream. It shows where the physical AI supply chain is exposed, why AI-scale buildouts make that exposure worse, and how AI itself can turn manufacturing, testing, and logistics data into security telemetry.
The session also addresses a growing executive concern: compute is now a geopolitical asset. Export controls, chip diversion, and national “sovereign AI” programs mean hardware provenance is becoming a governance, risk, and compliance issue, not just a procurement detail.
Attendees will leave with a broader definition of AI attack surface and a practical framework for bringing infrastructure, operations, and security into the same conversation.