The AI data centre, securing the physical foundation of the Australian AI boom

The AI Data Centre as Critical Infrastructure: Securing the Physical Foundation of the Australian AI Boom

Australia's AI ambitions are being poured — quite literally — into concrete, copper, and silicon. Hyperscale build-outs, sovereign AI compute initiatives, and a wave of new colocation and edge facilities are reshaping the national digital footprint at unprecedented speed. Yet most cyber security conversations still stop at the rack. 

In this presentation I will reframe the AI data centre as what it now demonstrably is: critical national infrastructure. It explores the threat landscape unique to AI-era facilities, the regulatory reality under the SOCI Act, and a pragmatic reference architecture for defending the physical substrate that the rest of the AI stack depends on.

The session will start with the overview of the scale of Australia's AI build-out, and the gap between current CISO scope and the realities of AI infrastructure. It then moves into four distinct threats that most existing security models miss: nation-state targeting of compute and model supply chains, OT convergence risks in building management and cooling systems, insider and commissioning risks during construction, and the new attack surface of high-density GPU environments.

From there, the talk addresses the regulatory reality confronting operators and their customers. Data centres are now formally recognised as critical infrastructure under Australia's SOCI regime, with CIRMP obligations, AESCSF expectations, and Essential Eight intersections all landing on the same operating model. 

The centrepiece of the session is a reference architecture that treats the AI data centre as four distinct trust boundaries — physical, OT and facility, IT, and AI fabric — and shows how segmentation, workload identity, and observability connect them without breaking operability.

(To Be Confirmed) Two anonymised case studies from the field ground the framework in real experience: a colocation tenant who discovered a building management system sitting on the same flat network as corporate WiFi and an AI training jump host, and an enterprise AI platform where shared service accounts on a model registry created blast radius no one had accounted for.