When IT Meets OT: The Data Center as Cyber-Physical System
Draw the two worlds first, because their collision is the whole point. Information technology, IT, is the domain of data: servers, storage, networks, the accelerators that train models, and the software that orchestrates them. Operational technology, OT, is the domain of the physical plant: the building management system, the power distribution and backup generators, the chillers and pumps that reject heat, and the industrial controllers that run all of it. For most of computing history these were separate worlds with separate staff, separate vendors, and, crucially, an air gap between them. That separation is gone.
Convergence happened for good reasons. Operators wanted to monitor power and cooling remotely, tune efficiency in real time, and correlate facility telemetry with compute load. So the control systems were connected to networks, given remote management interfaces, and wired into the same dashboards as everything else. The efficiency gains are real. The security consequence is that a domain built on the assumption of physical isolation is now reachable over a wire, and much of it was never designed to survive contact with a determined attacker.
OT carries structural weaknesses that IT security spent decades learning to manage. Industrial controllers often run for years or decades without patching, because taking them offline means interrupting the physical process they govern. Many speak legacy protocols with little or no authentication, built for a trusted, closed network. Their firmware is rarely inventoried and almost never scanned. In IT terms, the OT layer is a fleet of unpatched, hard-to-see, hard-to-authenticate devices that happen to control whether the building stays cool and powered. An adversary who understands this reaches for the OT layer precisely because the defender rarely watches it.
AI facilities sharpen the stakes because their physical demands are extreme. Racks of accelerators draw power densities that would have been unthinkable a decade ago, and every watt becomes heat that must be removed continuously. Advanced liquid-cooling loops sit directly against the silicon. Remove or degrade cooling and the hardware has seconds to minutes before it must throttle or shut down to avoid permanent damage. The physical plant is not a support function around the computer; it is part of the computer. That is what a cyber-physical system means, and it is why an attack on the chiller is an attack on the model.
Consider the disruption path concretely. An attacker who reaches the building management system does not need to touch a single byte of training data. Manipulate cooling setpoints, trip a power distribution unit, or falsify the sensor readings the safety logic depends on, and the facility's own protective systems will halt the compute to save the equipment. The training run dies, the scarce accelerator hours evaporate, and from the inside it can look like an equipment fault rather than an attack. The attacker exploits the facility's own safety design as a weapon.
The defensive principle is segmentation with intent. IT and OT should not share flat network reachability just because they share dashboards. Put controlled boundaries between them, allow only the specific flows monitoring genuinely requires, and watch those boundaries as closely as you watch the internet edge. Extend asset inventory and monitoring into the OT layer so it stops being a blind spot. And design the human response for cross-domain incidents: the team that investigates a strange login and the team that investigates a chiller fault must be the same team, or at least talk in real time, because in a converged facility those two events can be the same event seen from two sides. The organizing lesson holds: defend the whole system, because the adversary already sees it whole.