The Compute Side and Why It Is Not This

To see the facility nervous system clearly, you must see what it is not, and the contrast is compute-side orchestration: the AI factory. This is the layer that decides which workload runs on which processor, when, and for how long. It is Kubernetes and its GPU-aware schedulers, Slurm on a training cluster, NVIDIA's Base Command and the broader software stack that partitions accelerators, places jobs, checkpoints training runs, and recovers from node failure. Its object of concern is the computation. Its currency is FLOPs, job throughput, queue depth, and model iteration time.

The distinction is not academic hair-splitting. The two systems observe different state, actuate different mechanisms, and fail in different modes. Compute orchestration reads GPU utilization, memory pressure, interconnect saturation, job status. It actuates by scheduling, migrating, and killing workloads. When it fails, a training run stalls, a queue backs up, a model ships late. The facility nervous system reads temperatures, pressures, currents, and airflow. It actuates by shifting cooling setpoints, rebalancing power, flagging a failing power supply, or refusing a provisioning request that would exceed a thermal budget. When it fails, a breaker trips, a hot spot forms, hardware throttles or dies, and in the worst case the room goes dark.

Consider a single event through both lenses. A large training job launches and drives a rack of accelerators to full draw. The orchestrator sees utilization spike to target and reports success; the workload is running as intended. The facility system sees rack power jump, inlet temperatures climb, and a cooling loop approach its capacity. To the orchestrator, this is the good outcome. To the facility system, it is the onset of a constraint that may force cooling to respond or power to be rebalanced. Same physical event, two entirely different readings, two different sets of actions. Neither system can substitute for the other, because neither can see what the other sees.

This is why the industry's habit of collapsing them is so costly. When people say a data center is smart, they usually mean the compute side is well orchestrated, and they assume the building follows. It does not follow automatically. A facility can host a sophisticated AI factory on top of a physical plant managed by rudimentary tooling, and the mismatch shows up as stranded capacity, unexpected thermal limits, and outages that the orchestration layer never predicted because it was never watching. Conversely, a superbly instrumented facility can house pedestrian workloads. The two competences are independent variables.

The vendor landscape reflects the split, though it is consolidating in interesting ways. NVIDIA sits on both sides, which makes it a useful case. Its compute stack orchestrates the workload. Its Omniverse platform, by contrast, is used to build facility and industrial digital twins, simulations of the physical plant and its environment. That a single company ships tools for both surfaces does not merge them; it confirms they are distinct enough to warrant separate products. Cadence, through its Reality DC platform, comes at the facility from engineering simulation, modeling airflow and thermal behavior of the physical room. These are facility-side tools. They model the building. They say nothing about the schedule of jobs inside it.

Why does the boundary matter so much that a course is built around it? Because in accelerated computing the physical plant has become the binding constraint. For most of computing history the building was a supporting utility, provisioned once and largely ignored. Power densities were modest, cooling was air and slack was generous. That era is over. At current rack densities the facility is where the hard limits live: power delivery, heat rejection, and the physical integration of liquid cooling. The intelligence that governs those limits is therefore not a back-office function. It is a primary determinant of what the compute side can do.

So the two nervous systems coexist, coupled but distinct. The compute side orchestrates the work. The facility side sustains the conditions under which work is possible. DCIM and the facility digital twin are the intelligence of the second system. Keep them separate in your analysis, and you can reason clearly about either. Fuse them, and you will misattribute every failure and misjudge every operator. The rest of this course develops the facility side on its own terms, as the distinct surface it is.