The Power Loop and the Cooling Loop Need to Start Talking

Two systems, one building

In every data center, the two largest energy systems are power and cooling. They share the same building, the same workloads, and in many cases the same operating envelope. But they are typically designed by different teams, sourced from different vendors, and controlled by software systems that don't communicate with each other.

For most of the industry's history, that was fine. Densities were modest, workloads were stable enough that the two systems could each do their own job without much real-time coordination. The inefficiency at the boundary was absorbed by overhead.

That equilibrium is breaking down. At the densities driving AI infrastructure today, the cost of keeping power and cooling on separate tracks is now visible in the efficiency metrics operators track closely. The next generation of data center efficiency gains will come from making the two sides operate as one system, not from improving each side in isolation.

How the separation costs efficiency in practice

Inefficiencies show up most clearly during workload transitions. Consider an AI training job that starts up: GPUs ramp from near-idle to full power in seconds. The power side responds quickly because it has to. The cooling side, working from its own sensors and controls, often catches up a beat later. In that window, chip temperatures can rise toward throttle limits.

To prevent that, most operators run cooling with conservative setpoints. The room is held cooler than it strictly needs to be so chip temperatures stay clear of throttle limits even when cooling response lags behind power demand. That margin is, by definition, inefficiency. The cooling system is using energy to prepare for thermal events it does not yet see.

The same pattern runs in reverse. When workloads drop, cooling does not always wind down as fast as it could, partly because the controls are not getting a clean signal that power demand has fallen. Both systems carry overhead for events the other side already knows about.

What convergence actually looks like

Convergence is best understood as an operational design that gives the two systems shared awareness and shared response. Three elements have to be in place.

First, sub-second telemetry from both loops, available to a common control plane. The control side needs to see real-time power demand and thermal load together.

Second, bidirectional signaling. Power can tell cooling what is coming. Cooling can tell power what it can support. A workload spike that creates a margin problem in the separated model becomes a manageable event when the cooling side is pre-positioned.

Third, control algorithms that treat efficiency as a joint optimization rather than two parallel ones. Lowering chilled water temperature uses more energy on the cooling side but allows higher power densities on the IT side. With shared visibility, the optimum can be found dynamically instead of being set conservatively at design time.

In practice, the change is operational rather than physical. The sensors, controls, and connectivity are mostly available. The integration discipline is what tends to be missing.

This is where existing DCIM platforms come up short. They were built for supervisory monitoring rather than real-time control, with data latencies typically measured in minutes. Convergence requires sub-second response and bidirectional control.

Why this matters more now

Density is what is pushing convergence from a good idea to a requirement. AI training racks routinely draw 60 to 100 kilowatts, with hyperscale designs pointing higher. A few percentage points of efficiency on a 30 megawatt site is a real number. On a 300 megawatt site, the same points represent megawatts of compute the operator can or cannot deploy.

Higher density also brings tighter operating margins overall. New power architectures, including 800 volt DC distribution, enable the higher densities but also leave less headroom at the rack. A small mismatch between power delivery and thermal response no longer gets absorbed by the system.

The systems that used to operate independently can no longer afford to do so.

What nVent is doing

At nVent, we work on both sides of this loop. Our portfolio spans cooling distribution, manifolds, rear door heat exchangers, and rack systems, alongside power distribution products that sit on the same critical path. That breadth gives us a direct view into how the two systems hand off to each other.

Across more than 2GW of liquid cooling deployed, the pattern is consistent: sites that move toward coordinated control see real efficiency gains, and they often do it without major hardware overhauls.

Our current work is in the control plane between the two sides. That includes sub-second sensing, algorithms that treat power and cooling as a joint optimization, and operational intelligence that learns a specific site's patterns and adjusts accordingly.

We are building this in deployments today, with operators who are willing to treat their two largest energy systems as one.

What operators should be asking

For operators planning AI buildouts now, three questions are worth asking before the design is locked in.

Are your power and cooling vendors talking to each other at the design stage? If they are signing separate scopes of work that meet at a single page in the building specification, the integration risk is being moved from the supplier side to the operator side.

What sub-second telemetry do you have on both loops? If your monitoring resolves at minute or even second intervals rather than sub-second, you do not have the data to coordinate the two systems in real time. The instrumentation question is upstream of every other convergence question.

Are you procuring power and cooling as separate stacks or as a coordinated system? Procurement structure tends to dictate operational structure. Two RFPs typically yield two systems that do not coordinate.

Each of those questions reflects the same underlying shift.

Looking Ahead

The data center industry has operated for decades at a level of integration that matched the workload. Power and cooling did not need to coordinate in real time, and the boundary inefficiency was small enough to live with.

AI changes the calculus. The workloads are too volatile, the densities are too high, and the margins are too tight to keep treating the two systems as independent. The operators who lead on efficiency over the next few years will be the ones who treat power and cooling as one system.