Skip to content
Koidra
Book a demo

Cooling and power control

Data Centers

PUE is the number you are measured on, and cooling is where almost all of the controllable part of it sits. Yet the setpoints that decide it are still tuned by hand against a comfort margin nobody has revisited since commissioning. The margin is there because the cost of being wrong is measured in hardware, so it stays wide, and the fraction of a point it costs never appears on anyone's report.

Book a demo
Cold aisle in a data centre, server racks with status LEDs and perforated floor tiles running to a containment door

What goes wrong, and what it costs

Chiller short-cycling
Compressor wear and a power draw that spikes without moving the room temperature.
Setpoint drift after a maintenance visit
A margin widened during a callout and never narrowed again. It costs a fraction of a PUE point every hour for years, and nothing on any dashboard ever flags it.
Hot spots from airflow rather than capacity
The response is usually more cooling, when the actual fault is containment or a blanking panel.
Thermal runaway during a cooling failure
Minutes of headroom, and a manual response that depends on who is on call.

Where the energy goes

PUE is the through-line. Cooling is a thermal control problem with an unusually strict tolerance and an unusually large prize: stored thermal mass, slow actuators, fast disturbances, and a hard constraint on inlet temperature. Every tenth of a point of PUE is a permanent line on the power bill, and holding the facility closer to its constraint safely and continuously is where that tenth comes from.

Where the margin sits

We have no PUE figure of our own to publish yet, so here is the shape of the saving instead.

Where cooling energy is decided A temperature scale running from cool on the left to the equipment limit on the right. A recommended band sits inside a wider allowable band, which ends at the hard equipment limit. A setpoint set by hand sits low in the recommended band. The distance from there to the top of the same band is the margin continuous control can recover without widening the envelope. cooler, more energy warmer, less energy equipment limit recommended allowable set by hand and rarely revisited held continuously inside the same envelope the margin worth recovering
The equipment limit is fixed. The envelope inside it is a policy choice. The setpoint inside that envelope is where the energy is decided, and on most sites it was chosen once at commissioning. Closing part of that distance, continuously, with the envelope untouched, is the whole job of the control layer.

The same problem, from a domain with tighter tolerances

Data centre cooling and greenhouse climate are the same control problem wearing different clothes. Both are thermal environments with significant stored mass, slow actuators, fast external disturbances, humidity coupled to temperature, and an operating point that has to be held against a hard constraint.

The greenhouse version is arguably harder in one respect: the thing being controlled is alive, responds over days rather than seconds, and cannot be replaced if you get it wrong. That is the environment KoPilot was proven in, twice, in open competition against expert humans.

We are bringing the same closed loop to a facility where the constraint is inlet temperature rather than crop health.

What that means concretely

We have not run a data centre. There is no customer story on this page because there is no customer yet, and inventing one would be the fastest way to lose an audience that reads specifications for a living.

What transfers is the control approach, not a configuration. Physics-aware models, closed-loop setpoint optimisation, and a supervisory layer that knows what it is not allowed to do. What does not transfer is the equipment model, the safety envelope and the failure modes, all of which have to be built for the facility.

Where to start is the lower steps. Essentials brings power, thermal and maintenance data into one record, which is worth doing whether or not you ever automate the cooling. AI Advisory tells you where the margin is being wasted. Autonomous control is a conversation that follows evidence rather than preceding it.

If you are evaluating this

The honest qualifying questions are about your instrumentation and your appetite. Do you have per-unit power and thermal data at a useful rate? Is your containment sound enough that the control problem is actually control? And is there a facility where a supervised trial is possible before anything runs unattended?

If the answer is yes, we would like to talk. If it is no, the lower steps still pay for themselves.

What we can and cannot claim here

We do not yet run a data centre. What we do run is a harder version of the same physics: KoPilot twice won the international Autonomous Greenhouse Challenge, holding a live thermal environment against expert human operators for a full season. We would rather say that plainly than dress up a case study we do not have.

What we measure here

See Koidra on your facility data

Bring a week of readings from one hall. We will show you what the platform sees, what it would have flagged, and what it would have scheduled.

Book a demo