By Libra CEO Andrew Williams
AI did not change the physics of the data center. It changed the density, which at the rack amounts to the same thing. Racks that drew 10 kilowatts five years ago are being specified at 120 to 140 and climbing, and somewhere between 30 and 50 kilowatts — depending on who you ask and how much you are willing to spend moving air — air stops being a practical way to carry the heat away. The industry’s answer is liquid, and the workhorse of liquid cooling is the coolant distribution unit, the CDU: a cabinet that moves treated coolant between the facility’s water system and the cold plates sitting on the chips.
Most of the attention in liquid cooling goes to thermodynamics: flow rates, approach temperatures, and heat exchanger performance. That is the right first conversation. But walk through what ships inside a CDU and a second story appears: a CDU is as much an electronics product as a thermal one, and the electronics are increasingly where delivery schedules are won or lost.
The industry has arrived at this from its own direction. It is now common to hear CDUs described as the brains of a direct liquid cooling system — a revealing phrase for a product most people would file under plumbing.
What is inside the cabinet
Strip away the plumbing and a typical liquid-to-liquid CDU carries a substantial electronics bill of materials:
-
A controller and monitoring board. The supervisory brain; holding flow, pressure and temperature setpoints, sequencing the unit, and reporting telemetry to the building management and DCIM systems over whichever protocol the facility standardized on. This is an industrial-grade PCBA with real firmware, communications, and I/O.
-
Pump drive electronics. Variable-frequency drives modulate redundant pumps continuously to hold flow and pressure. Drive boards, gate drivers, and filtering — power electronics built to run hot for years. Redundancy is the point: liquid-to-liquid CDUs at data center scale typically ship with redundant pumps and redundant power inputs and deciding when to fail over between them is a control decision made on a board.
-
Dew point control. The controller has to hold supply temperature above the room’s dew point, which means it is running a closed loop with humidity as an input rather than a simple temperature setpoint. It is one of the few functions in the cabinet where a sensor drift or a control fault puts liquid onto energized electronics, so the sensing chain behind it deserves the scrutiny you would give a safety interlock, including calibration, redundancy, and a defined behavior when an input goes invalid.
-
The sensing chain. Temperature, pressure, and flow nodes throughout the loop, plus differential-pressure measurement across the filter. Each is a small analog chain (transducer, conditioning, calibration) and its accuracy is what decides whether an alarm arrives while there is still time to act or after flow has already fallen. Sensing is the least glamorous content in the cabinet and the most consequential to get right.
-
Leak detection. Often a sensing cable routed through the cabinet base with its own interface electronics, reporting not just that there is a leak, but where. In a room full of servers, leak detection is not an accessory; it is the product’s insurance policy, and the interlock logic that decides whether to isolate a loop is a board-level decision too.
-
HMI and I/O boards. Local display, alarm relays, interlocks.
-
Harness and box build. Power and signal looms connecting all of the above, and the assembled, wired, functionally tested electronics bay itself.
The heat exchanger, pumps, and manifold are precision thermal-mechanical hardware. The boards, sensors, harness, and controls assembly are electronics manufacturing. Two different disciplines, one delivery date.
This is a category still being written
There is a second reason the electronics deserve attention, and it has to do with how young this equipment category is.
The industry is standardizing liquid cooling in public and in real time: the Open Compute Project has published cold plate requirements, liquid distribution reference designs, and guidance on water-based transfer fluids, and several of those documents are still at revision 0 and 1.0. ASHRAE is moving in parallel. An addendum covering liquid-to-liquid CDUs is out for public review, and its data center committee has research underway on liquid cooling resiliency.
A category with unsettled specifications is a category whose products keep changing. And the part of a CDU that changes fastest is not the heat exchanger; it is the controls. Setpoint logic, communications, alarm behavior, redundancy handling, and telemetry all move with each product generation, and each of those moves is a board revision or a firmware revision or both.
That reframes the sourcing question. In a mature category, you buy boards on unit cost at steady-state volume, because the design has stopped moving. This category has not stopped moving. The more useful question to ask a board supplier right now is not what the assembly costs at ten thousand units. It is how many days pass between sending a revision and holding the first article because you are going to do that repeatedly for the next several years.
Why the electronics are the schedule risk
Cooling manufacturers are committing capacity years forward, and every unit of that committed capacity carries the electronics content above. Inside the OEM’s own factory, that content competes with the thermal hardware for the same floor space, the same skilled hours, and the same management attention.
Boards sourced offshore carry the familiar costs: weeks of transit plus customs and handling, buffer inventory to cover it, quality escapes discovered a container-ride too late, and engineering changes that cross an ocean before they cross a bench. On a product line where a customer’s commissioning date is contractual, the longest-lead item inside the cabinet sets the schedule. Increasingly, that item is a board.
The honest segmentation
Not every board in a CDU needs to be built domestically, and a credible partner will say so. Design-stable, high-volume commodity electronics may still favor offshore economics, and pretending otherwise is how suppliers lose credibility in the first meeting.
What belongs close to home is the work where the hidden costs bite hardest. Controller and drive boards still iterating with each product generation, for the revision-speed reason above. Sensor, leak-detection, and dew-point control assemblies, where a quality escape has consequences out of proportion to the value of the board and where the failure mode involves liquid meeting electronics. And controls box-build, where late-stage integration problems are found and fixed in hours across a bench instead of weeks across an ocean.
What to look for in an electronics partner
If you build cooling hardware and are weighing where the electronics content should live, the evaluation is straightforward: IPC Class 3 workmanship as the quality baseline, full lot and component traceability your customers’ QA teams can audit, demonstrated NPI-to-volume discipline so pilot builds become production without a supplier change, and enough proximity that an engineering change turns in days. Ask for the revision turnaround explicitly, in days, and ask what it was the last three times.
Libra Industries builds controller and monitoring PCBAs, drive electronics, sensor assemblies, harness, and functionally tested box-build from Dallas, Texas, and Dayton, Ohio, with the workmanship standards our defense and aerospace customers audit. For OEMs scaling fast, that scope can arrive as a complete dry-side electronics module that’s assembled, wired, and tested as one unit, ready to meet your thermal hardware. If liquid cooling is your ramp, the electronics inside it are a conversation worth having early.
