Who makes coolant distribution units for AI data centers

The CDU is the boundary of Layer 5: everything upstream (cold plates, manifolds, quick disconnects, hoses) is the technology-side loop it feeds, and everything downstream (chillers, cooling towers, the facility water loop) belongs to Layer 6. Inside, a plate heat exchanger keeps the clean, controlled technology-side coolant physically separate from facility water, while pumps circulate the technology-side flow and filters protect the cold plates from particulates. CDUs come in three form factors — in-rack units serving a single rack, in-row units serving several racks, and facility-class units in the 1–2.5 MW-plus range serving a whole row or hall — and the choice among them is a design decision driven by rack density and facility layout rather than a strict scale progression. As GPU racks moved from air to liquid cooling, CDU capacity requirements rose sharply, and controls (temperature setpoints, redundant pumps, leak response) became as important as the mechanical design. Investors should watch which vendors can scale from smaller in-rack or in-row units to facility-class CDUs, since that is a different engineering and manufacturing problem, and note that some server integrators build CDUs in-house for their own rack designs.