When a client tells us they have chosen "direct-to-chip liquid cooling" for a new AI hall, we have learned to ask the follow-up: chosen which CDU, from which vendor, under which service contract, running which dielectric, at what approach temperature, tied to which facility loop topology. Nine times out of ten, the answer is incomplete. The cold plate was the headline. Everything between the cold plate and the chilled-water tower was assumed.
That assumption is where the schedule slips.
The cooling distribution unit is the piece of equipment that translates between the server's demand profile (spiky, hot, sensitive to flow) and the facility's supply profile (steadier, cooler, tolerant of fouling). A correctly sized CDU with the wrong control philosophy will starve a GPU rack during a training checkpoint burst. A correctly controlled CDU with the wrong redundancy posture will take out a pod when a pump fails during commissioning. A correctly engineered unit from a vendor with a three-quarter backlog will arrive a quarter late and delay the whole handover.
The vendor field here is real, and it is consolidating fast. Vertiv, CoolIT, Motivair, Asetek, nVent, Schneider, and Airedale (Modine) are the names that show up on most RFPs. Below that tier, the quality curve drops sharply and the lead-time curve improves, which is exactly the tradeoff a tight schedule will pressure a client into making at the wrong moment.
Three questions we have found genuinely useful during diligence:
One. What is the approach temperature at full load, and who owns the risk if it drifts. An approach that looks fine at commissioning can walk out on you twelve months in as fouling accumulates and filter discipline slips. The warranty language on this is softer than most buyers realize.
Two. How does the CDU fail over, and what does the rack see during the failover. An N+1 on the pump is necessary but not sufficient. The question is whether the transition happens inside the server's thermal tolerance window without dropping a training run.
Three. What does the service footprint look like at the target site. A Motivair unit in Ashburn is a very different service proposition than the same unit in a remote build where the nearest field engineer is on a plane. For a neocloud buildout outside the primary metros, this question changes the vendor ranking.
Our operator lens: the CDU is the component most likely to determine whether a GPU hall comes online on schedule, and it is the component that is most often specified by reference rather than by engineering judgment. Getting it right is less about picking the "best" vendor and more about matching the vendor's real delivery capability to the specific site, the specific operating team, and the specific rack roadmap. That is an advisory problem, not a procurement problem.