DDC Solutions has stood up a live AI cooling showcase inside the operating ScaleMatrix facility in San Diego, with neocloud provider Cirrascale Cloud Services as the anchor tenant running GPU workloads, according to the announcement on PR Newswire. The stack is three Daikin businesses in one row: facility-level chiller infrastructure from Daikin Applied, a NEMA 3R rated DDC S-Series rack containment unit, and a Chilldyne negative-pressure CDU handling direct-to-chip. Rack-level fire suppression and DCIM monitoring sit on top. Keith Markley, chief executive of DDC Solutions, said the site is "built to give infrastructure buyers a real answer before they commit." Scheduled visits include a facility walkthrough, live workload observation, an architecture briefing, and training on hybrid cooling design.
DDC's own published material describes ScaleMatrix as the original creator of Dynamic Density Control technology and its San Diego facility as the testing ground for DDC's cooling products. This is the building where the S-Series was developed, which is closer to a factory acceptance test with customers watching than to an independent deployment. What survives that disclosure is still useful, and it is more than most vendor labs offer: a hall with real GPU load, real ambient, real chilled water, and a paying tenant whose SLA does not care about the demo schedule. Buyers should read it as a working reference site rather than as third-party validation, and should ask DDC for a customer deployment they can visit that nobody in the Daikin group helped design.
Leak risk is the objection that stalls liquid retrofits in occupied halls, ahead of capex. Most of the global fleet is already built, and liquid cooling has to work around it. Chilldyne runs its cold plate loop below atmospheric pressure, so a seal failure draws air inward and coolant stays inside the pipe instead of spraying energized hardware. The company claims a zero-leak record across every field deployment since 2014, and that claim is checkable, because a leak in a GPU hall generates an incident report. Daikin Applied bought Chilldyne in November 2025, and San Diego is the first place the chiller plant, the containment, and that CDU appear as one commissioned system instead of three datasheets a mechanical contractor is expected to reconcile on a slab.
The constraint is fixed and physical. A sub-atmospheric loop has about one atmosphere of driving pressure available, minus the coolant's vapor pressure, and that budget has to cover every cold plate, hose, quick disconnect, and fitting between the CDU and the die. Roughly ten metres of water column at sea level, less at elevation, less again as coolant temperature climbs and vapor pressure with it. No larger pump changes that ceiling. It sets how many racks one CDU can serve, how far it can sit from them, and how much pressure drop the cold plate design is allowed to spend. Cold plate geometry is where that budget gets consumed.
No kilowatts per rack. No CDU capacity. No approach temperature, no supply water temperature, no total showcase load, and no flow rate per rack. Every one of those numbers is on a DCIM screen inside the building, which makes the omission a choice rather than a limitation. Anyone booking a visit should collect all six on the floor, then ask the question a single populated row cannot answer on its own: what happens to per-rack flow when the second and third racks come onto the same manifold. Flow balancing is where direct-to-chip deployments go wrong after commissioning.
Expect the rest of the field to copy the format within two quarters. Vertiv, Munters, LG and Schneider all have integrated stacks and none of them can currently point a prospect at one running under a paying GPU tenant. Neocloud deals have been stalling on credit terms rather than thermal capability, which makes a neocloud willing to host a sales floor a scarce asset. The vendor that lines up an operator with no ownership or development ties to it, rather than a facility inside its own family tree, takes the reference away from Daikin.