The first useful question about an AI cooling purchase is where the heat will go after it leaves the processors. A qualified coolant distribution unit can manage an important part of that journey. It cannot, by itself, establish that a building will support the intended accelerator load. Vertiv's NVIDIA DSX Ready qualification is best understood as a starting point for that wider engineering review.
In its September 21, 2026 announcement, Vertiv described a 2.3 MW CoolChip CDU qualified through NVIDIA's program. The company calls it the first CDU qualified specifically through that program. The unit supports direct-to-chip cooling and rear-door heat exchangers, transferring heat from a server-side loop to facility water. Its position at the end of a row or building perimeter allows it to serve several racks.
Translate the rating into operating conditions
The 2.3 MW label is a thermal rating under specified conditions. It does not represent installed computing capacity or a universal rack count. Ask for performance curves at the planned supply and return temperatures, coolant mixture, flow and pressure. Pipes, valves and cold plates introduce resistance; a pump must deliver the necessary flow against that actual system.
NVIDIA's reference guidance describes a liquid-to-liquid design, at least 1.5 liters per minute per kilowatt of technology-loop flow, and N+1 CDU redundancy. These are design assumptions to examine against the selected platform. They do not eliminate local engineering or capacity margins.
Test the complete thermal path
A CDU separates facility water from the cleaner loop serving computing equipment, but both loops still require compatible materials, treatment, filtration and maintenance. The Vertiv product family describes redundant variable-speed pumps, filtration and leak detection. Buyers should establish which exact model and configuration provide each feature.
Heat still needs somewhere to go beyond the CDU. Chillers, dry coolers or cooling towers must work at the site's climate and water constraints. A strong component cannot compensate for undersized heat rejection. Total efficiency also depends on that full chain, rather than pump or fan savings alone.
Verify failure behavior and maintenance
N+1 is useful when the remaining equipment actually takes over the load during a failure. Commissioning should exercise pump loss, communications errors, electrical backup, alarm escalation and isolation for maintenance while checking server coolant limits. Counted spare units are insufficient evidence without an observed transition.
Service access belongs in the same review. Technicians need to replace pumps, clean filters and reach sensors without interrupting unrelated racks. Ask which supplier owns diagnosis when a server cold plate, facility-water condition and CDU controller interact. Written responsibility matters especially in a multi-vendor design.
Compare systems rather than badges
The qualification announcement uses NVIDIA's self-qualification process. It supplies a common screening framework, not independent approval of every site. Competing LG equipment and Schneider Electric's DSX work provide useful comparison points without making nominal megawatts the only criterion.
A practical purchasing decision combines matched performance data, a failover test plan, water requirements, maintenance space, service commitments and headroom for server refreshes. Named deployments and repeat orders can strengthen confidence later. For the current design, the decisive evidence remains a commissioned thermal system that delivers the planned load under the site's own conditions.