The same hardware behaves very differently depending on what the site is for. These are the four deployment patterns we see most often and how we approach each.
AI and HPC training clusters
Eight accelerators per node, 130 kW or more per rack, and a load profile that steps from idle to full in under a second. These sites need direct-to-chip or immersion — nothing else reaches the density — and they need a CDU with genuine temperature stability, because a supply temperature that wanders shows up directly as clock throttling.
We size these projects around the training job, not the nameplate. Peak simultaneous draw across a pod is the number that matters.
Colocation
The constraint is commercial rather than thermal. Tenants arrive with hardware you did not specify, sign for a power envelope, and expect to swap chassis without notice. Rear door heat exchangers work well here because they are tenant-agnostic: whatever goes in the cabinet, the door removes its heat.
Where tenants demand higher density, our rack fluid network is designed so that no rack units are consumed and any single rack can be isolated and drained without touching its neighbours.
Enterprise data centres
Usually a retrofit into a room built for 5 kW racks, with a live production load that cannot stop. These projects are staged: a passive door row first to prove the water loop, then a direct-to-chip pod once the operations team is comfortable, then expansion.
The engineering challenge is rarely the cooling. It is sequencing the work so the room never drops below its redundancy commitment.
Edge and constrained sites
No chilled water, no roof plant, no mechanical contractor. The KX-CDU A120 puts the entire loop inside the rack and rejects to room air, so a dense cabinet needs a power connection and nothing else. It is not the route to a hundred-kilowatt hall, and we will say so, but it removes every civil dependency from a single-rack deployment.