The project: a 2 MW GPU training hall for a mid-size AI infrastructure operator, deploying H100-class servers at 40 kW/rack. Every downstream step in this series is built against the cooling decision this step makes.
At 40 kW/rack, this facility sits well past the practical ceiling of computer room air handlers (CRAH) alone — a well-designed hot-aisle-contained CRAH system tops out in the 15-20 kW/rack range before airflow volume and fan power become impractical. The design decision at this stage is not whether to use liquid cooling, but how to split the load between liquid and air: this project uses direct-to-chip cold plates to remove roughly 80% of each server's heat (the CPUs and GPUs themselves) via a liquid loop, leaving the remaining 20% (memory, VRMs, drives, fans) to conventional CRAH-cooled room air. That 80/20 split is a realistic, commonly-cited starting assumption for current-generation GPU servers — Step 2 of the residential-style Manual J approach doesn't apply here, but the same principle does: get the split right before sizing anything downstream.
The liquid side routes through a Coolant Distribution Unit (CDU) — the interface between the facility's chilled-water loop and the sealed, closely-controlled liquid loop that actually touches the cold plates on each GPU. This isolation matters for two reasons: facility chilled water is treated for corrosion/scale control at a very different chemistry than what's safe against the cold plates' fine channels, and the CDU's secondary loop can run at tighter temperature setpoints than the facility loop without over-cooling (and condensing on) the rest of the room. For this 2 MW hall, the design uses row-based CDUs (one CDU per 4-6 racks) rather than a single hall-level CDU — this bounds the blast radius of a CDU failure to a handful of racks instead of the entire hall, a redundancy consideration Step 4 will size formally.
The remaining 20% air-cooled load still needs a real air-side design — hot aisle containment, adequate CRAH capacity for that reduced (but still substantial, roughly 400 kW total) load, and airflow rates sized to ASHRAE TC9.9 Class A2/A3 guidelines for the room's target inlet temperature. Getting the air-side sizing right here is what makes Step 3's UPS/generator sizing accurate — CRAH fan motors and controls are real electrical load too, not just a cooling-side concern.