Size the cooling for an AI GPU rack in seconds. Pick a rack build or enter kW per rack, set how many racks, and get the cooling capacity you need, an air-vs-liquid recommendation, CDU sizing and the equipment that fits — for H100/H200 HGX, B200 and 100 kW+ GB200 NVL72 racks.
Cooling is the constraint that decides whether an AI rack can even be deployed. Modern GPU racks concentrate more heat into one footprint than an entire traditional row, so before you rack a single server you need to know how much heat you must reject and whether air can do it or you need liquid. This tool turns a rack build into that answer.
It starts from a simple physical fact — essentially all electrical power a rack draws becomes heat — so the cooling load is the rack IT load. From there it scales to the room and picks the cooling approach by power density:
Rack density is what forces the move from air to liquid. A standard enterprise rack sits around 5–10 kW and is comfortably air-cooled. An H100 HGX rack lands near 40 kW; a B200 liquid rack around 80 kW; a GB200 NVL72 concentrates 72 GPUs into roughly 120 kW and a GB300 NVL72 pushes ~140 kW. Air simply cannot reject that much heat from one rack — which is why direct-to-chip liquid cooling is now standard for the highest-density AI.
Between the extremes, rear-door heat exchangers extend air-assisted cooling to about 40 kW/rack by rejecting heat to facility water at the back of the rack, with no plumbing into the servers. Past that, a coolant distribution unit (CDU) isolates a clean technical-water loop for the direct-to-chip cold plates and is sized with N+1 headroom so a unit can fail without dropping the rack.
Getting the approach right early is what lets you lock the longest-lead cooling plant — CDUs, chillers and rear-door units — before it gates the whole build. That is exactly the sizing this tool is built to accelerate, and the recommended equipment links straight into our catalog.
The density thresholds and cooling approach the calculator uses. Figures are planning bands used across the industry; confirm exact limits against your rack, coolant temperatures and ambient conditions.
| Density | Approach | Typical racks |
|---|---|---|
| ≤ 15 kW | Air cooling (CRAC / CRAH) | Standard enterprise |
| 15–40 kW | Rear-door heat exchanger | H100 / H200 HGX |
| 40–80 kW | Direct-to-chip liquid (CDU) | B200 liquid |
| > 80 kW | Integrated direct-liquid + CDU | GB200 / GB300 NVL72 |
Because virtually all electrical power a rack draws is converted to heat, a 100 kW GPU rack needs roughly 100 kW of heat-rejection capacity. At that density air cooling is not viable — you need direct-to-chip liquid cooling fed by a coolant distribution unit (CDU), sized with N+1 headroom so a single unit can fail without losing the rack.
Yes. A GB200 NVL72 rack integrates 72 Blackwell GPUs at roughly 120 kW per rack and ships with integrated direct-to-chip liquid cooling. There is no air-cooled configuration at that density — the facility must provide the chilled/technical water loop and CDU capacity to match.
Rack density is the deciding factor. Up to ~15 kW/rack, room air (CRAC/CRAH) is enough. From ~15–40 kW/rack a rear-door heat exchanger rejects heat to facility water with no server plumbing. Above ~40 kW/rack you need direct-to-chip liquid cooling with a CDU; above ~80 kW/rack (GB200-class) cooling is integrated liquid by design.
It depends on kW/rack and your facility. Air is simplest and cheapest but caps out around 15–20 kW/rack. Rear-door heat exchangers extend air-assisted cooling to ~40 kW. Direct-to-chip liquid is required for modern dense AI (40 kW+), is far more efficient (lower PUE), and is mandatory for GB200/GB300. This tool recommends the approach for your density and links the equipment.
It is a fast planning estimate using published rack loads and standard density thresholds — accurate enough to scope cooling capacity, choose air vs liquid, and budget CDU/chiller plant. Final mechanical design depends on your exact servers, coolant temperatures, ambient conditions and redundancy target. Our engineers confirm the real numbers when you request a quote.
Our solutions engineers size cooling, power and compute together and confirm real lead times — no payment, no commitment, quotes back in about one business day.
Our engineers scope power, cooling and compute together. No payment, and quotes come back in about a business day.