Skip to the main content

Platform library · shipping

What a NVIDIA GB300 NVL72 rack asks of an existing hall

Each figure below is the manufacturer’s own published value, used unchanged by the engine that screens halls. The sources are listed at the end. Where a figure was not published it says so; nothing is estimated to fill a gap.

Rack power
135 kW (peak 155 kW)
GPUs per rack
72
Captured by liquid
90%
Residual air load
13.5 kW
Maximum liquid inlet
45 °C
Technology-loop flow
130 l/min per rack
Rack mass
1,400 kg on 0.72 m²
Floor loading
1,944 kg/m²
Rack feed
205 A
Power domain
200-277 VAC in, 50 VDC out, 6-8 x 33 kW shelves

Arithmetic on those figures

Plain arithmetic on the published values above, shown so you can check it. It is not a measurement of any hall.

  • Liquid-captured heat per rack: 135 kW × 90% = 121.5 kW.
  • Loop flow per liquid-captured kW: 130 l/min ÷ 121.5 kW = 1.07 l/min per kW.
  • Mass over footprint: 1,400 kg ÷ 0.72 m² = 1,944 kg/m².
  • Whole racks in 1 MW of IT load at the published rack power: 7. This is IT load only, before any cooling or distribution overhead, and says nothing about whether a given hall can feed them.

Which of a hall’s constraints it presses on

  • rack feed tapoff rating: Each rack draws its feed current continuously; the tap-off box must be rated for it.
  • busway ampacity: The busway carries the sum of every rack's feed on its run.
  • cdu capacity site conditions: The coolant distribution unit has to move the liquid-captured heat at the hall's real facility-water temperature.
  • residual air load removal: The share not captured by liquid still has to be removed by the hall's air system.
  • floor loading: Rack mass over its footprint is a load the slab or raised floor has to carry.

Which one binds first in a particular hall depends on that hall’s numbers. The free qualifier finds out.

Sources

  • Lenovo Press LP2357, GB300 NVL72, 2026 [V]
  • NVIDIA NVL72 AI Factory Reference Architecture, 2026 [V]

Vendor-sourced values are marked as such in the provenance of every study they appear in. Platforms whose rack power is unpublished are deliberately absent from the library; see the platform library.