Vertiv 360AI High Density Reference Design #020: 6x 1.2 MW NVIDIA DGX SuperPODs (GB200 NVL72)

Overview

This 2025 design reaches 7 MW of solution capacity with six GB200 NVL72 SuperPODs of about 1.15 MW each. Total IT load is 6,912 kW = 6 x 1,152 kW. It holds 48 compute racks at 130 kW and 48 support racks at 14 kW, 108 racks in total including 12 reserved for wire management. Heat leaves 72% through direct-to-chip liquid and 28% through perimeter air.

Power follows the 4-to-make-3 pattern: groups of four UPS modules serve pairs of pods, and the design specifies four 1,200 kVA UPS. Cooling is N+1 with chillers, CRAHs and 1,350 kW CDUs. Racks carry OCP-compliant 33 kW DC power shelves (eight per compute rack); busways are 400 A (8 per pod), with two 60 A tap-offs per compute rack and two 40 A tap-offs per support rack.

The document is internally inconsistent: the design summary says 1,200 kVA UPS while the schematic shows 1,700 kVA modules and 6,000 A main switchboards, and the cooling UPS is 240 kVA 2N in the summary but 600 kVA A/B facility UPS in the schematic. Vendor values should therefore be treated as indicative and recomputed. It notes rack density may trend toward 1 MW per rack.

Key points

1. Total IT 6,912 kW = 6 pods x 1,152 kW; solution capacity 7 MW.
2. 48 racks at 130 kW + 48 racks at 14 kW; 108 racks including 12 for wire management.
3. Cooling split 72% direct-to-chip liquid / 28% perimeter air.
4. Power 4-to-make-3, cooling N+1; four 1,200 kVA UPS per pair of pods (schematic shows 1,700 kVA modules and 6,000 A MSBs).
5. Rack power shelf 33 kW DC, 8 per compute rack; 2 x 30 A rack PDUs per support rack.
6. Busway 400 A (8 per pod); 2 x 60 A tap-off per compute rack, 2 x 40 A per support rack.
7. Cooling UPS 240 kVA 2N (schematic shows 600 kVA A/B).
8. Vertiv projects AI density growing toward 1 MW per rack.

Using it in design

• Make the capacity block a first-class object: block = UPS group + PDUs + CDUs + chillers sized for N pods, so no power is stranded.
• Use the 1,200 vs 1,700 kVA conflict as a worked warning and always recompute vendor values.
• Default liquid fraction for GB200 NVL72 racks to 0.72-0.78.

Who should read it

Engineers planning multi-pod AI campuses above 5 MW.

Related standards

• OCP Open Rack V3
• Uptime Tier Standard: Topology
• ASHRAE TC9.9 Liquid Cooling Guidelines

Publication details

Authors

Vertiv

Publisher

Vertiv

Year

2025

Document number

AI Reference Design #020

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation