NVIDIA Enterprise Reference Architecture: NVIDIA GB300 NVL72 with NVIDIA Spectrum-X Networking (2-4-5-800 dual-plane)
NVIDIA
This March 2026 enterprise reference architecture uses a single NVL72 rack as the scalable unit and gives cluster layouts of 2, 4 and 8 racks for GB300 NVL72 with Spectrum-X Ethernet.
On this page
Overview
The document is written for building enterprise AI clusters from GB300 NVL72 racks, with one NVL72 rack as the scalable unit. Its hardware section lists the power shelves, DC busbar, liquid cooling and leak detection, followed by worked 2-, 4- and 8-rack layouts with their associated network racks.
The numbers are compact: a full GB300 NVL72 rack requires up to 142 kW, with eight 33 kW power shelves per rack, each holding six 5.5 kW PSUs. The rack is liquid cooled (MGX architecture) with integrated tray-level and rack-level leak detection. In-rack NVLink switches draw DC from the busbar while management switches run on AC, so two kinds of power coexist inside one rack.
The glossary makes a distinction that matters for layout: an OCP Open Rack V3 uses a 48 mm OU with a DC busbar, while an EIA-310-D rack uses a 44.45 mm RU with rPDUs. In mixed white-space layouts that difference affects power distribution and floor and containment planning. The cluster layouts assume power redundancy within each cabinet; otherwise an alternative rack layout is needed. This is a product reference and does not go into facility-scale cooling or electrical infrastructure.
Key points
- A full GB300 NVL72 rack requires up to 142 kW.
- Eight 33 kW power shelves per rack, each with six 5.5 kW PSUs.
- Rack is liquid cooled (MGX) with tray-level and rack-level leak detection.
- In-rack NVLink switches draw DC from the busbar; management switches use AC.
- Cluster layouts assume power redundancy within each cabinet; otherwise an alternative layout is needed.
- OCP Open Rack V3: 48 mm OU, power shelves and DC busbar; EIA-310-D: 44.45 mm RU with rPDUs.
Using it in design
- Add the GB300 NVL72 rack to the catalog at 142 kW max, OCP ORv3 (OU) form factor, 8 x 33 kW power shelves, liquid cooled with leak detection; model ancillary network racks as EIA/AC.
- Require dual power feeds within each NVL72 cabinet.
- Flag mixed OCP/EIA rows (busbar versus rPDU) for power distribution and floor/containment planning.
- Include leak-detection integration in the BMS points list.
Monthly newsletter · free
Keep your reading list current
Once a month: new library sources, changes to Turkish regulations and upcoming data centre events.
SoSo CAD
Check cooling capacity alongside your design
Review zone heat loads, available capacity and the capacity balance after the largest unit is lost in SoSo CAD.
Request a demo
Who should read it
Designers planning rack, power and network layout for GB300 NVL72 clusters.
Related standards
- OCP Open Rack V3
- EIA-310-D
Publication details
- Authors
- NVIDIA
- Publisher
- NVIDIA
- Year
- 2026
- 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
NVIDIA (2026). NVIDIA Enterprise Reference Architecture: NVIDIA GB300 NVL72 with NVIDIA Spectrum-X Networking (2-4-5-800 dual-plane). NVIDIA. https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory-with-gb300-nvl72-dual-plane-networking-architecture.pdf
Events on this subject
- SC26 – International Conference for High Performance Computing, Networking, Storage and Analysis
- Electricity Eurasia 2027
- NVIDIA GTC
All content and rights belong to the respective authors and publishers. SoSo CAD provides only the citation, its own summary and a link to the source, and is not responsible for the sources' content.