Cluster · NVIDIA · supercluster

NVIDIA DGX SuperPOD H100 (32-node reference)

32 DGX H100 nodes = 256 GPUs, 32 InfiniBand switches, one full-fat rail-optimised training pod. The AI startup unit-of-currency in 2023.

Key facts

Chips per instance: 256
Power (typical): 480 kW
Interconnect: NDR InfiniBand + NVLink Switch (400 Gb/s per GPU)
Cooling: air
Announced: 2022-03-22
GA / release: 2023-01-01

Chips inside

×256NVIDIA H100 Tensor Core GPUH100 SXM5 across 32 DGX H100 nodes

Where you'll find it

Data centers publicly running the chips inside this design. Per-rack counts are rarely disclosed, so these are presence rows: this campus runs these chips.

Stargate Norway (Kvandal / Narvik) · 100,000 NVIDIA GPUs targeted by end of 2026 (initial phase; specific SKU may be H100/H200/Blackwell mix)
Compass Datacenters Red Oak Campus (DFW III) · H100/H200-generation hyperscaler capacity at the Red Oak DFW III campus
ByteDance Pecem Data Center Campus (TikTok) · ByteDance TikTok/Doubao inference capacity in South America (H100/H200-class)
xAI Colossus (Memphis) · ~100,000 GPUs live in first cluster (Sep 2024); scaled toward 200,000 across the complex
Crusoe Childress Campus · Crusoe operates H100 GPU clusters at Childress
Meta Hyperion Data Center (Richland Parish) · Nvidia complement to MTIA for training + inference at Hyperion
Meta Prometheus Data Center (New Albany) · Nvidia complement to MTIA for training + inference at Prometheus

Vendor page: https://www.nvidia.com/en-us/data-center/dgx-superpod/

See also: all rack designs · chip catalog · NVIDIA.