COMPUTE ATLAS Supercomputer supply-chain graph
1540 systems 545 sites Sourced data

Parts/Accelerators/NVIDIA GB200 NVL72

Accelerators · Blackwell · 2024

NVIDIA GB200 NVL72

Rack-scale unit of 72 Blackwell GPUs and 36 Grace CPUs in a single NVLink domain, liquid cooled.

Why this part

The unit of purchase is now a rack. Seventy-two Blackwell GPUs and thirty-six Grace CPUs sit inside one NVLink domain, so software addresses the whole rack as a single accelerator with 13.4TB of high-bandwidth memory. That reframes what a node is, what an interconnect is for, and what a data centre has to supply: roughly 120kW into one rack, which almost no existing facility was built to deliver. The infrastructure constraint has moved from the chip to the building.

Systems using this part

Systems
14
Rmax underneath
27.37 PFlop/s
Units recorded
64,752
5 of 14 edges disclose a quantity
System Role Country Units System Rmax Supplied by Rank
Microsoft Fairwater (Wisconsin)
2026-06
Accelerator USA - - NVIDIA -
xAI Colossus 2
2026-01
Accelerator USA - - NVIDIA -
KDDI GPU Cloud
2026-01
Accelerator Japan - 22.21 PFlop/s NVIDIA -
YTL AI Cloud (Kulai)
2025-10-31
Accelerator Malaysia - - NVIDIA -
Stargate Abilene
2025-06
Accelerator USA - - NVIDIA -
Together AI / Hypertec Cloud GB200 Cluster
2025-01
Accelerator USA 36,000 - NVIDIA -
Nano 4 GB200
2025
Accelerator Taiwan - 5.16 PFlop/s NVIDIA -
Microsoft Fairwater (Atlanta)
Accelerator USA - - NVIDIA -
Horizon
Accelerator USA - - NVIDIA -
Mistral Campus AI (Bruyeres-le-Chatel)
Accelerator France 18,000 - NVIDIA -
Hon Hai Kaohsiung Super Computing Center
Accelerator Taiwan 64 - NVIDIA -
Carina
Accelerator USA 2,496 - NVIDIA -
Tyche-hsg
Accelerator USA 8,192 - NVIDIA -
CoreWeave Vennesla (Bulk N01)
Accelerator Norway - - NVIDIA -

“Units recorded” sums only the edges where a public source states a quantity. It is a floor, not a total, and should never be read as installed base.

Further reading