Specifications
GPUs
72x NVIDIA Blackwell Ultra GPUs per rack
CPUs
36x NVIDIA Grace CPUs
NVLink domain
36x GB300 Grace Blackwell Ultra Superchips in one fabric
NVLink generation
Fifth-generation NVIDIA NVLink with 9 NVLink switch trays
Per superchip
2x Blackwell Ultra GPU, 1x Grace CPU, 2x ConnectX-8 SuperNICs
Compute trays
18 trays (10 plus 8), 2 superchips per tray
Fabric networking
NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet
Data processing
NVIDIA BlueField-3 DPUs
Management
1x top-of-rack switch plus 2x management switches
Power shelves
6x 1RU shelves, 33 kW maximum output each
Cooling — liquid-to-air
Up to 80 kW per rack, heat rejected by fans
Cooling — liquid-to-liquid
Up to 2,500 kW, supporting 10 or more AI racks
Generational gain
1.5x more AI performance than GB200 NVL72
Workload target
Test-time scaling and long-thinking reasoning inference
Sibling platforms
Vera Rubin NVL72, GB200 NVL72, HGX and MGX systems
Build origin
Ingrasys AI server lighthouse factory, since 2002
Overview
The Ingrasys NVIDIA GB300 NVL72 is a rack-scale, liquid-cooled compute unit for AI reasoning workloads, assembled by Foxconn subsidiary Ingrasys. Each rack couples 36 NVIDIA Grace CPUs with 72 NVIDIA Blackwell Ultra GPUs across 18 compute trays, linked into a single 36-superchip NVLink domain by nine NVLink switch trays.
Every GB300 Grace Blackwell Ultra Superchip carries two Blackwell Ultra GPUs, one Grace CPU and two ConnectX-8 SuperNICs, so the rack ships with an integrated east-west fabric rather than a bolt-on network. NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet handle scale-out, while BlueField-3 DPUs and dedicated management switches complete the platform.
Power and thermal design are what make the density possible. Six 1RU power shelves deliver 33 kW each, and Ingrasys offers both a liquid-to-air configuration rated to 80 kW for retrofitting existing air-cooled halls and a liquid-to-liquid configuration rated to 2,500 kW that aggregates ten or more AI racks through facility water. NVIDIA positions the platform around test-time scaling — the long-thinking inference mode behind current reasoning models.
Key Benefits
Rack-scale simplicity — GPU, CPU, NVLink, NIC and DPU layers arrive factory-integrated and validated as one unit. Density with a cooling path — the same compute tray set is offered in liquid-to-air form for existing halls and liquid-to-liquid form for new AI factories. Fabric built in — 800 Gb/s ConnectX-8 superNICs, Quantum-X800 InfiniBand and Spectrum-X Ethernet are part of the reference rack. Reasoning-optimised — 1.5x the AI performance of GB200 NVL72 aimed squarely at inference-time scaling.
Applications
Large-scale AI factories and hyperscale training clusters, reasoning and agentic inference services requiring very large NVLink domains, sovereign and enterprise AI supercomputing, model pretraining and post-training pipelines, and liquid-cooled greenfield data centre builds designed around 800 VDC distribution.
Request a Quote — INGRASYS NVIDIA GB300 NVL72 — RACK-SCALE AI FACTORY PLATFORM
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →