Uvation MarketplaceMarketplace

NVIDIA GB300 NVL72
AI Data Center

One fully liquid-cooled rack with 72 Blackwell Ultra GPUs and 36 Grace CPUs, delivering up to 1,080 petaFLOPS of dense FP4 compute.

72
GPUs
20 TB
GPU memory
130
TB/s NVLink

An AI Factory Built as One Rack

Built on NVIDIA Blackwell Ultra, GB300 NVL72 links 72 GPUs and 36 Grace CPUs into one rack, scaling reasoning, inference and training efficiently.

72 GPUs Linked as One Giant GPU

Seventy-two Blackwell Ultra GPUs share 130 TB/s of NVLink bandwidth inside one fully liquid-cooled rack.

Massive Memory and Bandwidth for Giant Models

20 TB of GPU memory and 17 TB of LPDDR5X CPU memory keep data close to compute for larger models.

Built for the Era of AI Reasoning

Up to 1,440 petaFLOPS of sparse FP4 and 720 petaFLOPS of FP8/FP6 accelerate reasoning, long-context inference and training.

72 GPUs. One NVLink Fabric.
One Rack.

Fifth-generation NVLink connects all 72 GPUs and 36 Grace CPUs into one unified memory and compute domain.

NVIDIA GB300 NVL72 rack in a darkened data hall
72× GPUs
NVIDIA Blackwell Ultra GPUs
36× CPUs
NVIDIA Grace, 2,592 Arm cores
20 TB
GPU memory
130 TB/s
Total NVLink bandwidth

A platform for models that refuse to fit the old shape.

From frontier training to real-time reasoning, GB300 NVL72 runs demanding AI workloads at data-center scale.

AI Reasoning and
Real-Time Inference

1,080 petaFLOPS of dense FP4 compute across 72 Blackwell Ultra GPUs keeps reasoning and inference responses fast.

Large Language Model Training at Scale

720 petaFLOPS of FP8/FP6 compute and 20 TB of GPU memory help teams train larger models in less time.

Higher AI Factory
Output at Scale

37 TB of fast memory, including 17 TB of LPDDR5X CPU memory, keeps large AI workloads within one rack.

Engineered to dominate

Rack-Scale AI Power

GB300 NVL72 combines Blackwell Ultra compute, NVLink bandwidth and 800 Gb/s networking per GPU to deliver peak performance for AI factories.

Exploded view of the NVIDIA GB300 NVL72 rack
1.44 EF
FP4 compute (sparse)
720 PFLOPS
FP8/FP6 Tensor Core (sparse)
1,080 PFLOPS
FP4 compute (dense)
800 Gb/s
Networking per GPU
Exploded view of the NVIDIA GB300 NVL72 rack
1.44 EF
FP4 compute (sparse)
720 PFLOPS
FP8/FP6 Tensor Core (sparse)
1,080 PFLOPS
FP4 compute (dense)
800 Gb/s
Networking per GPU

The rack, itemized

SUBSYSTEM

SPECIFICATION / CONFIGURATION DETAILS

Processor

36× NVIDIA Grace CPUs; 2,592 Arm Neoverse V2 cores

Operating System

Not specified (NVIDIA DGX GB300 system specifies NVIDIA DGX OS / Ubuntu )

Chipset

NVIDIA GB300 compute tray with 2× GB300 Grace Blackwell Ultra Superchips

Accelerators / GPU

72× Blackwell Ultra GPUs; 1,440 sparse / 1,080 dense PFLOPS FP4; 720 PFLOPS FP8/FP6; 130 TB/s aggregate NVLink bandwidth

Memory

37 TB total fast memory (20 TB GPU HBM3e memory; 17 TB LPDDR5X CPU memory)

Primary Storage

Not specified

Secondary Storage

Not specified

Security

Not specified

Chassis

Fully liquid-cooled rack-scale architecture

Power Supply

8 power shelves of 33 kW (six 5.5 kW PSUs each); up to 142 kW per rack

Ports

ConnectX-8 connectivity up to 800 Gb/s per GPU; supports Quantum-X800 InfiniBand or Spectrum-X Ethernet

Slots / PCIe Architecture

9 × L1 NVIDIA NVLink switch systems; fifth-generation NVLink

Bezel

Not specified

Form Factor

Rack-scale 72-GPU NVLink system

Network Card

NVIDIA ConnectX-8 SuperNIC

Warranty

N/A

Front view of the NVIDIA GB300 NVL72 rack

Memory, Fabric and
Networking at Rack Scale.

37 TB of fast memory per rack

20 TB of GPU memory and 17 TB of LPDDR5X CPU memory give 72 GPUs and 36 CPUs one fast-memory pool.

800 Gb/s per GPU via ConnectX-8 SuperNICs

Each GPU gets its own 800 Gb/s ConnectX-8 SuperNIC, supporting Quantum-X800 InfiniBand or Spectrum-X Ethernet for scale-out.

130 TB/s NVLink fabric across 72 GPUs

One NVLink domain links all 72 GPUs at 130 TB/s, so data moves between GPUs without crossing the data-center network.

Nuclear-Ready, Plug and Play AI Factories

Bypass grid constraints with self-contained, high-density inference ready infrastructure directly adjacent to clean, base load nuclear power.

Global Deployment Engine

Any Infra, Anywhere in the world.

Global Shipping and Deployment Capability treats advanced infrastructure as mobile, modular freight.

Packaged complex computing nodes, cooling manifolds, and power distribution units into certified, standard form factors.

360° Service Ecosystem

Unifying Legacy with Next-Gen Compute.

Vertically integrated service ecosystem that serves as an operational bridge between legacy environments and next-generation, AI-native infrastructure.

Finance and Leasing

Capital-as-a-Service for High-Density Compute

Convert upfront capital expenditures (CapEx) into predictable operating expenses (OpEx).

Uvation allows enterprises and sovereign entities to deploy multi-megawatt AI factories and advanced GPU clusters without liquidity constraint.