GB300 NVL72
72×GPUs
Blackwell Ultra + Grace, rack-scale
- + 37 TB fast memory
- + 130 TB/s NVLink fabric
ENTERPRISE · AI FACTORIES
From one rack to an AI factory.
We spec, source and integrate datacenter-class AI compute, and deliver it as racks that boot, not boxes you have to assemble.
NET · TOPOLOGY◐ LIVE◐ DRAG TO ROTATE
Current-gensilicon, authorized channels
Rack-scaleintegration & fabric
Multi-dayburn-in as standard
On-siteracking, install, power-on
THE SILICON
SPECS · PUBLIC NVIDIA DATARack-scale systems we deliver
DGX B200
1.44TB HBM3e
8× Blackwell, one system
- + 72 PFLOPS FP8 training
- + Node → multi-rack pod
HGX B300
2.3TB HBM3e
8-GPU Blackwell Ultra baseboard
- + 5th-gen NVLink fabric
- + Into your OEM chassis
WHAT SHIPS
HGX-CLASS GPU RACK · SCHEMATICA rack, not a pile of boxes.
RACK ELEVATION42U
FABRICSPINE SWITCH
FABRICLEAF · NVLINK
COMPUTEGPU NODE
COMPUTEGPU NODE
COMPUTEGPU NODE
COMPUTEGPU NODE
CONTROLHEAD NODE
COOLINGDLC MANIFOLD
POWERPDU ×2
SEALED · CABLED · TESTED
Compute, fabric, cooling and power land pre-integrated. We plumb the liquid loop, cable the fabric, image the nodes and burn it in, so what you get is a rack that powers on.
- Compute4× HGX nodes, 32 Blackwell GPUs
- FabricNVLink + InfiniBand NDR, non-blocking
- CoolingDirect-to-chip liquid, factory-plumbed
- Power3-phase 415V feeds, N+1 redundant PDUs
FACILITY & FABRIC
POWER · COOLING · NETWORK · SOFTWAREDatacenter-class by default.
415V feeds · N+1 PDUs · liquid-ready to ~132 kW/rack
Direct-to-chip DLC + rear-door, air-assist fallback
NVLink · InfiniBand NDR · 400 GbE spine-leaf
Drivers · Slurm / Kubernetes · multi-day burn-in
THE SOFTWARE
SIGNED · IMMUTABLE · OTABoots into a working cluster.
Every rack ships a signed, reproducible image with drivers, CUDA, scheduler and telemetry already wired. There's no day-one setup: power on, the self-test passes, you allocate.
- OSUbuntu LTS · hardened, CIS-benchmarked kernel
- DRIVERSNVIDIA · DOCA · fabric manager
- RUNTIMECUDA · cuDNN · NCCL · containerd
- SCHEDULERSlurm · Kubernetes, multi-tenant ready
- OBSERVABILITYPrometheus · Grafana · NVIDIA DCGM
CONTROL PLANE
PROVISION · MONITOR · METEROne console for the whole fleet.
Every deployment ships wired into one console. Provision nodes, watch every GPU, meter usage and power. It runs on the stack you already trust: Slurm, Kubernetes, Prometheus, Grafana and DCGM, and it's yours to keep.
Zero-touch provisioningSigned order to live cluster, no ticket queues.
One schedulerTraining and inference on the same fabric, one API.
Single-tenantDedicated compute, network and storage. No neighbours.
Rack-level telemetryLive metrics per rack, node, GPU and job.
Fleet overview updated 14s ago
+ DeployGPUS1,024
POWER16 MW
PUE1.12
JOBS38
RUNNINGcluster-01H200 · 1,024util 96%
RUNNINGcluster-02H100 · 256util 78%
DEPLOYINGcluster-03H200 · 512burn-in · eta 34m
POWER DRAWlast 24h · 16 MW
HOW WE BUILD
Racks that boot. Not boxes.
- 01PLANPower, cooling & rack density
- 02CONFIGURECompute matched to your models
- 03FABRICNVLink · InfiniBand · 400G
- 04IMAGEDrivers, scheduler, orchestration
- 05BURN-INMulti-day, documented, then ship
DEPLOYMENT SIZES
Start where you are
Desk
One engineer, serious models
- DGX Spark / RTX PRO
- Ships from stock
- Self-serve checkout
MOST COMMON
Rack
A team running inference and fine-tuning
- 4 to 8 GPU nodes
- Single-rack fabric
- Specced and burned-in
- Quote on request
AI Factory
Production training at scale
- NVL72 pods
- Multi-rack fabric
- On-site racking & install
- Dedicated engineer
NEXT STEP
Planning serious compute?
Tell an engineer what you're training or serving, and you'll get a spec and a straight answer.
Talk to an engineer →