Creative Compute
0

HPE AI Factory at Scale

Rack-scale systemProject engagement onlyManufacturer documented

The largest tier of HPE's AI Factory portfolio: liquid-cooled rack-scale systems built around NVIDIA Grace Blackwell and successor platforms, for training and serving models at the frontier of what is currently possible.

The problem it solvesYour models are large enough that the constraint is no longer which GPU to buy, but how much power and heat a room can take.

Ask Carl about this

No payment is taken online. Every request is an enquiry answered by an engineer.

Representative system image

Where does this sit?

  1. 01Local
  2. 02Workstation
  3. 03Departmental
  4. 04Enterprise
  5. 05Cluster
  6. 06Rack-scale

Rack-Scale & Supercomputing. Integrated racks where compute, interconnect, storage, power and liquid cooling are engineered as a single system. Typically bought by a datacentre operator, a national facility or a large ai platform. Datacentre with liquid cooling and high-density power.

Fit

Is this the right thing for you?

Who it's for

  • AI cloud and platform operators
  • Large enterprises building an internal AI capability at scale
  • National and academic supercomputing facilities

What it runs

  • Frontier model training above one trillion parameters
  • Large multi-tenant inference platforms
  • Converged HPC and AI workloads

When it is the wrong answer

  • Anyone who has not yet run a production workload on a single eight-GPU server
  • Facilities without liquid cooling or high-density power
Where it goes
Datacentre with direct liquid cooling and high-density power
Complexity
specialist
Cooling
Liquid cooled
Specification

What the numbers mean

Every figure below is explained in plain English. Switch to the technical view for the bare specification.

Available rack-scale platforms

NVIDIA GB200 NVL72 by HPE

Shipping — HPE announced first shipment February 2025

MeansA 72-GPU NVLink domain in a single liquid-cooled rack.

NVIDIA GB200 NVL4 by HPE

Available — a smaller four-GPU NVLink module for converged HPC and AI

MeansThe entry point into rack-scale NVLink without committing to a full rack.

NVIDIA GB300 NVL72 by HPE

Available — 72 Blackwell Ultra GPUs and 36 Grace CPUs, fully liquid-cooled

MeansAimed at models above one trillion parameters and AI reasoning workloads.

MattersHPE cites 1.5x denser FP4 throughput and 2x attention performance against Blackwell.

NVIDIA Vera Rubin NVL72 by HPE

Announced — general availability date not published

MeansOn the roadmap, not orderable to a date.

WhenOnly relevant if your project timeline is a year or more out.

Facility

Cooling

Direct liquid cooling

MeansHeat is carried away by liquid rather than air.

MattersMost existing enterprise datacentres cannot do this without building work.

Rack power drawUnverified

Awaiting verification

MeansDepends entirely on the configuration; HPE quotes this per design.

Verification

Manufacturer documented — HPE AI rack-scale systems page, GB300 NVL72 product page and QuickSpecs, HPE/NVIDIA March 2026 press release. Reviewed 2026-08-31.

Manufacturer source
Before you commit

Practical considerations

The things that catch people out after the hardware has already been ordered.

  • Lead times at this scale are measured in quarters and are allocation-driven. The hardware conversation is rarely the long pole.
  • The facility work almost always costs more and takes longer than expected. Start it in parallel.
  • HPE Vera Rubin NVL72 is announced rather than shipping; do not build a plan around a date nobody has published.
Completeness

Nothing arrives working on its own

What else will I need?

  • Direct liquid cooling capability, or a plan and budget to install it
  • A power feasibility study for the building, not just the rack
  • Structural loading confirmation for the floor
  • An operations team, or a managed service arrangement
AI networkingAI storageInstallationAsk Carl what I’m missing
How this is bought

This is a project, not a checkout

Solutions at this scale are designed before they are priced. Here is the sequence.

  1. 01

    Requirement conversation

    What you want to run, for how many people, on which data. No hardware discussed yet.

  2. 02

    Site and power review

    Rack space, power feeds, cooling and network capacity checked against what the platform needs.

  3. 03

    Architecture proposal

    A written design covering compute, networking, storage, software and services, with alternatives.

  4. 04

    Quotation

    Priced against the agreed design, including installation and support. Nothing is charged until you accept.

  5. 05

    Delivery and commissioning

    Racked, cabled, commissioned and handed over with documentation.

Talk it through with Carl first
Goes with

What normally sits alongside it

Relationships documented by the manufacturer, or by us during a deployment.