Datacenter as a Service

Own every rack. We handle the rest.

NVIDIA Blackwell clusters from 500 kW to 50 MW in the United States and Europe, on your balance sheet and under your control. We handle procurement, colocation, power, installation, configuration and maintenance.

  • 500 kW – 50 MW
  • GB300 NVL72
  • HGX B300
  • HGX B200
  • InfiniBand

Owned hardware placed in service in the US may qualify for 100% first-year bonus depreciation under Section 168(k). How it works

Why own

Capacity that’s yours. Advantage that lasts.

Ownership pays off for teams with sustained AI workloads: AI companies, enterprises and neoclouds.

Control

Your models, your data

Weights, training data and fine-tuning stay on hardware you own, set up the way your team works.

Capacity

On your schedule

Every GPU is dedicated to your work and ready when your roadmap needs it.

Value

An asset that keeps working

As new GPU generations arrive, your systems move to inference and smaller jobs and keep earning their place.

Services

Six jobs. One operator.

Procurement

Sourcing

Allocation from NVIDIA OEMs, with every system burned in and tested before it ships.

Colocation

Space

Placement in colocation facilities in the United States or Europe, chosen to meet your data-residency needs.

Power

Capacity

Secured power for the full build, with liquid cooling for racks that draw over 120 kW each.

Installation

Build

Racking, cabling and the InfiniBand fabric, delivered as one working cluster.

Configuration

Bring-up

Firmware, operating system and cluster software, validated and handed over ready for workloads.

Maintenance

Operations

Hardware monitoring, part replacement and OEM warranty claims, handled for the life of the cluster.

Ownership

Own it alone. Or own it together.

Both models run on the same hardware with the same operating team. Only the number of owners changes.

Dedicated

One company, one cluster

Your company owns the entire cluster and makes every configuration choice.

  • An InfiniBand fabric reserved for your team
  • A software stack built to your specification
  • Expansion planned around your roadmap

Shared

One cluster, many owners

A fund and its portfolio companies, an accelerator or a consortium owns one larger cluster. Each member gets a dedicated, isolated allocation for its models, weights and data.

  • Group-scale purchasing for every member
  • Shares sized to each member’s workload
  • Every member screened to the same standard

Hardware

Three systems. One fabric.

Rack-scale

GB300 NVL72

72 Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled NVLink domain.

GPU memory
~20 TB HBM3e
NVLink
130 TB/s
Network
ConnectX-8, 800 Gb/s
Request a quote for GB300 NVL72

8-GPU node

HGX B300

Eight Blackwell Ultra GPUs per node. Scales from a few racks to a full cluster.

GPU memory
2.3 TB HBM3e
NVLink
1.8 TB/s per GPU
Network
ConnectX-8, 800 Gb/s
Request a quote for HGX B300

Networking

InfiniBand

NVIDIA Q3400-RAQuantum-X800 XDR switch, 144 × 800 Gb/s
NVIDIA MQM9790Quantum-2 NDR switch, 64 × 400 Gb/s
ConnectX-8SuperNIC, up to 800 Gb/s
ConnectX-7Adapter, up to 400 Gb/s

Hardware lead times

Hardware in weeks. Power-on on schedule.

Lead times run from confirmed order, and every proposal sets out a dated plan from order to power-on.

InfiniBand & NICs
48 h
HGX B200
2–4 wk
HGX B300
4–6 wk
GB300 NVL72
6–8 wk

Process

Specify, build, operate.

  1. Specify

    Share the capacity, systems, region, timeline and ownership model you need. We return a written proposal with configuration, site and schedule.

  2. Build

    Procurement and site work run in parallel, so your racks arrive at a facility that is ready for them.

  3. Operate

    Your team runs workloads from handover while we maintain the hardware underneath.

Export controls

Controlled hardware. Verified owners.

Our systems are subject to the U.S. Export Administration Regulations (EAR) and OFAC sanctions, and may be classified under ECCNs 3A090, 4A090, 5A002.z and 5A992.z. Every customer agrees to the restrictions below, and they extend to any end customer given access.

Destinations

Restricted destinations

No access from, or for anyone located or headquartered in, Belarus, Cuba, Iran, North Korea, Russia, Syria, or the Crimea, Donetsk or Luhansk regions of Ukraine. Destinations requiring an EAR license, which may include Country Groups D:1, D:4 and D:5 (including China, Hong Kong and Macau), are excluded unless that license is obtained.

Parties

Prohibited parties

No customer, affiliate, owner, director, officer, end customer or user may appear on the OFAC SDN List or the BIS Denied Persons, Entity, Military End-User, Military-Intelligence End User or Unverified Lists, or be 50% or more owned or controlled by a listed person.

End use

Restricted end uses

No nuclear, chemical or biological weapons, missile or UAV end use. No military, military-intelligence or WMD end use or end user, including training AI models for such use on behalf of parties headquartered in Country Group D:5.

Know your customer

Five KYC requirements on every order.

  • Identity and full ownership, including ultimate parent and headquarters
  • Intended end uses and the locations the hardware will be accessed from
  • Disclosure of any end customers; no undisclosed agents or intermediaries
  • Signed end-use statements and letters of assurance as required by us, the OEM, NVIDIA or the data center operator
  • 30 days’ written notice before any change of control

Services may be suspended immediately on reasonable suspicion of a breach, and any breach is grounds for termination without a cure period.

Quote

Every cluster is scoped individually.

For a shared cluster, one request can cover the whole group. We respond within one business day.

FAQ

Common questions.

Who owns the hardware?

You do. The GPUs, servers and networking are your company’s assets, and we operate them on your behalf.

How does a shared cluster work?

The group agrees an ownership structure up front. Your proposal then sets out each member’s share, its allocation on the cluster and the KYC each member completes.

What does Section 168(k) mean for US customers?

Because you own the hardware, a US company may be able to deduct the full cost of qualifying equipment placed in service in the United States in its first year. Eligibility depends on your circumstances, so confirm with your tax advisor.

How does payment work?

Hardware is prepaid, and payment for large orders can be held in escrow until delivery. Colocation, power and maintenance come together as one monthly service fee.

What warranty and support is included?

Every system carries the OEM warranty with next-business-day parts.