A GPU cluster in
your own data center.
We design the cluster around what your room can take: power per rack, cooling and floor loading. Then we build it on site, and your team or ours runs it.
- Site
- Your data center
- Owner
- You
- Build
- Lightyard
- Operator
- Your team or us
Sized to what the room can take.
We start from measurements, not from a catalog. The cluster is designed to fit the room, and any upgrade the room needs is listed before you commit to hardware.
- Power at the rack positionsWhat the room can deliver to each position decides how many nodes or racks go in.
- Air or liquidHGX B300 nodes can run on room air. GB300 NVL72 racks need liquid cooling, where the room allows it.
- The floor and the route inFloor loading, and the path from the loading dock to the room, are checked for GPU racks.
The cluster stays inside your perimeter.
For enterprises, sovereign programs and operators with a hall already running. The hardware sits in your building, under your access and security rules.
- Access
- Your rules
- Operator
- Your team or us
- Your access rulesOur engineers work on site under your badge, escort and safety procedures.
- Data that can't leaveThe cluster sits in your data center. What connects to it is your decision.
- Run by your team or oursYour staff can run it after handover, or we operate it to an SLA in the contract.
What we measure before we design.
The assessment comes first and is written up for you. It decides the size of the cluster and lists the upgrades the room needs, if any.
- 01Power
- Supply to the room
- Distribution to rack positions
- Redundancy
You get Power available per rack position
- 02Cooling
- Air capacity
- Liquid cooling options
- Heat rejection
You get Which platforms the room can cool
- 03Structure
- Floor loading
- Raised floor or slab
- Route from the dock
You get Where GPU racks can stand
- 04Network
- Paths to the room
- Uplinks
- Space for the RoCE v2 fabric
You get A network plan for the cluster
From a site survey to a running cluster.
Step 1: Survey the room
We measure power, cooling, structure and network paths on site.
You getA written assessmentStep 2: Design to fit
The cluster is sized to the room, with any upgrades listed line by line.
You getA design and a bill of materialsStep 3: Build on site
We rack and cable the nodes, configure the fabric and burn in each one inside your room.
You getA burned-in clusterStep 4: Hand over or operate
Your team accepts it and runs it, or we keep running it under an operations contract.
You getA cluster in service
How an on-prem cluster can be set up.
On your own site you own the cluster. We can build it and hand it over, build it and run it, or run one you already have.
| Way to work with us | Owner | Design | Source | Build | Operate | Software |
|---|---|---|---|---|---|---|
| Cluster deploymentWe design, source and build it, then hand it over burned in. Your team runs it. | You | |||||
| Managed clusterYou own it, we run it. We can build it first, or take over a cluster that's already live. | You | |||||
| SoftwareYour team runs every stage, on our stack. | You | |||||
| Reserved clusterWe own and run it. You reserve it for 60 months, with rollover into the next generation. | Lightyard |
Organizations with a data center of their own.
03Enterprises & sovereign
Private AI inside your own building.
You bring the room and the rules. We design and build to both.
Private and sovereign AI05Asset owners
GPU density in a hall you already run.
You bring the hall. We size the cluster to it and can operate it.
Clusters for asset owners
If the room isn't enough.
When the room can't take the load, a retrofit of another building, a modular site or a colocation hall can carry the cluster instead.
- 02 On-prem · this page
Your data center, with the cluster designed around it.
Questions we get on the first call.
Something missing? Ask us directly. We usually reply within a few hours.
We list the upgrades it needs. If they're large, a retrofit or a modular site on the same land may fit better.
Only for NVL72 racks such as GB300 NVL72. HGX nodes such as HGX B300 are air-cooled, so many rooms can take them as they are.
Yes. We review it first, then design the cluster around it, or source new hardware alongside it.
Our engineers rack, cable and burn in on site, under your access and safety rules.
Yes. We operate it to an SLA in the contract and agree the on-site work, and who does it, per project.
No, not through us. The cluster sits inside your data center, and what connects to it is your decision.
Access to the room, the site's drawings and power data, and a contact on your facilities team.
Solutions, sites and who we build for.
- AI factoryEvery stage, from the site to a running cluster.
- Cluster deploymentDesigned, sourced, built and handed over burned in.
- Managed clustersYou own the cluster. We build it or take it over, and run it.
- Reserved clustersWe own and run it. You reserve it for the term.
- SoftwareRun every stage yourself, on our stack.
- White-label GPU cloudYour GPU cloud under your brand, built and run by us.
- ColocationA Tier III facility in the region you choose.
- On-premYour data center, with the cluster designed around it.
- RetrofitA building or mining site upgraded for GPU density.
- ModularA prefabricated site on land with power.
- NeocloudsCapacity built to your standard, ready to sell.
- Venture capitalOne cluster, shared by your portfolio companies.
- Enterprises & sovereignYour own cluster in your region.
- Private equityGPU capacity on portfolio sites, or a neocloud as a platform.
- Asset ownersHardware, a hall or a power site, turned into a running cluster.
- InvestorsAn operated GPU asset, without building an operator.