Your own GPU cluster.
With a way out.
Dedicated NVIDIA Blackwell clusters, reserved for five years and run for you. Keep an exit option, or roll into the next generation when it ships.
- Term
- 60 months
- Minimum
- 8 nodes or one NVL72 rack
- Exit option
- In the contract
- Rollover
- Next generation
A new GPU generation every year, on contracts that run five.
Buying five years of capacity has meant betting on one generation of hardware. Every contract carries an exit option and a rollover into the next generation, so a five-year rate doesn't tie you to one generation.
How the contract is structured.
Commit to what you need now, keep a way out, and move to new hardware when it ships.
Start at the minimum, add capacity in waves
Reserve 8 nodes or one NVL72 rack, or more. Larger clusters are handed over in monthly waves of 32 nodes, each wave burned in before handover.
Keep a way out
If your plans change, the contract has an exit option. Your proposal sets the terms.
Roll into the next generation
When new silicon ships, part of your commitment can move to it instead of staying on older hardware.
Bare metal or managed
Take it as bare metal, with root on every node, or managed, with us also running the software layer on top. Same single-tenant hardware either way, and either way we monitor it, replace failed parts and run the cluster to an SLA.
Who reserves a cluster.
Dedicated capacity in your region, at a long-term rate, on current GPUs.
- Capacity you don't have to findReserved for you for the whole term, with no waiting for allocation on a shared cloud.
- Margin on every GPU-hourA long-term rate, lower than on-demand and fixed for the term, sized and placed where you choose.
- Current hardwarePart of the commitment can roll into the next generation when it ships.
An operated compute asset, with a path to newer GPUs.
- Operated for youWe run the cluster to an SLA. You resell the capacity through your own channels or partners.
- Rollover to newer GPUsRollover moves part of the commitment to new silicon instead of leaving you on old hardware.
- Escrow and transferPrepayment is held in escrow with an independent agent, and the contract can be transferred with our consent.
Committed supply, delivered ahead of your commitments.
- Supply on a scheduleCover a commitment or a slipped delivery with capacity handed over in dated waves.
- A cost base you can planOne 60-month commitment under capacity you resell by the hour or the month.
- Root on every nodeRun your own scheduler, images and billing on single-tenant hardware. Firmware and BMC stay with us.
A dedicated cluster in your region, with an exit option and rollover.
- No data center to runDedicated hardware in a Tier III facility, operated for you. Or on your own site.
- Exit or roll overExit, or roll over when the next generation ships.
- Your regionCapacity placed in the region your residency rules require.
Regions and facilities.
Tier III facilities across the US, Europe, Asia and South America. Your cluster stays in the region you choose.
A cluster in a module of its own.
A cluster can run in a prefabricated module: racks, power, liquid cooling and network in one enclosure. Whether yours does depends on the region, and the proposal says which.
- Single tenant per moduleIts own power, cooling and network, with no other customer's racks inside.
- Liquid-cooledCoolant to every rack and dry coolers outside, sized for NVL72 racks.
- Room for anotherWhere the pad has space, added capacity can be a second module beside the first.
The contract terms, summarized.
Exit, rollover, prepayment and transfer, summarized in plain language before you open the MSA.
Lightyard is a service of Aircloud LLC, Mountain View, California, US. Member of NVIDIA Inception.
Summary of principal terms
Reserved GPU capacity · Illustrative
- Capacity
- [n] nodes, NVIDIA HGX B300, handed over in [n] waves
- Region
- [Region], Tier III facility
- Term
- 60 months
- Exit
- An option in the contract, on terms set in your proposal
- Rollover
- Part of the commitment can move into a newer GPU generation when it ships
- Prepayment
- Required. Paid into escrow, credited in equal parts against monthly invoices
- Transfer
- Possible with Lightyard's consent
Questions before scoping a cluster.
Anything else, ask. We usually reply within a few hours.
The minimum is 8 HGX nodes (64 GPUs) or one NVL72 rack (72 GPUs). Most clusters are 32 to 128 nodes; larger ones are scoped case by case. NVL72 systems are reserved as whole racks of 72 GPUs. If your number is different, ask anyway.
It depends on the GPU, the size and the region. HGX B300 is about 10 to 14 weeks from signing. Larger clusters are handed over in monthly waves of 32 nodes. Every proposal states delivery in weeks from signing, and we say so if we cannot meet your timing.
Every node in the cluster is dedicated to you for the term: every GPU, the NVLink domain between them, the host and the fabric. No other customer's work runs on it, and we never split a node into fractions.
Root on every node, no hypervisor, your OS image and your scheduler. Firmware and BMC stay with us.
We also run the software layer on top, such as a scheduler for containers and jobs, or inference endpoints.
Same single-tenant hardware either way, and either way we monitor it, replace failed parts and run the cluster to an SLA.
Prepayment is required, and it lowers the rate. It is paid into escrow with an independent escrow agent, which releases it in stages as supply, installation and acceptance milestones are met. From the first full month, it is credited in equal parts against each monthly invoice over the term. There is no security deposit.
The contract includes an exit option, on terms set out in your proposal. The term is 60 months, and 60 months is the only term we write. Part of the commitment can also move into a newer GPU generation when it ships.
In a Tier III facility in the region you choose: US West, US Central, US East, Europe, East Asia, Southeast Asia and South America. We commit to the region and the facility tier. Enterprises can also have a cluster on their own site.
We monitor the nodes, replace failed parts and run the cluster to an SLA in the contract, in bare metal and managed alike. Support runs for the full term.