Compute · Dedicated GPU nodes
Whole NVIDIA nodes reserved for your team on a 1, 2 or 3-year term and configured to your workload. Run them on bare metal or on managed Kubernetes and Slurm. Serve your own models, or resell the capacity under your brand.
How it works
One capacity agreement covers the whole path, from the spec to the last day of the term.
Before handover · we build
From handover · your term
Before handover · we build
01GPU, node count, network, storage and term. We come back with a proposal.
One capacity agreement for the whole term.
We rack, cable and power the nodes, then stress-test every GPU before you see it.
You run your own tests against criteria agreed at signing. Nodes that fail are fixed or replaced.
From handover · your term
05The nodes join your cluster and your term starts.
We run the hardware and the site. You run your models, or sell the capacity.
Term, price, delivery date, uptime commitments and support are agreed per deal.
Talk to sales →Your stack, or ours
Your ML engineers get the environment they already use. Your compute team gets nodes that are theirs alone.
The hardware
Tell us the GPU, the cluster size and the network your models need. We configure the nodes around your workload, not a fixed catalog.
Terms
Choose 1, 2 or 3 years for each deal. A shorter term keeps you on current hardware; a longer one lowers the rate.
Keep the same nodes for another term, at a rate set at renewal.
Take the next GPU generation on a new term, on the same site.
Return the nodes when the term ends.
Why teams trust the nodes
We generate and firm the power, build the sites and run them ourselves. Nobody else sits between your nodes and the people who fix them.
2022
First site energized. Operating ever since.
~25 MW
Of compute built and running today.
4 sites
Built and run by our own field team.
2.5 GW
Of sites in development.
Placeholder: customer logos and a short case study go here once customers can be named.
Sites we build and run ourselves.
Los Toldos II · one of our operating sites
Also on the same site
Add interruptible compute on top of your dedicated nodes, or bring GPUs you already own.
The variable layer
30–50% below dedicatedto confirm
Runs on the site’s variable power. It pauses only for energy reasons, with notice first, then resumes. Suits batch work, evals and checkpointed training.
Fits batch inference, evals, background agents
Your GPUs, our power
Per kW-month · power includedrate to confirm
Powered colocation for hardware you own. We supply the energy, the site and the operations, without the wait for grid-connected colo.
Fits platforms with their own GPU fleet
FAQ
We do. We own and operate the GPUs and the sites. Your team gets exclusive use of the nodes for the term, without putting them on your balance sheet.
Yes. Tell us the GPU, node count, network and storage your models need, and we configure the nodes around them.
Yes. Take the nodes as bare metal with full root, or run them on managed Kubernetes or Slurm, with shared storage.
1, 2 or 3 years, chosen per deal. At the end you can renew, move to the next GPU generation, or hand the nodes back.
Our crew swaps the failed part from spares kept on site. Replacement times and uptime commitments are agreed in your contract.
Yes. You can resell, white-label or bundle it under your own brand, at your own prices.
Yes. Tell us the node count you need now and we will give you a delivery date. More nodes can join on the same site later, on their own term or ending with your first.
Yes. Interruptible compute runs on the site’s variable power at a lower rate, with notice before any pause. It suits batch tiers and work that can wait.
In Latin America, on sites we build and operate. We share the exact site and network routes during scoping. On bare metal, only your team has access to the operating system and your data.
Tell us about the workload, the GPUs, how many nodes, when you need them and the term. We reply with a proposal.
We reply within 48 hours. [email protected]