GPU clusters and reserved capacity

The largest machine you can launch yourself on DeployGPU is one VM with 8 GPUs. Anything bigger, and anything with a fixed term at a fixed price, is a quote: you describe the GPU, the count, the interconnect and the dates, and a person replies by email. This page covers where that line sits and what to put in the request.

What you can launch yourself today

Every offer is a single VM with 1, 2, 4 or 8 GPUs, billed by the hour with no commitment. These are the 8-GPU nodes in stock right now:

An 8-GPU SXM node, where the Interconnect column on the GPU pages reads NVLink (SXM), connects all 8 cards through NVLink, which is what tensor-parallel inference and most single-node training runs need. The multi-GPU page explains how to check the topology once you are on the machine. If one node is enough, launch it from the marketplace and skip the form.

When to ask for a quote instead

The self-serve platform stops at one VM, so a request makes sense when you need any of these:

• More than 8 GPUs working together, across several nodes.
• InfiniBand or RoCE networking between nodes.
• A term of a month or more at a fixed price, rather than on-demand rates that move.
• A GPU that is rarely in stock on demand, such as B200 or B300, reserved for your dates.
• A specific region, or storage and compliance needs that a marketplace offer does not state.

Nothing in that list can be launched from the dashboard or the API today. There is no self-serve multi-node cluster, no self-serve InfiniBand and no reserved pricing you can buy yourself, so these requests are handled one at a time.

What happens after you send it

The request goes to the DeployGPU team by email. A person reads it, works out what capacity can be arranged for your GPU, count and dates, and replies with what is possible and what it would cost. Sending the form does not reserve anything or charge you anything, and the terms of any arrangement are set out in the reply rather than on this page.

A useful request answers the questions a provider will ask anyway: the workload (training, fine-tuning or inference), the model size, how many GPUs and of which type, whether the nodes must talk over InfiniBand, when you want to start and for how long, where the machines should be, and how much storage sits next to them. Clusters above 16 GPUs are quoted on terms of 12 months or longer.

What a quote does not include

No part of DeployGPU carries an uptime SLA today, and that includes arranged capacity unless a quote says otherwise in writing. For smaller needs, the on-demand platform is often the faster path: an 8-GPU node from the table above is usually running within minutes, billed by the hour, and stopped when you are done. The H200 and B200 pages show which large-memory GPUs are in stock right now.