Start a GPU cluster yourself

Until this week, a GPU cluster on Runcrate meant asking us for a quote. Now you start one yourself, the way you start an instance: from the dashboard, your agent, the CLI or the API.

A cluster is 2 to 64 GPU machines that you create, connect to, pay for and terminate as one thing. Every machine has the same GPUs, they reach each other over SSH by name (node-0, node-1, ...), and on virtual machines and bare metal they share a /shared folder. Node 0 is the head.

Pick the software

Software	What you get
None	The machines, the private network, a hostfile and ready-made torchrun, mpirun and DeepSpeed commands
Slurm	Submit with sbatch and srun, watch the queue with squeue
Kubernetes	GPU scheduling, and a kubeconfig you use from your own computer
Ray	The head on node 0 and every node joined

The cluster is ready when the software has checked itself on every node: every Slurm node idle, every Kubernetes node Ready with its GPUs schedulable, every Ray node in ray status. You get an email when it is.

Ask your agent or the CLI

Connect Runcrate to your agent (one line in Claude Code or Codex, one click in Cursor), then ask in plain words:

The agent lists the offers with cluster_offers, creates the one you pick with create_cluster, and checks get_cluster until it is ready. From a terminal, the CLI does the same:

1runcrate clusters offers \
2  --gpu L40S --nodes 4 \
3  --stack slurm
4runcrate clusters create \
5  --name evals \
6  --offer <offer-id> \
7  --nodes 4 --stack slurm
8runcrate clusters get evals

The dashboard has the same flow under Clusters, and the API takes an Idempotency-Key, so a retried create never starts a second cluster.

Whole or not at all

A cluster starts whole or not at all. If any machine fails, stops or is late for a step while the cluster starts, the whole cluster fails at once, every machine is released, and nothing is billed. Nothing is started again without you: the cluster page shows the offers available now, and you choose.

Billing starts when every machine has joined, from the same moment for all of them, and ends when you terminate the cluster. On the Billing page a cluster is one line that opens to a row per machine.

Priced by the machine

Each offer shows the price per hour for the whole cluster, with the price per machine under it. A location that reports how many machines it has free sells up to that many; one that does not stops at 8 machines, because one machine short would fail the whole cluster. For more than that, Request a quote sits under the sizes.

Know the network

Self-serve clusters run on Ethernet, with an encrypted private network on top that moves about 1 to 2 Gbit/s between machines. That suits independent work on every node, Slurm job arrays, Ray tasks, Kubernetes workloads and LoRA fine-tuning. It is not built for training one full model across machines, where GPUs on different machines exchange gradients every step. Run the NCCL test after you create a cluster to see the bandwidth you get.

Read the clusters docs for every option, or create one now.