Every instance is a real server with root access. The console gives you power control, live logs and telemetry; everything it does is also available over the API.
Watch the node build in real time — scheduler placement, image pull, driver load, network attach. When something goes wrong you see it happen rather than getting a generic failure.
GPU utilisation, VRAM occupancy, board temperature, power draw and host CPU, sampled every few seconds and streamed to the console.
Stop an instance to halt billing while keeping the root volume, or destroy it to release the hardware and erase the disk.
Capture a root volume at any point and restore it onto a fresh instance later. Charged at $0.04/GB/month.
Block storage that outlives the instance. Detach from one machine, attach to another in the same region, keep your datasets and checkpoints.
Upload public keys once and inject them into every instance at boot through cloud-init. Password authentication is disabled by default.
Six states, and you are only billed for one of them.
| State | What is happening | Billed |
|---|---|---|
| queued | Waiting for capacity in the chosen region | No |
| allocating | Scheduler has matched a host, reserving the card | No |
| provisioning | Root volume created, GPU passed through | No |
| imaging | Disk image written and resized | No |
| booting | Kernel starting, drivers loading, network attaching | No |
| running | SSH is up and the GPU is yours | Yes |
| stopped | Halted, root volume retained | Storage only |
| terminated | Hardware released, disk erased | No |
Anything the console can do, a script can do. Tokens are scoped per project.
Full endpoint reference in the documentation.
Create an account and deploy an instance — the provisioning log streams live while it builds.
Get started