New B200 spot capacity is live in US East from $1.69 per GPU-hour. See availability
Documentation

Everything on this platform is five objects.

Learn them once and the console, the CLI and the API all make sense — they are three ways of touching the same things.

Last reviewed 15 September 2026 · something wrong or missing?

A GPU instance here is a virtual machine with a GPU passed through to it, a template as its operating environment, an optional persistent disk, and a tier that decides whether we may take the hardware back. Nothing else is hiding behind the interface.

The five objects#

Instance
One virtual machine with one or more GPUs of a single model, created from a template, in one region. It has a reference like i-7f3a2c, an IPv4 address and an SSH endpoint. It is billed per started minute while it runs.
GPU model
The hardware: memory, bandwidth, architecture, node sizes and prices. All 47 models are in the catalogue, and the same data feeds the console, this documentation and the price pages.
Template
A container image plus the runtime configuration that makes it useful — PyTorch, vLLM, ComfyUI, ROCm, a bare CUDA base, or any OCI image you bring. Images are pre-pulled on the nodes, which is why an instance is reachable in under 60 seconds.
Volume
A persistent network disk, 10 GB to 4,000 GB, attached to one instance at a time. It survives stops, reclaims and tier changes — the piece that makes spot capacity usable. $0.08 per GB-month.
Region
3 locations — US East (Virginia), US West (Oregon), EU Central (Frankfurt). Prices are identical everywhere; pick the one closest to your data. A volume belongs to its region and cannot cross.
Tiers are not objects, they are promises.

Spot, on-demand and reserved run on the same nodes with the same images. What changes is whether we may reclaim the hardware, and what that costs you per hour.

The life of an instance#

  1. Configuring. You choose a model, a node size, a region, a template and, optionally, a disk. The console shows the hourly rate before anything starts.
  2. Provisioning. The node is allocated, the image is prepared and your SSH key is installed. Nothing is billed during this step.
  3. Running. The instance is reachable over SSH and on its exposed ports. Billing starts at this moment, per started minute.
  4. Reclaim notice (spot only). 2 minutes before a stop, the notice appears on the metadata endpoint, on your webhook and in the console. See interruption handling.
  5. Stopped. Compute billing ends. The volume, the IP and the configuration are kept, so the instance can be started again — by you, or automatically on the next free capacity.
  6. Terminated. Only when you say so. Local NVMe scratch is wiped; the persistent volume survives until you delete it yourself.

How you are billed#

Two lines, and no third one. Compute is the hourly price of the model divided by 60, charged for every started minute the instance is running. Storage is $0.08 per GB-month, prorated, charged while a volume exists.

EventComputeStorage
ProvisioningNot billedBilled if the volume exists
RunningPer started minuteBilled
Reclaim notice windowBilled as running timeBilled
StoppedNot billedBilled
Terminated, volume keptNot billedBilled
Volume deletedStops at that minute

A deployment requires at least one hour of the instance's rate on your balance — the platform will not start something it cannot bill. Credit is prepaid from $40, never expires, and a monthly spend cap can stop instances before a ceiling is crossed. The details, with a calculator, are on the pricing page.

Ways to drive the platform#

Console

Everything the platform can do, including billing, support and account settings. cloud.spotgpus.com

CLI Private beta

Launch, watch and stop from a terminal or a Makefile. Command reference

REST API Private beta

The same objects as JSON, with webhooks for lifecycle events. API reference

The CLI and the API are in private beta.

Keys are created in the console, and access is enabled per account on request while the interface settles. The reference pages describe exactly what is being rolled out — ask for access and we will tell you where the queue stands.

Where to go next#