🚀 One API for every frontier model on FastInfra

Browse models

Transparent per-token pricing → compare providers on FastInfra

View pricing

OpenAI-compatible chat completions → start in minutes

Read the docs

⚡ Server Auction → Enterprise bare metal from $78.15/mo (173 in stock)

Browse deals

Spot GPUs · Interruptible · Billed per minute

Spot GPU instances at a discount to on-demand

FastInfra spot GPU instances are the same dedicated pods as GPU Cloud, rented from spare capacity at a large discount to the on-demand price and billed per minute. The provider can reclaim a spot pod at any time, so they suit checkpointed training, batch inference, rendering and experiments that can resume. The configurator shows the live spot and on-demand price side by side for every GPU type, and a reclaimed pod stops billing the moment it is interrupted, with a clear preempted status in the console.

Same pods, same images, same volumes as on-demand GPU Cloud. Spot pods run on spare capacity at a steep discount and can be reclaimed; billing stops the moment that happens.

Up to 70% off on-demand

Spot pods are priced from spare capacity. The deploy page shows the live spot and on-demand rate side by side so the saving is visible before you commit.

Preempted, not billed

If the provider reclaims a spot pod it moves to a preempted status and billing stops immediately. Attached storage volumes are kept so you can relaunch from the last checkpoint.

Same tooling

SSH, JupyterLab, custom images, environment variables and the GPU pod API all work the same. Spot is a per-pod toggle, not a different product.

Deploy a spot pod

Spot is a toggle on the normal deploy page. Use a start command that resumes from a checkpoint on a storage volume so an interrupted job picks up where it stopped.

1. GPU Cloud → Deploy → pick a GPU → choose "Spot" under pricing tier
2. Attach a storage volume for checkpoints (kept after preemption)
3. Start command, e.g.
   python train.py --resume-from /workspace/ckpt --save-every 10m
4. If the pod shows "preempted", click Relaunch — billing stopped at the interruption

Frequently asked questions

What is a spot GPU instance?

A spot GPU instance is a dedicated GPU pod rented from spare capacity at a discount to the on-demand price. In exchange, the provider may stop it at short notice to reclaim the hardware, so workloads should checkpoint and be able to resume.

How much cheaper are spot GPUs than on-demand?

Typically 40 to 70 percent cheaper depending on the GPU type and current demand. The configurator shows the live spot price next to the on-demand price for each GPU so you can see the exact saving before you deploy.

What happens when a spot pod is interrupted?

Billing stops immediately and the pod shows a preempted status in the console. Attached volumes are kept so you can relaunch and resume from the last checkpoint; the container disk is not preserved.

Which workloads suit spot GPUs?

Anything that tolerates a restart: fine-tuning with regular checkpoints, batch embedding or inference jobs, video rendering, hyperparameter sweeps and development boxes. Latency-sensitive production serving should stay on on-demand pods.

Can I mix spot and on-demand pods?

Yes. Spot is a per-pod choice at deploy time; the same account, volumes and API keys work for both, and you can run a steady on-demand pod next to spot workers for burst capacity.

Reviewed by Ashish Kumar, Head of Product · Updated October 21, 2026. Prices and counts on this page come from the live catalog.

Burst capacity without the on-demand bill

Run checkpointed training, batch inference and experiments on spot pods and keep on-demand pods for the work that cannot pause.