Docs

GPUBank documentation

How to rent a GPU, connect to it and pay only for the seconds it runs. The quickstart is right below.

CLI 1.4API v1Updated 1 November 2026

Launch your first GPU in five minutes.

You need an email address, $25 in any supported coin and an SSH key. No compute is billed until your machine passes its health check.

1
Create an account
Sign up at console.gpubank.io/register with an email and a password of 12 characters or more. We send a six-digit code to confirm the address. Signing in with GitHub or Google works too.
2
Top up your balance
In the console, open Billing › Top up and choose an amount from $25. Pay in Bitcoin, USDT, Ethereum, Monero or another listed coin. The balance is credited once the payment confirms on chain, usually within a few minutes, and a $100 top-up is credited as $135.
3
Add an SSH key
If you do not have one yet, create it on your computer, then paste the public key in Account › Security › SSH keys. Every machine you launch gets it.
bash
```
ssh-keygen -t ed25519 -C "[email protected]"
cat ~/.ssh/id_ed25519.pub
```

Launch a GPU

From the console: Pods › Launch, then pick a GPU, a region, a template and a disk. Or from your terminal:

bash

pipx install gpubank
gpubank login
gpubank launch --gpu h100-sxm --region eu-paris --disk mydisk --template pytorch-2.8

output

✓ instance inst_7c1e created
✓ allocated GPU H100 SXM in eu-paris         12.34 s
✓ attached disk mydisk (200 GB)               28.11 s
✓ health check passed, billing started        47.80 s
connect: ssh [email protected]

billing starts here

5
Connect and check the GPU
bash
```
ssh [email protected]
nvidia-smi
```
What you are paying now. The meter started at 47.8 s, when the health check passed. Run gpubank stop inst_7c1e and compute billing stops; the disk and the IP address are kept, and the disk costs $0.03 per GB-month until you delete it.
6
Open a port and move your files
Expose a web app on the public HTTPS address of the machine, and copy your data with the tools you already use.
bash
```
gpubank port open inst_7c1e 8188        # https://par1-g07-8188.gpubank.io
rsync -avP ./dataset/ [email protected]:/workspace/dataset/
```
Files under /workspace live on the persistent disk and survive every stop. Everything else sits on the local NVMe and is wiped when the machine stops.

Guides for what people run on GPUBank.

Run ComfyUI

Start ComfyUI with your models on a persistent disk and open it in the browser.

RTX 4090 or RTX 50906 min

Serve a model with vLLM

An OpenAI-compatible endpoint for Llama, Qwen or Mistral, with an API key in front.

RTX PRO 6000 or H100 SXM8 min

Fine-tune with Unsloth

LoRA or QLoRA on Llama 3.1 8B to 70B, with checkpoints written to your disk.

A100 SXM or H100 SXM10 min

JupyterLab

Notebooks with the GPU visible in PyTorch and JAX, reached over HTTPS.

L40S or A100 SXM4 min

Ollama

Pull a model, chat over the API and keep the weights on your disk.

L40S or RTX 50904 min

Bring your Docker image

Launch any OCI image from a public or private registry, GPU passthrough on.

Any GPU5 min

Expose a port

Serve a web app or an API on an HTTPS address, with or without a token in front.

Any GPU3 min

Transfer files

scp, rsync, Hugging Face downloads with your token, and S3-compatible buckets.

Any GPU5 min

Move a disk to another region

Copy a persistent disk from Paris to Frankfurt or Ashburn and launch there.

Any GPU3 min

Checkpoints on interruptible GPUs

Catch the five-minute notice, save state and resume on the next card.

Any GPU7 min

Start from a tested image.

Each template pins the driver, CUDA and libraries, and is relaunched on every GPU model before we mark it verified. All 25 are re-checked at least every three days.

Template	Tag	First start	Verified
PyTorch 2.8 · CUDA 12.8	pytorch:2.8-cu128	about 2 min	2026-11-01
Ubuntu 24.04 · CUDA 12.8	ubuntu:24.04-cu128	about 1 min	2026-11-01
JupyterLab 4	jupyter:4.3-cu128	about 1 min	2026-10-30
ComfyUI · model sync	comfyui:0.3-cu128	about 4 min	2026-10-30
vLLM 0.11 serving	vllm:0.11-cu128	about 3 min	2026-10-31
SGLang	sglang:0.4-cu128	about 3 min	2026-11-01
Ollama	ollama:0.12	about 2 min	2026-10-31
Text Generation Inference	tgi:3.3	about 3 min	2026-10-30
Axolotl fine-tuning	axolotl:0.9-cu128	about 2 min	2026-10-31
Unsloth	unsloth:2025.9-cu128	about 2 min	2026-10-30
faster-whisper API	whisper:faster-1.2	about 1 min	2026-11-01
InvokeAI	invoke:5.6	about 3 min	2026-10-31
Kohya trainer	kohya:25.1	about 3 min	2026-10-30
PyTorch on ROCm 6.4	rocm-pytorch:2.8-rocm6.4	about 3 min	2026-10-31
Ubuntu 22.04 · CUDA 12.4	ubuntu:22.04-cu124	about 1 min	2026-10-30
Debian 12 · CUDA 12.8	debian:12-cu128	about 1 min	2026-11-01
CUDA 12.8 devel	cuda:12.8-devel	about 2 min	2026-11-01
Ubuntu 24.04 · ROCm 6.4	rocm:6.4-ubuntu	about 1 min	2026-10-30
VS Code in the browser	code-server:4.99	about 1 min	2026-10-31
TensorFlow 2.19	tensorflow:2.19-cu128	about 2 min	2026-10-30
SD WebUI Forge	forge:2.x-cu128	about 3 min	2026-10-31
Open WebUI + Ollama	open-webui:0.6-ollama	about 2 min	2026-11-01
llama.cpp server	llamacpp:server-cu128	about 1 min	2026-10-31
TensorRT-LLM · Triton	trtllm:0.19-triton	about 6 min	2026-10-31
ComfyUI · video (Wan 2.2)	comfyui-video:0.3-wan2.2	about 5 min	2026-11-01
Your Docker image	any OCI image	your build	on launch

Core concepts.

Pod: An interactive machine with one to eight GPUs, SSH and a public IPv4. You start, stop and resume it.
Job: A command that runs to completion on a GPU, then stops by itself. Billed for its running seconds only.
Endpoint: An OpenAI-compatible inference URL that scales its workers with traffic. In beta.
Disk: Persistent storage in one region, billed at $0.03 per GB-month. It survives every stop and can move to another region.
Scratch: The local NVMe of the machine, included in the price and wiped when the machine stops.
Region: Where the GPU and the disk physically are: eu-paris, eu-frankfurt, us-ashburn or ap-singapore.
Meter: Counts running seconds from the passed health check to the stop, and draws them from your balance.
Balance: Prepaid, in US dollars, topped up in crypto. Every running second is drawn from it, and a low-balance alert warns you first.

The CLI, the API and Terraform do what the console does.

CLI

launch, ls, ssh, stop, resume, rm, port open, disk mv, logs. Installed with pipx, authenticated with gpubank login.

Python SDK and Terraform

pip install gpubank, then gpubank.Client() reads GPUBANK_TOKEN. Terraform resources: gpubank_instance, gpubank_disk, gpubank_spend_cap.

REST API

Base URL https://api.gpubank.io/v1, with a bearer token created in Developers › API keys. Every console action has an endpoint.

From inside a machine, http://169.254.169.254/gpubank/v1/ returns the instance id, the GPU, the region and, on interruptible machines, the pending notice.

bash

curl https://api.gpubank.io/v1/instances \
  -H "Authorization: Bearer $GPUBANK_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"gpu": "h100-sxm",
       "region": "eu-paris",
       "disk": "mydisk"}'

Webhooks

Signed with HMAC-SHA256 in the GPUBank-Signature header, retried for 24 hours until your endpoint answers 2xx.

instance.readyinstance.stoppedinstance.interruption_noticeinstance.faileddisk.movedbalance.lowspend_cap.reached

json

{
  "type": "instance.interruption_notice",
  "created": "2026-10-04T09:41:12Z",
  "data": {
    "id": "inst_7c1e",
    "stops_at": "2026-10-04T09:46:12Z"
  }
}

Limits and quotas.

Limit	Default	Can it be raised?
GPUs per machine	1, 2, 4 or 8 (up to 4 on the RTX 5090 and RTX 4090)	Fixed by the hardware
GPUs running at once, per account	16	Yes, on request
API requests	600 per minute per key	Yes, on request
Size of one persistent disk	10 TB	Yes, on request
Disks per region	20	Yes, on request
Webhook endpoints	10 per account	No
SSH keys	50 per account	No

To raise a limit, open a ticket from the console with the limit, the new value and what it is for. Most requests are handled the same day.

When something does not work.

The launch says queue.

That GPU is sold out in the region. Queue the launch, set an availability alert, or move your disk to a region with stock.

SSH asks for a password.

The key on your computer is not the one in Account › Security. Check with ssh -v which key is offered.

nvidia-smi shows no GPU inside your own image.

Your image needs the NVIDIA container runtime libraries; start from one of our templates or set NVIDIA_VISIBLE_DEVICES=all.

My files are gone after a stop.

They were outside /workspace, on the local NVMe, which is wiped at stop. Keep anything you need under /workspace.

A top-up has not arrived.

It is credited after the network confirmations shown on the payment page. If it is still pending an hour later, open a ticket with the 8-character top-up reference.

Status

Live state of every region and the incident history, in the console.

Changelog

What changed in the console, the CLI and the API, dated, newest first.

Something wrong in these docs?

Open a ticket from the console with the page and what you expected.