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

From nothing to an SSH prompt.

Five steps, about five minutes, and around $0.03 if you follow it on the cheapest GPU in the catalogue.

Last reviewed 15 September 2026 · something wrong or missing?

This walkthrough uses the console. Everything here has an equivalent in the CLI and the API, both in private beta.

1. Create an account#

Sign up at cloud.spotgpus.com with an email address and a password. You will be asked to confirm the address; that is the only personal data the platform needs to start. Turn on two-factor authentication in Settings straight away if the account will hold credit — it takes a minute and it protects a balance.

2. Add credit#

Billing is prepaid: you top up a balance and it is drawn down per minute. The minimum top-up is $40, there is no top-up fee, and credit does not expire.

In Balance & billing, choose an amount and a coin. The page tells you what the credit buys before you pay — for example, $40 is about 72 hours on a spot H100 or 1333 hours on a RTX A4000. Send the exact amount to the address shown, on the network shown, and the balance updates when the transaction confirms.

The network matters more than the coin.

USDT on Tron and USDT on Ethereum are different networks. Send on the network printed next to the address, or the deposit will need manual recovery.

3. Add an SSH key#

Paste your public key in SSH keys. It is installed on every instance you launch, so you never receive a password. If you do not have one:

on your machine
$ ssh-keygen -t ed25519 -C "spotgpus"
$ cat ~/.ssh/id_ed25519.pub

4. Launch an instance#

Open Deploy, then work down the screen:

  • GPU model — pick from the catalogue. If you are only trying the platform out, a RTX A4000 at $0.03 an hour is enough to see everything work.
  • Node size — 1 GPU unless a model does not fit in one. Each model page has the memory arithmetic.
  • Tier — spot for anything that can be interrupted with 2 minutes of notice, on-demand otherwise.
  • Region — the one closest to your data. Prices are identical.
  • Template — PyTorch, vLLM, JupyterLab, a bare CUDA base, or your own image.
  • Persistent disk — optional, but the thing that makes a reclaim survivable. 100 GB costs $8.00 a month.

The summary shows the hourly rate and the per-minute equivalent. Launch, and the instance is usually reachable in under 60 seconds; billing starts at that moment, not before.

5. Connect and work#

terminal
$ ssh root@<address shown in the console>
# The GPU is already visible to the container.
$ nvidia-smi --query-gpu=name,memory.total --format=csv
name, memory.total [MiB]
H100 SXM, 81,920 MiB

# Datasets on scratch, checkpoints on the persistent volume.
$ ls /mnt/scratch /mnt/vol

Templates with a web interface — JupyterLab, ComfyUI, code-server — publish a URL in the instance page. Ports you open yourself are reachable on the instance address.

6. Stop it#

Press Stop in the console. Compute billing ends at that minute. The instance keeps its disk, its IP and its configuration, so starting it again picks up where you left off. Terminate it when you are done for good; a persistent volume outlives the instance and keeps being billed until you delete it.

Make it spot-proof#

If the job is longer than an afternoon, spend ten minutes on the three habits that make interruptions a non-event:

  1. Write checkpoints to /mnt/vol, never to /mnt/scratch.
  2. Poll /v1/notice — or handle the webhook — and flush a final checkpoint when it fires.
  3. Make your entrypoint resume from the newest checkpoint instead of starting from zero.

The full pattern, with payloads and code, is in interruption handling.