How to rent a cloud GPU (and not pay for idle time)

The rule I follow when renting a GPU is simple: pick it by memory, launch it only when the work is ready, and stop it the moment the results are off the machine. Get those three right and the provider you pick matters much less. This guide walks through the whole loop on DeployGPU, from choosing the card to getting your money back for the unused minutes.

1. Work out how much GPU memory you need

GPU memory decides what runs at all, so start there and let price decide between the cards that fit. A rough map:

The models section has the arithmetic for specific models, and the KV cache explainer covers why context length and concurrent users need memory on top of the model weights. The smallest GPUs DeployGPU launches as VMs have 48 GB. Consumer cards like the RTX 4090 appear only as Vast.ai containers in the dashboard, labelled testing only.

2. Create an account

Sign up with Google, GitHub, an email magic link, or an email and password. No card is needed to sign up. At signup DeployGPU also creates an SSH key called Default Key and offers its private half for download once. Save it, or plan to add your own key in step 4.

3. Add credits

DeployGPU is prepaid. Add credits on the billing page by card, from $5. Each top-up carries a card processing fee of 5% + $0.30, so adding $10 costs $10.80 and puts $10.00 on your balance. To launch, your balance must cover one hour of the offer you pick, because launching reserves the first hour.

Prepaid credit is also your spending cap. Without auto top-up, a forgotten instance can never cost more than what is on the balance. The pricing page has worked examples.

4. Add your SSH key

If you did not keep the Default Key, add your own public key on the SSH keys page before launching. Keys are installed when the machine is created, not later. Ed25519 keys work on every provider. The SSH guide shows how to make one.

5. Pick an offer

The marketplace lists live offers you can launch right now, with the hourly price, the provider, the region, and the CPU, RAM and disk that come with the GPU. Filter by GPU and by number of GPUs, then sort by price. Every price includes any platform fee, and the deploy page breaks it down before you confirm.

Offers come in fixed sizes of 1, 2, 4 or 8 GPUs, and stock moves by the minute. If the offer you wanted is gone, the next one down the list is usually a similar machine at a similar price.

6. Choose what the machine runs

On the deploy page, the Template setting decides what is set up for you:

The apps take longer because DeployGPU waits until the app itself answers before it shows Running. Their links open only for your DeployGPU account, in a browser. The templates page compares them.

7. Launch and connect

Deploy, then watch the Deployments page. When the status reads Running, copy the SSH command from the deployment row, or click Open Application for a template. On some providers SSH needs another minute after Running before it accepts connections, so a Connection refused straight away is not a failure yet.

If a launch fails on the provider's side, the reserved hour comes back to your balance automatically, and an email tells you the reason and the refund. There is no automatic retry on another provider, so pick another offer and launch again.

8. Work, and save as you go

Everything on the instance disk is deleted when the instance stops, whether you stop it, your balance runs out, or the provider loses the machine. Push code to git and copy results off while you work, not only at the end:

rsync -avz --progress -e "ssh -p 22" ubuntu@YOUR_VM_IP:~/results/ ./results/

9. Stop it

Stop the instance from the Deployments page, or through the API, as soon as your files are safe. The unused seconds of the current hour go back to your balance. There is no idle timer and no maximum runtime, so a GPU you forget keeps billing until you stop it or the balance reaches zero. The stop guide has a checklist and a script that stops anything running longer than a limit you set.