How to rent a DGX Spark
Create an account with Google or GitHub, add prepaid credit, add your SSH public key, and start a Spark from the console. From there you connect over SSH as root into your own container with Docker and CUDA already installed. The whole flow takes a few minutes before the machine itself is booting; this page walks through each step and the billing decision at the end.
1. Create an account
Sign in at console.gpuwerk.com with Google or GitHub. There's no separate password to set for the account itself. This step alone doesn't commit you to anything or start any billing.
2. Add prepaid credit
Rental is prepaid, not invoiced after the fact. Adding credit uses Stripe. From our pricing page, verbatim: "Top up $10 the first time and get $20 in credit (a dollar-for-dollar match of up to $10 on your first top-up)." That is enough to try the console before deciding how much credit to load. A top-up is a prepaid balance, not a fixed-duration booking: it's spent down as instances run, at $0.79/hour for a single Spark or $1.79/hour for the two-node, 256 GB cluster, billed per minute.
3. Add an SSH key
Every instance is reachable only over SSH; there's no password fallback. If you don't already have a key pair, our SSH setup guide covers generating one with ed25519 and loading it into your agent, which takes about two minutes. Once you have a public key, paste it into the console. Keep the private key on your own machine; only the public key goes to GPUwerk.
4. Start a Spark
From the deploy screen, pick a single Spark or the cluster, and an image if you want one preconfigured with a model already serving, such as the Qwen3 Coder 30B image used in our first-evaluation guide. You can add prepaid credit before configuring a Spark. The console shows live availability when you deploy; if nothing is free, you can join the waitlist. Adding credit does not reserve a machine. Once deployed, wait for the instance to reach ready status, then copy the SSH connection command from the instance details.
5. Connect over SSH
Connecting drops you into your own container as root, with Docker and CUDA already installed, so you can run a serving stack you choose or use whatever the image already started. If you're setting up a passwordless alias so the connection is one word to type, that's also covered in the SSH guide. From here the machine behaves like any Linux box you have root on: install what you need, or use what's already running.
6. Stop or terminate when you're done
These are two different actions with two different costs, and mixing them up is the most common way people keep paying without meaning to.
- Stop keeps your workspace on the same node, bills 75% of the running rate to hold that reservation, and lets you resume later on the same machine, same SSH port and hostname. Model presets may need a few minutes to warm back up.
- Terminate ends billing completely and deletes the workspace permanently, which is why the console asks you to type the instance name to confirm. If you want to stop paying, terminate rather than stop.
Copy anything you need off the instance before you terminate it. GPUwerk only keeps a periodic recovery copy of /workspace for hardware failures, refreshed roughly every six hours, not a backup service. If credit runs out or a budget is reached, the instance is stopped and the machine released, with the workspace saved off it for 7 days so topping up and pressing start restores it; past that window it's deleted the same as a manual terminate.
Plan the session before you start it
A short evaluation costs less than you might expect: at $0.79/hour, eight hours of running time is $6.32. Decide roughly how long your test will take and load enough credit to cover it comfortably, since a top-up that runs out mid-session ends the instance rather than pausing it. Our first-evaluation guide walks through planning a specific model test end to end, including exact commands to run once you're connected.
What this costs against buying instead
If you're trying to decide whether to rent at all versus buy the hardware outright, that's a separate question from the mechanics on this page. Our price page lays out the EU purchase range next to the rental rate, and the rent-vs-buy calculator works out the break-even point for your own usage pattern.
FAQ
How do I sign in to the console?
With Google or GitHub sign-in at console.gpuwerk.com. There is no separate password to set or remember for the console account itself.
How much do I need to add to start?
Top-ups are prepaid credit against the $0.79/hour on-demand rate, billed per minute. Top up $10 the first time and get $20 in credit: your first top-up is matched dollar-for-dollar, up to $10.
What's the difference between stopping and terminating?
Stopping keeps your workspace on the same node and bills 75% of the running rate to hold that reservation. Terminating ends billing entirely and deletes the workspace for good, which is why it asks you to type the instance name to confirm.
Do I need my own inference server, or is one preinstalled?
Both are possible. You get root SSH into your own container with Docker and CUDA preinstalled, so you can install anything yourself, and some images come preconfigured with a model already served, such as the Qwen3 Coder 30B image used in our first-evaluation guide.
Related reading: current availability, full specs, what it costs to buy instead, and where to buy one in Europe if renting isn't the answer for your workload.