Rent a GPU

Renting runs entirely in your browser — there's nothing to install.

Before you start

You only need two things:

• An account — sign in with Google or a one-time email code.
• Credits — rentals are billed per minute, so you need a balance to begin. A short 1-hour balance is required up front to open a session; you're only charged for the minutes you actually use. Top up by card, by a USDC transfer from an exchange or any wallet, or from a connected crypto wallet (USDC/USDT) on the Add Credits page — see Credits & Billing for details.

1. Browse available GPUs

Open the Rent page. You'll see every GPU that's online and ready, with its specs, location, and per-hour rate. Filter and sort to find the right card — from an RTX 4090 to a B300.

GPUs marked Spot are cheaper, but their owner can take the machine back at any time. If that happens you get 2 minutes' notice (a countdown on My Rentals and a notification) before the rental stops, and you pay only until it stops. Spot machines show "Spot · reclaimable" instead of a reliability score. Use the On-demand / Spot filter to see one kind or the other; for long jobs on a spot machine, checkpoint often.

2. Open a GPU

Click a GPU to open its page. Each node has a shareable link (/rent/<id>), so you can send a specific GPU to a teammate. If a node was taken in the meantime, the page tells you and points you to other available GPUs.

Everything about the machine is on this page before you rent it, so check it first:

• Measured performance — every node is benchmarked by Paralon: memory bandwidth against the card's spec, FP32/FP16 throughput, and when it was last verified. Bandwidth is the best single health signal; a card well below its spec will run your job slower.
• How the card is set up — the PCIe link it actually runs at and the power limit set by the provider, next to the factory default. A lower power limit means lower sustained speed.
• The host — CPU, RAM, disk, location, and the node's and provider's reliability score from past rentals (spot machines show "Spot · reclaimable" instead).

What you see here is what you get: the rental runs on exactly this machine, as listed.

3. Choose access and confirm

On the GPU's page:

1. Select an access type — Jupyter Lab (a browser IDE with Python, a terminal, and a file browser) is selected by default.
2. Accept the Terms of Service with the checkbox.
3. Click Start rental. This is the deliberate, billed step — you'll see the per-hour rate on the button. That rate is locked for your whole session; it can't change while your rental is running. Billing is pay as you go, per minute, and you can stop anytime.

4. Open your session

Your container starts in a few seconds. When it's ready you get a Jupyter Lab URL — click it to open your environment in a new tab. From the terminal you can pip install, run training, use nvidia-smi, and more.

Working in your session

A few things worth knowing:

• Terminals and notebooks keep running even if you close the browser — they live on the instance, not your tab.
• The instance can restart, which stops running processes (your files on disk are kept). For long jobs, checkpoint to disk so you can resume.
• Everything is wiped when the rental ends — download anything you want to keep before you stop.
• No Docker inside the session. A rental is an unprivileged container, so docker run and anything else that needs root on the host won't work. Install what you need with pip, conda or apt.
• Access is through Jupyter. Its terminal and notebooks are how you reach the machine; other ports are not exposed to the internet.
• Disk space is the "Disk free" shown on the GPU's page. Large models and datasets need to fit there.

If your credits run out

Your balance is checked every minute while a rental runs. When it can no longer cover the next minute, the rental stops automatically and you get a notification. Since a stopped rental is wiped, keep enough credit for the whole job and checkpoint long runs. Top up on Add Credits; see Credits & Billing.

If the machine stops responding

Paralon checks every running rental. If yours stops responding:

• After 2 minutes you get a notification (and so does the provider, who is asked to fix it).
• After 5 minutes billing stops: you don't pay for time the machine is down.
• After 15 minutes the rental is ended. The failure counts against the node's reliability score, which is why the score on each GPU's page is worth a look before you rent.

Messaging the provider

While a rental is running you can talk to the machine's provider: open Message provider on the rental in My Rentals (also under Messages). Use it for anything about the machine, such as a restart, a stuck session, or asking to keep it up longer. Messages are checked before delivery and contact details or off-platform payment offers are blocked; both sides see each other only by role and GPU model. Paralon support can read and join every conversation. Threads are deleted 30 days after the rental ends.

Something wrong?

If the machine runs slower than its page shows, or something else isn't right:

1. Compare with the GPU's page. The measured bandwidth, PCIe link and power limit there are what this machine was listed with.
2. Message the provider from the rental (see above) for anything on their side, like a restart.
3. Open a ticket under Support in your dashboard, category A rental, with the rental ID (on My Rentals) and a screenshot if it helps. We check the node and answer there. See Getting help.

Stopping

Stop the rental from My Rentals whenever you're done. You're billed only for the minutes you used.

Prefer code?

Everything on this page is also available over HTTP — browse GPUs, start a rental, get its connection URL, and stop it, all programmatically with the same API key. See the Rental API.

Need a hand? Join the community on Discord.