Root on a real machine
Each rental is a full virtual machine, not a shared container. You log in as root over SSH with your own key.
A full virtual machine with root SSH access and one or more GPUs. Pick the GPU, the count, an image and the hours; it’s usually ready to SSH into in about a minute.
Each rental is a full virtual machine, not a shared container. You log in as root over SSH with your own key.
Ubuntu 22.04 with CUDA 12.4, or Ubuntu 24.04 with CUDA 13.0 for the RTX 5090. NVIDIA driver, Docker, Python 3, git and tmux are already installed.
Add a disk mounted at /data. It is kept across restarts of the machine and deleted when you release it, unless you choose to keep it.
Prepaid from your balance. Unused whole hours come back if you release early.
See the price before you pay. It holds for two minutes.
ext4, kept across restarts. Deleted on release unless you keep it.
Full root over SSH, with the keys on your account.
Pick a GPU model, the GPU count, an image and how long you need it.
Check the quote and pay from your prepaid balance.
Connect as root with the host and port from the instance page.
Release from the console when you’re done. Unused whole hours are refunded.
ssh root@<host> -p <port>
nvidia-smiRun long jobs in tmux and keep checkpoints on /data across restarts.
Try a model on a real GPU for a few hours, then release and get the unused hours back.
Docker is installed, so you can run the containers you already use.
After you pay, a machine is usually ready to SSH into in about a minute.
The GPU models are listed on the pricing page. Live prices and availability are shown in the console before you pay.
You pay up front from your prepaid balance, by the hour, day or month. The only other charge is optional storage for disks you keep. See the pricing page for how refunds work.
In Asia-Pacific (East).