A QuantaCloud GPU VPS is an Ubuntu 22.04 virtual machine with one to eight NVIDIA GPUs, the NVIDIA driver and Docker installed, and SSH access as the ubuntu user. You pay from prepaid credit an hour ahead, and the unused seconds come back when you stop. It differs from a classic VPS in the one way that matters most: stopping it deletes it, disk included.
What you get on the VM#
Every GPU VPS starts from the same image, Ubuntu 22.04 with the NVIDIA driver and Docker, launched from the Bare Metal template.
| Part | What you get |
|---|---|
| GPUs | 1, 2, 4 or 8 NVIDIA GPUs with 48 to 141 GB of memory each, from the RTX A6000 to the H200 NVL. The GPU catalog lists them |
| CPU and RAM | Fixed per configuration: for example 16 vCPUs and 180 GB of RAM with one H200 NVL, and 126 vCPUs and 800 GB with eight A100 SXM4 (2026-09-27 catalog) |
| Disk | One local disk included in the hourly price: 256 GB with one RTX A6000, up to 5,000 GB on 8-GPU configurations (2026-09-27 catalog) |
| Network | A public IP address and SSH on port 22. No egress or ingress charges |
| Login | SSH as ubuntu with your key. You do not log in as root |
| Software | The NVIDIA driver and Docker. You install your own stack on top |
| Regions | us-east-1 in Virginia, and us-midwest-1, us-midwest-2 and us-midwest-4 in the Midwest |
Bare Metal is the template's name, not the hardware: every on-demand GPU VPS is a VM. If you are after GPU server hosting with a physical machine to yourself, that is a dedicated GPU server, which we build to order.
Checks I run on a fresh VM#
The one thing I always check on a new GPU VM is that the driver, CUDA and Docker agree before I install anything. Five commands cover it:
nvidia-smi
nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv
sudo -n true && echo "passwordless sudo works"
sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
lsblk -d -o NAME,SIZE,ROTA,MODEL && df -h /
The fourth line is NVIDIA's own Container Toolkit test: if it prints the same GPU table as the first, your containers can see the GPUs.
If the driver or CUDA version is older than your framework needs, the driver and CUDA version guide explains how to read the numbers.
How it differs from a typical VPS#
The biggest difference is that there is no stop and start: Stop terminates the VM and deletes its disk.
| What a VPS usually means | QuantaCloud GPU VPS | |
|---|---|---|
| Billing | A monthly plan | Prepaid credit, charged an hour ahead, unused seconds refunded when you stop |
| Stop and start | Power off, keep the disk, start again later | Stop terminates the VM and deletes its disk |
| Snapshots and backups | Usually on offer | None |
| Extra storage | Attachable volumes | None: one fixed local disk |
| Resizing | Change the plan in place | Launch a new VM with another configuration |
| When it ends | When you cancel | When you stop it, or when your balance cannot cover the next hour |
Treat the VM as disposable: keep code in Git, script the setup so a new VM is ready with one command, and copy results to your own machine before you stop. The SSH and VS Code guide covers SSH, VS Code Remote and port forwarding.
Hourly and monthly cost#
A GPU VPS costs its hourly rate for every hour it runs, so a month is the rate times the hours. The table converts the 2026-09-27 price of each one-GPU VM into an always-on month of 730 hours, which is 8,760 hours a year divided by 12 (our calculations).
| GPU | Lowest live price per GPU-hour | One-GPU VM on 2026-09-27 | 730 hours at that price |
|---|---|---|---|
| RTX A6000 (48 GB) | $0.48/GPU-hr | $0.48/hr | $350.40 |
| RTX 6000 Ada (48 GB) | $0.78/GPU-hr | $0.79/hr | $576.70 |
| L40 (48 GB) | $0.94/GPU-hr | $0.94/hr | $686.20 |
| L40S (48 GB) | $1.09/GPU-hr | $1.09/hr | $795.70 |
| A100 SXM4 (80 GB) | $1.49/GPU-hr | $1.50/hr | $1,095.00 |
| RTX PRO 6000 Blackwell (96 GB) | $2.39/GPU-hr | $2.39/hr | $1,744.70 |
| H100 PCIe (80 GB) | $2.59/GPU-hr | $2.59/hr | $1,890.70 |
| H200 NVL (141 GB) | the console price | $3.43/hr | $2,503.90 |
Prices checked 6 Oct 2026, 01:40 UTC
Running the VM only while you work changes the arithmetic: 8 hours a day for 22 working days is 176 hours, $84.48 on one RTX A6000 at the 2026-09-27 price (our calculation: 176 x $0.48). The catch is the stop rule. Every morning starts from a clean VM, so this pattern pays off only when your setup is a script you can rerun and your model downloads are quick, which downloading Hugging Face models fast covers. The pricing page has the billing rules in full.
When a GPU VPS is the right fit#
A GPU VPS fits work measured in hours to weeks that you can rebuild from a script. That covers serving a model from your own Docker image, as in deploying vLLM with Docker, a development box you reach through VS Code, batch jobs, and fine-tuning runs that write checkpoints you copy off.
It is the wrong fit for data that must survive a stop, and for production traffic that needs a service commitment: there is no SLA. For months of always-on use, reserved capacity built to order is the better model, from one dedicated server upward. Send a capacity brief.
Launch one#
- Create an account and add at least $5 of credit (adding credits).
- Add your SSH public key, Ed25519 or RSA, under SSH Keys, or save the managed key's private key when the console offers it (SSH keys).
- Pick a GPU offer in the marketplace, choose the Bare Metal template and deploy (deploying a GPU).
- Connect with
ssh -i ~/.ssh/your-key ubuntu@your-vm-ip(connecting over SSH).
Frequently asked questions#
Do I get root access?
You log in as ubuntu, not root. The sudo line in the checks above shows whether ubuntu can run administrative commands without a password.
Can I run my own Docker image?
Yes. The Bare Metal template ships with Docker, so pulling your image and running it with --gpus all is the usual way in, once the Docker check above passes. There is no custom-image template, and nothing you pull is kept after the VM stops.
Can I keep the disk when I stop the VM?
No. Stop terminates the VM and deletes its disk, and there are no volumes or snapshots. Copy what you need first with rsync or scp.
Is it a bare-metal server?
No. Every on-demand GPU VPS is a VM, whatever the template is called. Physical servers are available as reserved builds.
What happens if my balance runs out?
The VM is terminated at the first hourly charge your balance cannot cover, and its disk is deleted. Auto top-up on the Billing page avoids it.
My rule for a GPU VPS: rent it by the hour when you can rebuild it from a script, and ask for a reserved build when losing the machine is not an option.
Launch an Ubuntu GPU VM