GPU VPS, billed by the hour
Updated September 23, 2026
We rent NVIDIA A16 and A40 cards by the hour from $0.085/hr, paid from a crypto balance with no identity checks. Drivers and CUDA are already in the image, so nvidia-smi answers on first boot. Destroy the server and the charge stops the same hour.
Renting a GPU usually means a monthly commitment, a card on file and an identity check. This is the opposite: you top up with crypto, take a card for the afternoon, and destroy it when the job is done.
The entry point matters more than the top end. A fractional A16 with 2 GB of VRAM costs $0.085/hr - a $20 top-up runs it for more than ten days, which is what makes a GPU worth trying at all if you are not a lab with a budget.
What you can rent and what it costs
Fractional A16: $0.085/hr for 2 GB of VRAM, $0.17/hr for 4 GB, $0.33/hr for 8 GB. The card is physically shared, your slice of VRAM is not - it is assigned to your machine alone.
Whole and fractional A40: from up to a 48 GB card at $2.397/hr. Bigger VRAM is what decides whether a model fits at all, so the jump from 8 GB to 48 GB is the one that matters, not the clock speed.
Every card is billed hourly from your balance, exactly like a regular server. There is no reservation, no minimum term and nothing to cancel: destroying the server ends the charge in the same hour.
Drivers, images and what boots
The image ships with the NVIDIA driver already loaded - nvidia-smi answers on the first boot, and you do not install anything. We learned this the hard way: adding drivers on top of the provider image pulled a newer userspace than the loaded kernel module and broke nvidia-smi outright, so now we deliberately touch nothing.
You get root over SSH with the password you set at deploy, a dedicated IPv4 and the same panel as any other server. What you build on top - PyTorch, CUDA toolkit, a Jupyter notebook, an inference server - is yours to choose.
The one rule worth knowing before you start
A GPU server is deleted, not archived, if your balance reaches zero. Regular servers are snapshotted first and can be restored for days; on a GPU machine the disk is hundreds of gigabytes and storing it would cost more than the server earns, so we do not keep it.
In practice: keep enough balance for the run, and copy results off the machine as you go. You get a low-balance warning by email well before anything is removed, and the panel shows how many hours your balance covers at the current burn rate.
Which cards we actually hand out
Cards live on Vultr in our mix, and the catalogue only lists what the provider genuinely allocates to our account - half of their GPU line-up exists in the price list and is never handed out, and we do not sell that. Available VRAM sizes today: 2 GB, 4 GB, 8 GB, 32 GB on A16, and 2 GB, 8 GB, 48 GB on A40.
Every GPU size is deployed through a live end-to-end test before it goes on sale: an actual server created through our own provisioning, logged into over SSH, with nvidia-smi checked. A card that exists at the provider but has not passed that test is visible in the panel and cannot be ordered.
GPU sizes & hourly pricing
| Card | vCPU | RAM | NVMe | $/mo from | $/hr from |
|---|---|---|---|---|---|
| A16 · 2 GB VRAM | 2 | 8 GB | 50 GB | $62.05 | $0.085 |
| A16 · 4 GB VRAM | 2 | 16 GB | 80 GB | $124.1 | $0.170 |
| A16 · 8 GB VRAM | 3 | 32 GB | 170 GB | $240.9 | $0.330 |
| A16 · 32 GB VRAM | 12 | 128 GB | 700 GB | $963.6 | $1.32 |
| A40 · 48 GB VRAM | 24 | 120 GB | 1400 GB | $1750 | $2.40 |
Drivers are preinstalled, so nvidia-smi answers on first boot. Billing is hourly and stops when you destroy the server.
Deploy in 33 cities across 6 continents
🇳🇱 Amsterdam · 🇩🇪 Frankfurt · 🇫🇷 Paris · 🇬🇧 London · 🇬🇧 Manchester · 🇪🇸 Madrid · 🇵🇱 Warsaw · 🇸🇪 Stockholm · 🇫🇮 Helsinki · 🇩🇪 Nuremberg
🇺🇸 New Jersey · 🇺🇸 Chicago · 🇺🇸 Dallas · 🇺🇸 Atlanta · 🇺🇸 Miami · 🇺🇸 Seattle · 🇺🇸 Los Angeles · 🇺🇸 Silicon Valley · 🇨🇦 Toronto · 🇲🇽 Mexico City
🇧🇷 São Paulo · 🇨🇱 Santiago
🇯🇵 Tokyo · 🇯🇵 Osaka · 🇰🇷 Seoul · 🇸🇬 Singapore · 🇮🇳 Bangalore · 🇮🇳 Mumbai · 🇮🇳 Delhi NCR
🇦🇺 Sydney · 🇦🇺 Melbourne
🇿🇦 Johannesburg · 🇮🇱 Tel Aviv
Frequently asked questions
How much does a GPU server cost per hour?
From $0.085/hr for a fractional NVIDIA A16 with 2 GB of VRAM. Larger slices and whole A40 cards go up to $2.397/hr for 48 GB of VRAM. Billing is hourly and stops when you destroy the server.
Do I need to verify my identity to rent a GPU?
No. An email address is enough, the same as for any other server here, and you pay from a crypto balance. There are no documents, no phone number and no card on file.
Are NVIDIA drivers preinstalled?
Yes. The image comes with the driver loaded, so nvidia-smi works on first boot and you install nothing. Do not add drivers on top - a newer userspace against the loaded kernel module breaks nvidia-smi.
What is a fractional GPU?
One physical card split between several machines, with each machine getting its own slice of VRAM. Your VRAM is yours alone; the compute units are shared. It is what makes a GPU affordable at cents per hour instead of dollars.
What happens to my data if my balance runs out?
A GPU server is deleted rather than archived, because storing a disk that size would cost more than the server earns. You get a low-balance email in advance - keep enough balance for the run and copy results off the machine as you go.
Can I pay for a GPU server with Bitcoin or Monero?
Yes. Top up the balance with Bitcoin, Monero, USDT or any of 300+ coins, then deploy from the balance. The same balance covers GPU and regular servers.