GPU servers with one whole NVIDIA card each.
Each plan passes one NVIDIA card through to the server, beside dedicated vCPUs and an NVMe disk. Every plan includes 20 TB of traffic a month, and runs in Milan and Riga.
From €499.00 a month
Features
Cards for training, tuning and inference.
A whole card, passed through
Each card belongs to one server through PCIe passthrough, with no slicing and no neighbours. All of its memory is yours.
Three cards
The L4 for inference and video, the L40S for fine-tuning and rendering, and the H100 for training large models.
Driver and CUDA installed
The Ubuntu and Debian GPU images carry the NVIDIA driver, CUDA and the container toolkit. Your own image works as well, with your own driver.
Reached from every region
The private network brings the cards to the rest of a project. Riga is 5 ms from Tallinn, and Milan 9 ms from Vienna.
Snapshots between runs
Snapshot the disk when a run ends and delete the server. Create it again from the snapshot for the next run.
GPU node pools
Add GPU servers to a Kubernetes cluster as a node pool, with the NVIDIA device plugin set up for you.
Pricing
Three plans, one card each.
Each plan’s card, its hourly rate and the most a month of it costs. Every plan includes an IPv4 address, an IPv6 /64 and 20 TB of traffic.
GPU servers
| Plan | GPU | Memory | A month, at most |
|---|---|---|---|
g-l4 | NVIDIA L4 24 GB | 48 GB | €499.00 |
g-l40s | NVIDIA L40S 48 GB | 96 GB | €949.00 |
g-h100 | NVIDIA H100 80 GB | 192 GB | €1,799.00 |
GPU servers run in Milan and Riga. Each further TB of outbound traffic costs €1.00.
Code
A card in one command.
Create a GPU server like any other, with a GPU plan and a region that holds the cards.
- The NVIDIA driver and CUDA in place from the first boot.
- Volumes for datasets, attached before the job starts.
- Snapshots of the disk to pick up the next run.
$ lc server create train-1 --plan g-l40s --region it1 \
--image ubuntu-24.04-nvidia --ssh-key me
Creating train-1 in it1 ... running
IPv4 203.0.113.41
IPv6 2001:db8:4c1::41
$ ssh root@203.0.113.41 nvidia-smi -L
GPU 0: NVIDIA L40S (UUID: GPU-5f1c2a7e-0b9d-4e61-a3c8-2d7e9f04b1a6)
Questions
Questions about GPU servers.
Why only Milan and Riga?
Can I share one card between servers?
Does a stopped GPU server cost anything?
Works with
Cloud servers
AMD EPYC vCPUs, an NVMe disk and 20 TB of traffic a month. Shared vCPUs for most work, and dedicated ones for steady load.
Object storage
S3-compatible buckets in each region, for backups, media and datasets. Billed by the TB stored, with no charge per request.
Kubernetes
Conformant Kubernetes with the control plane run for you and not billed. The nodes are cloud servers at server prices.
Start with one server and one command.
Choose a region and a plan, and add an SSH key. The server answers SSH 38 seconds after the command, at the price on its spec sheet.