On-demand cloud GPUs, hosted in Europe.
Launch B200, H200, H100 and more as GPU instances: 1 to 8 GPUs, SSH access, full control of your stack, billed per second. Sovereign AI Cloud: GPU instances hosted in Europe, operated by a French company.
1 to 8 GPUs per instance · SSH access · Billed per second · Read the docs ↗
Your GPU instance in three steps.
01
Pick your configuration
Choose the GPU type and from 1 to 8 GPUs per instance, in the console or through the API.
02
Add your SSH key and launch
Provide your SSH key and launch the instance. You get the machine's IP to connect.
03
Run your workloads
Install your frameworks and run training, fine-tuning or inference with full control of the environment.
Full control of your cloud GPU, without the hardware.
GPU instances for teams that want to run their own stack, from a single experiment to production workloads.
Control
Your environment, your stack
Virtual machines you configure end to end: PyTorch, vLLM, ComfyUI or your own tools.
Hardware
The latest NVIDIA GPUs
Blackwell, Hopper, Ampere and Ada Lovelace GPUs, from cost-efficient inference cards to B200.
Pricing
Billed per second
An hourly rate per GPU, charged to the second while your instance runs. No commitment required.
Storage
High-performance storage
Fast NVMe and persistent storage to keep your GPUs fed with data and your work safe between sessions.
Automation
Console or API
Launch instances from the console, or send their parameters directly through the API from your own tools.
Sovereignty
Hosted in Europe
Sovereign AI Cloud: GPU instances hosted in Europe, operated by a French company.
Explore each GPU: B200 · H200 · H100 · A100 · L40S
Want to run models on RTX PRO 6000 or B300? Use Serverless GPU, on our own servers in Paris and Lyon.
Cloud GPU pricing and configurations.
Hourly price per GPU, billed per second, with the vCPUs and RAM included in each configuration.
GPU
Price per GPU
vCPU
RAM
VRAM
$7.99/hr
30
184 GB
180 GB
$5.99/hr
44
182 GB
141 GB
$3.99/hr
32
185 GB
80 GB
$2.29/hr
22
120 GB
80 GB
$1.79/hr
20
60 GB
48 GB
NVIDIA RTX 6000 Ada
$1.49/hr
10
60 GB
48 GB
NVIDIA RTX A6000
$0.99/hr
10
60 GB
48 GB
Pay as you go, or reserve capacity.
Start on demand with no commitment. When your needs become predictable, secure your GPUs for the years ahead.
On demand
Launch and stop instances whenever you need. You pay the hourly GPU rate, billed per second, only while your instance runs.
Reserved
Commit to a fixed number of GPUs for 3 months to 5 years (the minimum term depends on the GPU) to secure capacity for training runs and steady production workloads.
Hybrid
Combine a reserved baseline with on-demand instances for peaks, managed from the same console.
One instance or a whole cluster?
Both are hosted in Europe and operated by a French company. The difference is scale and commitment.
Cloud GPU
GPU Clusters
You get
A GPU virtual machine with 1 to 8 GPUs
A dedicated multi-node GPU cluster with InfiniBand networking
Commitment
None, billed per second
Reserved, 3 months to 5 years
Access
SSH access to your instance
Managed Kubernetes or Bare Metal deployment
Best for
Development, fine-tuning, single-node training, batch inference
Large-scale training and high-volume inference
Cloud GPU for inference, fine-tuning and training.
Inference
Serve models in production with consistent, low-latency responses, from chatbots to computer vision.
Explore inference →Fine-tuning
Adapt open-source models to your data and your domain on a multi-GPU instance.
Explore fine-tuning →Training
Train language, vision and deep learning models on the latest NVIDIA GPUs, then scale to a cluster.
Explore training →Questions about Cloud GPU
What is a cloud GPU?
A cloud GPU is an NVIDIA GPU you rent on demand instead of buying hardware. With Arkane Cloud you launch a virtual machine with 1 to 8 GPUs, connect over SSH and pay per second. Instances are hosted in Europe.
How is Cloud GPU billed?
You pay an hourly rate per GPU, billed per second, only while your instance is running. Each configuration includes its vCPUs and RAM. No commitment is required.
How many GPUs can I get per instance?
From 1 to 8 GPUs per instance, depending on the GPU model. For multi-node workloads, our GPU Clusters give you a dedicated cluster.
How do I access my instance?
Add your SSH key when you launch the instance, then connect over SSH. You install and run your own stack: frameworks, libraries and tools.
What storage is available?
You choose an OS image when you create the instance, and you can add block volumes for your data. A volume is attached to one instance at a time: detach it, then attach it to another instance to reuse your data. When you delete an instance, its volumes are kept and detached by default. Volumes are billed per GB while they exist, even when detached: see storage pricing.
Can I reserve GPUs for a longer period?
Yes. You can reserve capacity for 3 months to 5 years, depending on the GPU, or combine a reserved baseline with on-demand instances. Talk to our sales team to plan it.
Which GPU should I choose?
L40S and RTX 6000 Ada are cost-effective choices for inference and development. H100 and H200 suit large models and production workloads. B200, with 180 GB of memory, handles the largest models and long contexts. To run models on RTX PRO 6000 or B300, on our own servers in Paris and Lyon, use Serverless GPU.
Where are the GPUs hosted?
Cloud GPU instances are hosted in datacenters in Europe. Sovereign AI Cloud: GPU instances hosted in Europe, operated by a French company.
Launch your first GPU instance.
Pick a GPU, add your SSH key and start working. No commitment.
Billed per second · 1 to 8 GPUs per instance · Hosted in Europe