GPUs for AI model training, hosted in France.
AI training needs GPU memory, fast links between GPUs and an environment you control. Prototype LLMs and vision models on NVIDIA GPU instances hosted in Europe, then scale to reserved multi-node clusters on our own servers in Paris and Lyon for pre-training and continued training.
1 to 8 GPUs per instance · Multi-node clusters with InfiniBand · Managed Kubernetes or Bare Metal deployment
What is AI model training?
AI training is the process of teaching a model from data: the model processes large numbers of examples and adjusts its weights until its predictions improve. For large language models and vision models, this means running the same computation across many GPUs for hours, days or weeks.
Two forms are common. Pre-training builds a model from scratch on a broad corpus and needs multi-node GPU clusters. Continued training starts from an existing open-source model and extends it with domain or language data. In both cases, three things decide the outcome: GPU memory, bandwidth between GPUs, and how efficiently jobs are scheduled.
How to train an LLM, in five steps
- Choose a starting point: an open-source model for continued training, or random weights for pre-training.
- Prepare your dataset: clean, deduplicate and tokenize it.
- Validate your code on a single Cloud GPU, billed per second.
- Scale out: up to 8 GPUs per instance, then a multi-node GPU Cluster with distributed training.
- Evaluate and deploy the model on Serverless GPU.
Pre-training
Continued training
Starting point
Random weights
An existing open-source model
Data
Large, general-purpose corpus
Domain, language or company data
Compute
Multi-node GPU clusters
One node to a small cluster
Typical setup
Reserved GPU Clusters
Cloud GPU or GPU Clusters
What makes training hard, and how we solve it.
Moving from a notebook experiment to a multi-node training run raises the same four questions for every team.
Compute
Scale without re-platforming
Start on one GPU to validate your pipeline, move to 8 GPUs on a single instance, then to a multi-node cluster, without changing provider.
Interconnect
Fast links between nodes
Distributed training depends on how fast GPUs exchange data. Our GPU Clusters connect nodes with InfiniBand networking.
Data residency
Your datasets stay in France
Arkane Cloud is a Sovereign AI Cloud: GPU Clusters run on our own servers in Paris and Lyon, and Cloud GPU instances are hosted in Europe. Clusters are single tenant, isolated for your organization: no noisy neighbors.
Operations
Orchestration handled for you
Managed Kubernetes or Bare Metal deployment: we handle provisioning, resource allocation and maintenance. You keep full access.
The right GPUs for each stage of training.
Prototype on on-demand instances, train at scale on a reserved cluster, then serve the trained model. One provider, from first experiment to production.
Per second
Cloud GPU
On-demand NVIDIA GPU instances hosted in Europe, 1 to 8 GPUs, SSH access, billed per second. Install your own training stack, run your experiments and stop when you are done. No commitment.
Best for: prototyping, single-node training, experiments
Explore Cloud GPU →Reserved
GPU Clusters
Dedicated multi-node clusters with B300 or RTX PRO 6000 GPUs, InfiniBand networking, managed Kubernetes or Bare Metal deployment, on our own servers in Paris and Lyon. Single tenant, reserved for 2 to 5 years, with priority access to the latest GPUs.
Best for: pre-training and large distributed runs
Explore GPU Clusters →Per second
Serverless GPU
Serve the model you trained: deploy any vLLM-compatible Hugging Face model, private repos included, behind an OpenAI-compatible endpoint with scale-to-zero.
Best for: putting your trained model in production
Explore Serverless GPU →Recommended GPUs
On reserved GPU Clusters, our own servers in Paris and Lyon: B300 (288 GB HBM3e) for pre-training, multi-node runs and the largest models · RTX PRO 6000 Blackwell (96 GB) for mid-size models and continued training. To prototype, use Cloud GPU: NVIDIA GPU instances hosted in Europe.
Orchestration
Managed Kubernetes for containerized ML workflows, or Bare Metal deployment for full control of your training stack. We run provisioning and maintenance; you get full access and monitoring.
Frameworks and storage
Bring your usual stack: PyTorch with DDP or FSDP, DeepSpeed and Megatron-LM, with NCCL over InfiniBand for multi-node communication on GPU Clusters. Keep datasets and checkpoints on block volumes attached to your instances.
GPU memory rule of thumb
Full training with Adam in mixed precision needs roughly 16 to 20 bytes per parameter, before activations (indicative). A single GPU with 80 to 96 GB fits full training of models up to a few billion parameters, on a Cloud GPU instance. A 7B model needs about 112 to 140 GB: a multi-GPU Cloud GPU instance, or one B300 (288 GB) in a reserved GPU Cluster. A 70B model needs over 1 TB, sharded with FSDP or DeepSpeed across a reserved B300 GPU Cluster.
From first training run to production.
Prototype
Validate your data pipeline and training code on a single Cloud GPU, billed per second.
Scale up
Run full single-node jobs on up to 8 GPUs per instance, with SSH access.
Train at scale
Launch multi-node runs on a reserved GPU Cluster with InfiniBand, managed Kubernetes or Bare Metal deployment.
Deploy
Push your weights to Hugging Face and serve them on Serverless GPU, billed per second.
Which product for your training workload?
A quick guide to match each stage of a training project with the right product.
Workload
Recommended
Why
Comparing base models before training
Test open-source candidates per token before committing GPU time
Prototyping a training pipeline
A single GPU billed per second, no commitment
Continued training of an open-source LLM
Up to 8 GPUs on one node, or a cluster for larger models
Pre-training a model from scratch
Multi-node, InfiniBand, reserved capacity with managed Kubernetes or Bare Metal deployment
Serving the trained model
Deploy from Hugging Face, billed per second, scale to zero
Questions about AI training
What is AI training?
AI training is the process of teaching a model from data. The model processes many examples and adjusts its weights until its outputs improve. For LLMs and vision models, it runs on GPUs, often many of them in parallel.
Should I train a model or fine-tune an existing one?
Fine-tune when an open-source model already does most of what you need and you want it to follow your domain, format or tone: it takes far less data and GPU time. Train, from scratch or with continued training, when you need a new language, broad new knowledge or your own architecture. See LLM fine-tuning for the fine-tuning workflow.
How do you train an LLM?
Choose a starting point (an open-source model for continued training, or random weights for pre-training), prepare your dataset, validate your code on one GPU, then run distributed training across several GPUs or nodes. Evaluate the result and deploy it for inference.
What is distributed training?
Distributed training splits a training job across several GPUs or several servers so that larger models and datasets fit and runs finish sooner. Across servers, the network between nodes matters: our GPU Clusters use InfiniBand.
Which GPU is best for AI training?
It depends on model size. On reserved GPU Clusters, B300 (288 GB HBM3e) suits pre-training, large multi-node runs and the largest models, and RTX PRO 6000 Blackwell (96 GB) suits mid-size models and continued training. To prototype, Cloud GPU offers NVIDIA GPU instances hosted in Europe, from 1 to 8 GPUs per instance.
Where is my training data processed?
GPU Clusters run on our own NVIDIA GPU servers in colocation datacenters in Paris and Lyon, France, and are single tenant and isolated for your organization. Cloud GPU instances are hosted in Europe.
How is GPU training billed?
Cloud GPU instances are billed per second, with no commitment. GPU Clusters are reserved for 2 to 5 years, on a quote. Serverless GPU, used to serve the trained model, is billed per second.
Other use cases
Inference
Serve open-source and custom models through an OpenAI-compatible API.
Fine-tuning
Adapt open-source models to your data and your domain.
Agentic AI
Power agents and tool-calling workflows with sovereign inference.
Coding
Use open-source coding models in your IDE with our VS Code extension.
Train your next model in France.
Start on a single GPU today, and request a dedicated cluster when your runs outgrow one node.
Cloud GPU billed per second, no commitment · GPU Clusters reserved 2 to 5 years