Sovereign AI infrastructure in Europe

Sovereignty is decided layer by layer

Sovereign AI means you can train and run your models on infrastructure you control: where the hardware sits, who operates it, who holds the keys, and which law applies. It is not a label on a datacenter. It is a set of engineering decisions, taken on every layer of the stack.

OPSaaS is an engineering team based in France. We design, deliver and operate sovereign GPU clusters in Europe, from the datacenter constraints to the model endpoint, for companies, public bodies and operators who need AI capacity under European law.

The questions to ask on each layer

Layer	The question	A sovereign answer
Datacenter	Where is the hardware, and under which law?	A datacenter in the EU, run by a European entity, with contracted power.
Fabric and network	Who can reach the cluster?	Private fabric, tenants separated in the network itself, admin access from the EU, no vendor remote access by default.
Storage	Who can read the data?	Encryption at rest and in transit, keys in a KMS or HSM you control, backups in the EU.
Cloud and orchestration	Who runs the control plane?	Open-source control planes (OpenStack, Kubernetes, Slurm) that you can operate yourself, with no dependency on a managed service abroad.
AI platform	What leaves the cluster?	Models, datasets and telemetry stay inside. Registries and package mirrors are local.
Model serving	Where do the prompts go?	Inference served from your region, on open-weight models you host, with logs under your retention policy.
Operations	Who has root?	A European operations team, named accounts, just-in-time access, every action audited.

Hardware is the honest limit: GPUs are designed in the United States and manufactured in Asia. Sovereignty is not autarky. It is control over what runs on that hardware, who can access it, and a supply chain and firmware baseline you can trace.

Three levels of sovereignty

Not every workload needs the same level, and each level costs more than the one before.

1. Data residency. Data and compute stay in the EU. The operator may be non-European. This covers many GDPR-driven requirements.
2. Operational sovereignty. The cluster is in the EU, operated by a European entity and European staff, and you hold the encryption keys. This is a common target for organizations regulated under NIS2 or DORA.
3. Full sovereignty. No exposure to extraterritorial law, hosting qualified where the rules require it (SecNumCloud in France), air-gapped operation when needed. For defense, sensitive public data and critical operators.

Sector rules add their own requirements, such as HDS certification, issued by an accredited certification body, for hosting personal health data in France. We help you classify each workload first, then design for the level it needs. See how we build security into every layer.

What an AI factory is

An AI factory is a datacenter built to produce AI: it trains models and serves tokens at scale, the way a plant turns energy and raw material into a product. The term covers NVIDIA reference designs and the public AI Factories the European Union co-funds with its member states around EuroHPC supercomputers.

In practice, an AI factory is:

• rack-scale GPU systems such as NVIDIA GB300 NVL72, or 8-GPU HGX B300 nodes, at 40 to 130+ kW per rack;
• direct liquid cooling, with CDUs and usually a facility water loop;
• a rail-optimized fabric at 400G or 800G, InfiniBand or Ethernet for AI;
• parallel storage sized for checkpoints and dataloaders;
• an orchestration and serving stack: Kubernetes or Slurm, Ray, vLLM.

A private AI factory is the same thing at your scale, from a few nodes to several megawatts. We deliver the cluster under one contract, in your datacenter or in a European site we help you find and qualify: see turnkey GPU clusters.

What we do

• Assess. Classify workloads and data, and choose the sovereignty level each one needs.
• Design. Site, power, cooling, fabric, storage, isolation between tenants, key management and access model, written down in architecture decision records.
• Deliver. The cluster, turnkey, as the prime contractor: sizing, quote, hardware sourcing, bring-up and acceptance tests.
• Operate. From France, in European hours, with 24/7 on-call on demand. Runbooks, configuration and change history stay in your Git.
• Stay open. Kubernetes, OpenStack, Slurm, Ceph, vLLM: open-source components you can take over at any time. No lock-in to us either.

For the practical side, read our guide to deploying a GPU cluster in Europe.

Frequently asked questions

What is sovereign AI?

Sovereign AI is the ability to train and run AI models on infrastructure you control: hardware located where you decide, operated by people you choose, with keys you hold, under a law you accept. It applies to a country, a public body or a company.

Is a European region of a US cloud sovereign?

It gives you data residency: the data stays in the EU. The provider remains subject to US law, including the CLOUD Act, which can require it to disclose data it controls wherever that data is stored. Whether that matters depends on your data and your regulator.

Do we need SecNumCloud?

Only if your workload requires it, which is the case for some French public bodies and operators handling sensitive data. SecNumCloud is a qualification issued by ANSSI for a specific cloud service, not for a provider as a whole. OPSaaS does not hold one: we design to its requirements, and the qualification belongs to the hosting provider’s service.

Can we run open-weight models on our own cluster?

Yes. Open-weight models such as Mistral, Llama or Qwen run on your own GPUs behind vLLM, SGLang or TensorRT-LLM, with authentication, quotas and logs under your control. Prompts and outputs never leave your cluster.

How big is a private AI factory?

It starts with a few 8-GPU HGX nodes for inference and fine-tuning, which fit in an existing datacenter. A cluster for training at scale is in the order of 1,000 GPUs and 2 MW of IT power, with a dedicated fabric and, for rack-scale systems, liquid cooling.

Who can build a sovereign AI infrastructure in Europe?

OPSaaS, an engineering team based in France, designs, delivers and operates sovereign GPU clusters in Europe as the prime contractor, with its datacenter and hardware partners. See turnkey GPU clusters or talk to an engineer.