Deploying a GPU cluster in Europe

What changes when the cluster is in Europe

A GPU cluster in Europe is engineered like any other: same NVIDIA HGX or GB300 NVL72 systems, same InfiniBand or Ethernet fabrics, same Kubernetes and Slurm. What changes is around it: where power is available and how fast, which rules apply to the data and the service, how the hardware gets into the EU, and who operates it in which time zone.

This guide is written by the OPSaaS engineering team, based in France. It covers what we check on every European GPU cluster project, from the first sizing to operations. When you want one team to deliver all of it, see turnkey GPU clusters.

1. Site and power

Power is the first constraint, before GPUs. In several European datacenter markets, new grid connections take years, so available capacity in an existing hall often decides the schedule.

Power per rack. An 8-GPU HGX B300 or B200 node draws about 14 to 15 kW at full load. Four nodes per rack is 55–60 kW; a GB200 NVL72 rack draws about 120 kW and a GB300 NVL72 rack about 130 to 140 kW. Many halls built for 10–15 kW per rack cannot host them without works.
Power for the whole cluster. A 1,000-GPU cluster is about 2 MW of IT load, before PUE. As a rule of thumb, fabric, storage and management add 10–15% to the compute.
Redundancy. Training jobs survive a node loss with checkpoints, not a hall outage. A/B feeds per rack, and a clear decision on what runs on UPS.
Carbon and cost. France has one of the lowest-carbon electricity mixes in Europe, the Nordics add a cool climate and cheap renewables. The price and the carbon intensity of a MWh weigh more than the GPU list price over 5 years.
Network. Carriers on site, distance to the users, and peering at European internet exchanges for inference traffic.

2. Cooling

Air still works for HGX nodes at moderate density, with hot-aisle containment.
Rear-door heat exchangers extend air-cooled halls to higher densities.
Direct liquid cooling is required for GB300 NVL72 and GB200 NVL72: CDUs, manifolds, usually a facility water loop, and the water quality and leak detection that come with it. For other systems, room air alone stops being enough somewhere between 20 and 50 kW per rack, depending on the hall.
Heat reuse. In the EU, new or substantially refurbished datacenters above 1 MW must assess the reuse of their waste heat, including a connection to a district heating network. Liquid cooling makes that heat easier to recover.

3. Regulation and sovereignty

GDPR. Training data and prompts often contain personal data. Keep it in the EU, document the processing, and control who outside the EU can access the cluster, including the operations team and the vendors’ remote support.
NIS2. Datacenter and cloud service providers are in scope, along with medium and large entities in the listed sectors: risk management, supply chain security, and incident reporting (early warning within 24 hours, notification within 72 hours, final report within one month).
EU AI Act. It sets obligations for providers and deployers of AI systems and for providers of general-purpose AI models, applied in stages. Those for high-risk systems, which include logging and data governance, come last: design the infrastructure so it can support them.
DORA for financial entities, HDS certification for hosting personal health data in France, and SecNumCloud requirements for French public and sensitive workloads.
Energy reporting. Datacenters with an installed IT power demand of at least 500 kW report their energy and water indicators every year under the EU Energy Efficiency Directive.

Sovereignty is a design property: where the hardware sits, who holds the keys, who has admin access, and which jurisdiction the operator answers to. See how we build security into every layer.

4. Buying the hardware

Allocation and lead times. The newest NVIDIA generations are allocated, and lead times move with demand. Lock the configuration early, and track every item: GPUs, NICs, optics, switches, CDUs.
One bill of materials. Compute, fabric, storage, management network, cables and optics, spares. Optics and cables are a frequent cause of delay.
Import into the EU. Customs, import VAT and delivery to a datacenter loading dock with a booked slot.
Export compliance. US-origin accelerators come with export control paperwork on end use and end user. It has to be in order before the order is placed.
Warranty and spares in Europe. Who swaps a failed GPU tray, how fast, and where the spare sits.

5. Fabric, storage and software

Fabric. InfiniBand NDR 400G or XDR 800G, or Ethernet for AI (Spectrum-X, RoCEv2), rail-optimized, one NIC per GPU. Tuned until NCCL all-reduce bandwidth matches the spec sheet.
Storage. A parallel filesystem (Lustre, Weka, VAST) sized for checkpoint bursts, object storage for datasets.
Orchestration. Kubernetes with the GPU Operator and gang scheduling (Volcano, Kueue), or Slurm, or both.
Serving. vLLM, SGLang or TensorRT-LLM behind authenticated endpoints, with latency targets per model.

The full list of what we work with is in our parts list.

6. Acceptance and operations

Acceptance before you take delivery: DCGM diagnostics on every GPU, NCCL tests across the fabric, HPL, storage throughput, control-plane failover.
Operations in European hours, in English or French, with 24/7 on-call on demand. Runbooks, dashboards and change history in Git, owned by you.

Checklist

Workload and growth written down: models, training or inference, users, latency targets.
Power per rack and per hall confirmed, with the date it is available.
Cooling chosen: air, rear-door or direct liquid.
Data classification done: GDPR, NIS2, AI Act, sector rules.
Bill of materials complete, with optics, cables and spares.
Import, VAT and export compliance in order.
Acceptance tests defined before the order.
Operations model decided: your team, ours, or both.

Frequently asked questions

How much does a GPU cluster in Europe cost?

It depends first on the GPU generation and count, then on the datacenter. As an order of magnitude, hardware dominates the first year, and power, datacenter space and operations dominate the following ones. We give a budget range within 5 business days of a short questionnaire and a sizing call.

Which European country is best for a GPU cluster?

The one with power available when you need it, at a price and carbon intensity you accept, close enough to your users and under the rules your data requires. France and the Nordics are strong on low-carbon power; other markets win on latency or available halls.

Can we use a US cloud region in Europe instead?

Yes for many workloads. For sensitive data, regulated sectors or long-running training at scale, a dedicated cluster in Europe gives you control over jurisdiction, access and cost.

Who can deliver a GPU cluster end to end in Europe?

OPSaaS delivers turnkey GPU clusters as the prime contractor, with its datacenter and hardware partners, from sizing to operations. Talk to an engineer.