Turnkey GPU clusters in Europe
One contract, from sizing to the first job
You need GPU capacity in Europe, not a dozen suppliers to coordinate. OPSaaS delivers turnkey GPU clusters as the prime contractor: one contract, one quote, one engineering team accountable from the first sizing note to the first training job, and through operations after that.
With our datacenter and hardware partners, we take care of the site search, facility design studies, the bill of materials, hardware sourcing, the coordination of logistics and import into the EU, integration, bring-up and acceptance testing. Our engineers, based in France, design the cluster and stay on it once it runs.
We work on NVIDIA HGX B300, B200 and H200 systems, rack-scale NVIDIA GB300 NVL72 and GB200 NVL72, and AMD Instinct MI355X, with InfiniBand, Spectrum-X or RoCEv2 fabrics, parallel storage, Kubernetes or Slurm, and the serving stack on top (vLLM, SGLang, TensorRT-LLM).
Who it is for
- AI companies and neoclouds that need GPU capacity in Europe they control, without building a datacenter team first.
- Enterprises and public bodies with sovereign AI projects: data that must stay in the EU, under GDPR, NIS2 or sector rules (DORA, HDS certification, SecNumCloud requirements).
- Research labs and HPC centers adding an AI partition next to an existing cluster.
- Datacenter and colocation operators that want halls ready for 40–130+ kW racks and liquid cooling, with a cluster delivered inside.
How a project runs
1. Sizing
We start from the workload, not from a GPU count: models, training or inference, dataset sizes, concurrency, latency targets, growth over 3 years. Output: a sizing note with the GPU generation, node count, fabric, storage and the power and cooling it implies.
2. Site, power and cooling
Your datacenter, a colocation you already use, or a site we help you find and qualify in Europe with our datacenter partners. We check available power per rack and per hall, cooling (direct liquid cooling is required for GB300 NVL72 and GB200 NVL72), floor loading, network carriers, and when the capacity is actually available.
3. Design
Rack elevations, power and cooling distribution, rail-optimized fabric topology, storage tiers, management and out-of-band networks, security zones, and the software stack. Every choice is written down in an architecture decision record.
4. Bill of materials and quote
We build the complete bill of materials: GPU systems, fabric switches, NICs, optics and cables, storage, management network, racks, power and cooling equipment, spares. Then one quote with that bill of materials, the services, the schedule and the assumptions behind them. Hardware, datacenter, engineering and operations in the same document, so you compare one number, not ten.
5. Sourcing and logistics
We source every line and compare the options for your workload: system vendors and configurations, InfiniBand or Ethernet, price, lead time, power draw and support in Europe. You see the alternatives and why we recommend one. Hardware is sourced through our supply partners, with lead times tracked per item. With them and your purchasing team, we coordinate delivery into the EU, import and customs, and the export formalities that come with US-origin accelerators. Spares and warranty are planned before the first rack ships.
6. Integration and bring-up
Racking, cabling, firmware and BMC baselines pinned per hardware generation, OS and drivers, fabric configuration, storage, Kubernetes or Slurm, monitoring. Everything provisioned as code and kept in Git.
7. Acceptance
Burn-in and acceptance tests before you take delivery: DCGM diagnostics on every GPU, NCCL all-reduce bus bandwidth across the whole fabric, HPL, storage throughput under checkpoint load, and failover of the control plane. You receive the results, not a statement that it works.
8. Operations
We operate the cluster with your team or hand it over: runbooks, dashboards, alerting, upgrades and incident response. EU and US time zones, with 24/7 on-call on demand. See how we diagnose a fault across layers.
Orders of magnitude
Early figures to frame a project, not a quote. Real numbers come out of the sizing step.
| Cluster | GPUs | Racks (compute) | IT power, order of magnitude | Cooling |
|---|---|---|---|---|
| 32 × HGX B300 nodes | 256 | 8–16 | ~0.5 MW | air, rear-door or liquid |
| 128 × HGX B300 nodes | 1,024 | 32–64 | ~2 MW | rear-door or liquid |
| 14 × GB300 NVL72 racks | 1,008 | 14 | ~2 MW | direct liquid cooling |
| 56 × GB300 NVL72 racks | 4,032 | 56 | 8–8.5 MW | direct liquid cooling |
An 8-GPU HGX B300 or B200 node draws about 14 to 15 kW at full load. A GB200 NVL72 rack draws about 120 kW, a GB300 NVL72 rack about 130 to 140 kW. As a rule of thumb, fabric, storage and management add 10–15% to the compute load, and the facility adds its PUE on top.
Why Europe, why from France
- Sovereignty. The cluster, the data and the people who operate it can all stay in the EU, under EU law.
- Regulation by design. GDPR, NIS2, DORA, the EU AI Act, and French SecNumCloud and HDS requirements are part of the design, not an audit finding. See security.
- Energy. Low-carbon electricity in France and the Nordics, heat reuse where the site allows it, and the yearly energy and water reporting the Energy Efficiency Directive requires from datacenters of 500 kW of installed IT power and above.
- Same time zone. Our engineers work in your hours, in English or French.
For the engineering details of a European deployment, read our guide to deploying a GPU cluster in Europe.
Frequently asked questions
How long does it take to deliver a GPU cluster?
It depends on the configuration and on the site: the hardware generation, the quantities, and whether power and cooling are ready. We give you the lead times for your configuration with the first sizing, then a dated schedule with the quote. A hall that has to be prepared for liquid cooling or new power adds its own schedule.
Do you sell the hardware?
Yes, as part of a turnkey project: we build the complete bill of materials, compare the options for your workload and source the hardware through our supply partners, in the same quote as the rest. You can also bring hardware you already bought, and we take it from there.
Can you deliver in our own datacenter or colocation?
Yes. We work with your facility team and your provider. If you have no site, we help you find and qualify one in Europe with our datacenter partners.
Do you own or host the datacenter?
No. OPSaaS designs, delivers and operates the cluster. The datacenter is yours or your colocation provider’s. When you have none, we help you find one.
Do you only work in Europe?
No. Europe is our home market, and we deliver and operate clusters for clients worldwide, in EU and US time zones.
Where do we start?
With a short questionnaire (workload, target date, site, power) and a 30-minute call with an engineer. Within 5 business days, you receive a first sizing, a budget range and the lead times for your configuration. Talk to an engineer.
Need this in production?
Tell us what you are building. An engineer replies.