Power
Sized to your subscribed load
Power feed, redundancy, generators and fuel autonomy, sized for the power you reserve.
AI clusters
Next-Generation AI-Native Datacenters for your GPU clusters
We host your GPUs in France, from a few racks to several megawatts. The land is secured, the equipment is in stock and our works team is ready: the first tranche is delivered in 4 to 8 weeks.
Estimate your cluster's power
NVIDIA
Vera Rubin NVL72 72 GPUs per rack, 220 kW
Vera Rubin Ultra NVL144 (Kyber) 144 GPUs per rack, 600 kW
GB300 NVL72 72 GPUs per rack, 140 kW
GB200 NVL72 72 GPUs per rack, 132 kW
HGX B300 40 GPUs per rack, 72 kW
HGX B200 40 GPUs per rack, 65 kW
HGX H200 48 GPUs per rack, 61 kW
HGX H100 48 GPUs per rack, 48 kW
RTX PRO 6000 Server 80 GPUs per rack, 63 kW
AMD
Instinct MI355X 80 GPUs per rack, 112 kW
Instinct MI325X 40 GPUs per rack, 65 kW
Network, storage, CPU nodes, other hardware
- Mellanox Quantum-X800 InfiniBand Counted per rack, 15 kW
- Mellanox Spectrum-X Ethernet Counted per rack, 12 kW
- Baie stockage Counted per rack, 8 kW
- Baie réseau et management Counted per rack, 6 kW
- Baie nœuds CPU Counted per rack, 12 kW
Other hardware kW per rack racks
Your configuration
Pick a hardware type to start.
The offer in figures
Capacity
40 MW IT available immediately, on more than 50 MVA of connected power
Density
150 kW per rack, a common density on our sites
Lead time
4 to 8 weeks for a first tranche up to 1 MW
Energy
1.35 PUE contractually capped
What we cover
Power, cooling and site operations are our responsibility
Density
One GB300 rack draws as much as ten classic colocation racks
:quality(65))
The situation
Power becomes the main constraint
In a classic colocation room, the limit is floor space. With GPU racks above 100 kW, the limit becomes the power and cooling available per square metre.
Our answer
Rooms designed for AI workloads from the outset
Our rooms are built for racks at 150 kW and above: power distribution and cooling are sized at construction.
Cooling
Liquid cooling adapted to your hardware
Rear-door, direct-to-chip or immersion: the architecture is defined with you during the technical study.
NVIDIA Vera Rubin
PAR9 can host Vera Rubin NVL72 and Kyber racks
Our PAR9 site in the Paris region is connected at 36 MVA and accepts racks well above 150 kW.
Vera Rubin NVL72
Each rack holds 72 Rubin GPUs and 36 Vera CPUs and draws about 190 to 230 kW. NVIDIA has been shipping it since the second half of 2026.
A 45 °C water inlet
Vera Rubin is fully liquid-cooled, directly on the chips, with a 45 °C water inlet. The PAR9 cooling loop accepts this temperature.
Kyber: 600 kW per rack on 800 VDC
NVIDIA plans Vera Rubin Ultra in the Kyber rack for the second half of 2027, at about 600 kW per rack powered by 800 VDC. PAR9 can host these racks.
Location
Two ways to site a cluster in France
Hub: Paris and Marseille
Close to networks and your teams
Peering, carrier interconnection, proximity to your teams and your customers. Suited to inference and production.
Edge: Lille, Grenoble, Castres
Available power and energy cost
More available power, more favourable energy costs and headroom for the next tranches. Suited to training and long-term deployments.
Energy
Electricity billed at market price, PUE capped by contract
Spot
Indexed on the market
Electricity rebilled at the month's spot price
Electricity consumed is rebilled monthly at the market spot price plus market, with no multi-year smoothing.
1.35
PUE cap
Datacenter overhead capped at 1.35
The PUE is contractually capped at 1.35: technical equipment consumption above this threshold is borne by us.
MW
What you reserve
Reserved power, energy billed on meter readings
The contract covers IT power made available at all times. Energy consumed is rebilled separately, on meter readings.
Handover
Everything is ready before your order
Your room does not wait for a site, for its equipment to be manufactured or for a team to be hired.
Sites already connected to the grid
The sites that will host your cluster are already secured and connected: 36 MVA in the Paris region at our PAR9 site, 10 MVA in Castres and 5 MVA in Lille, for 40 MW of IT power available immediately.
Your room's equipment, in stock
Chillers, gas and diesel generators, UPS systems and battery racks, transformers, raised floor, access control and building management systems are in stock in our logistics and reuse centres in western and northern France, in addition to our regular stock in the Paris region.
An in-house works team
Our works team of about 50 employees, on permanent and apprenticeship contracts, fits out your room: power distribution, cooling and room infrastructure. It is available to start your project.
Lead times
Three steps to power-on
1.
You describe your hardware
The configurator works out power, density and rack count from your hardware. The estimate takes five minutes.
2.
An engineer calls you back
Within 48 working hours, with the available capacity, the site and the delivery date.
3.
You deploy
400 kW to 1 MW in 4 to 8 weeks. Beyond that, 3 months. Your teams get access from delivery.
No middleman
You deal directly with the datacenter operator
DC2SCALE operates its own datacenters with its own teams, and its works team fits out the rooms. The estimate is available without an NDA and an engineer answers your request within 48 business hours.
Projects of 1 MW and more
A multi-megawatt project is prepared with a named contact
Give the target power and the timeline. Your sales contact calls you back within 48 business hours, arranges the site visit and signs an NDA before the discussion if your project requires it.
Your sales contact 
Your cluster
Describe your cluster project
In a few minutes the configurator gives you the power, the number of racks and the budget range of your project. An engineer then calls you back within 48 working hours.
:quality(50))