Circular Datacenter
Cooling

Removing heat, at AI scale.

Rack densities have changed by an order of magnitude. Cooling infrastructure built for conventional IT no longer keeps up.

Data center cooling continuously removes the heat produced by computing equipment while holding temperature and humidity within acceptable ranges. It combines a cooling plant, a distribution network and in-room delivery. Artificial intelligence workloads have moved the problem: at densities of several tens of kilowatts per rack, air alone is no longer sufficient and direct liquid cooling becomes the reference.

Chiller being lifted into position on a data center

A change of scale, not an evolution

For twenty years, cooling a data center meant blowing cold air under a raised floor for racks drawing a few kilowatts. Today's training and inference workloads dissipate per rack what an entire row once did. At those densities the heat capacity of air becomes the limiting factor: the required airflows exceed what ductwork and acoustic comfort allow.

Direct liquid cooling answers that physical wall by bringing fluid into contact with the components. It changes the nature of the infrastructure: no longer only air handling units, but hydraulic loops, in-room distribution, water quality and leak-tightness above live electronics.

For an existing site, the practical consequence is that an AI-oriented retrofit is first a hydraulic project. It is also what makes anticipation pay: making a building liquid-ready before the need arises costs a fraction of the same work carried out later, in operation.

Free cooling, where the climate allows

In the French climate, outdoor temperatures allow cooling without compression for a large part of the year. Exploiting that resource - by direct outdoor air exchange, by coil on a water loop, or adiabatically - cuts consumption by margins few other measures reach.

The trade-off is not only about energy. Free cooling requires roof or façade area, admits a degree of humidity that must be controlled, and calls for controls that handle mode transitions cleanly. We assess this choice on site data - the actual weather file for the location and the expected load profile - rather than on national averages.

Where this expertise applies

The data center packages this expertise covers are shown in full, the others are dimmed. Click a package for detail.

RTE / ENEDIS INTAKE

Technical scope

From plant to containment, in design as well as installation.

Cooling plant
Air- or water-cooled chillers, new or refurbished, including handling and positioning.
Free and adiabatic cooling
Sizing on site weather data, heat exchangers, towers, and control of transitions between modes.
Hydraulic networks
Primary and secondary circuits, pumps, valves, balancing, water treatment and quality, insulation.
Direct liquid cooling
Technical loops, coolant distribution units, rack connection, and building infrastructure ready to accept DLC.
In-room delivery
Computer room air handlers, air handling units, ductwork and air distribution.
Containment
Hot or cold aisle containment, blanking, floor and cable penetration sealing.
Controls and supervision
Local controls, reporting to the building management system, measurement points, alarms and efficiency monitoring.

How a cooling engagement runs

Thermal design rests on site data: that is what separates sizing from estimating.

  1. Load profile and site data

    Current and target load, rack densities, local weather file, roof and footprint constraints.

  2. Cooling scenarios

    Costed comparison of the viable architectures, with their consequences for consumption, footprint and headroom for higher density.

  3. Design and simulation

    Thermal balances, hydraulic sizing and, on sensitive points, airflow simulation of the rooms.

  4. Procurement

    New or refurbished. On high-capacity chillers, reuse shortens lead time appreciably.

  5. Installation

    Positioning, hydraulic and electrical connection, insulation, pressure testing and flushing.

  6. Balancing and commissioning

    Network balancing, control tuning, load testing and verification of transitions between modes.

Frequently asked questions

At what density does liquid cooling become necessary?

Air remains workable up to around ten kilowatts per rack with good containment, and can be pushed beyond at the cost of rising airflow and noise. Current AI training workloads far exceed those figures, which makes direct liquid cooling unavoidable. The real threshold depends as much on room geometry as on the stated density: it is a matter for simulation, not a general rule.

Can an existing data center be made liquid-cooling ready?

Yes, and it is a growing part of our work. It requires bringing a hydraulic loop to the rooms, verifying that the structure accepts the water loads, and treating leak detection above live electronics seriously. Carrying out that work before it is needed costs considerably less than doing it later, in operation.

Is free cooling worthwhile everywhere in France?

It delivers a gain across the country, but its size varies markedly with the local climate and with the water temperature the installation accepts. The warmer the water the loop tolerates, the more operating hours become available - that parameter matters more than latitude. We assess the trade-off on the site weather file.

Does a refurbished chiller hold its performance?

Provided it is properly requalified: refrigerant circuit inspection, leak-tightness, replacement of wearing parts, documented performance testing. We only return machines to service whose efficiency has been measured, not merely found to run. The decisive advantage remains lead time on high capacities.

Who are you, exactly?

Three very different people land on these pages. Take the door that fits you.

Related expertise