AI-ready datacenters: density, liquid cooling and power

Start from your hardware, not from facility brochures. A rack of current-generation GPU servers draws tens of kilowatts; accelerator roadmaps point higher with every cycle. When a facility quotes a maximum rack density, ask what it means in practice: sustained draw or peak, per isolated rack or across a contiguous row, and with what cooling method attached.

Averages mislead here. A hall that supports 40 kW on four scattered racks is a different proposition from one that supports 40 kW across the forty adjacent racks a cluster actually needs. Contiguity — power and cooling delivered to one physical block — is what makes a deployment work.

Above roughly 30–40 kW per rack, air cooling stops being practical and liquid becomes the enabler. The methods differ meaningfully. Rear-door heat exchangers retrofit onto existing racks and extend air-based designs. Direct-to-chip (D2C) circulates coolant through cold plates on the processors and carries the density mainstream today. Immersion submerges entire systems and suits specialised deployments.

“Liquid-ready” deserves scrutiny: it can mean anything from operating liquid-cooled clusters today to having pipework provisions in a design document. Useful questions: which method is operational on site today, at what density, for which customers, and what is the lead time to deliver it on your footprint?

AI deployments turn power conversations from kilowatts into megawatts, which puts grid reality at the centre of site selection. In constrained metros — Amsterdam, Dublin, parts of Frankfurt and London — multi-megawatt connections carry multi-year queues. Power-rich markets such as the Nordics, Spain and parts of France can deliver faster, at lower energy cost and with cleaner power.

The training/inference split matters commercially: training tolerates distance, so it can chase cheap, available, renewable power; inference follows users. Many organisations end up with exactly that pair of sites. Whatever the market, the diligence question is the same: how much power is contracted with the utility today, and on what schedule does the rest arrive?

High-density equipment is heavy — liquid-cooled racks can exceed common floor-loading limits, especially on raised floors in older buildings. Confirm structural capacity per square metre, plus practical logistics: delivery paths, freight elevators, staging space for pallets of servers, and crane access where needed.

Cluster networking adds its own requirements: high-fibre-count east-west cabling between racks, room for spine switches, and — if you train across sites or ingest large datasets — serious external capacity. For inference, proximity to internet exchanges and cloud on-ramps keeps latency and egress costs manageable.

A megawatt-scale AI deployment produces a megawatt-scale stream of heat, and European regulation increasingly treats that as a resource: Germany’s Energy Efficiency Act pushes operators toward heat reuse and efficiency reporting, and the EU’s Energy Efficiency Directive drives disclosure across member states. Facilities with district-heating connections or documented reuse plans are ahead of that curve.

For your own reporting, the carbon intensity of the local grid and the operator’s renewable sourcing (PPAs versus certificates) flow directly into the footprint of every training run. Nordic hydropower, French nuclear and Iberian renewables each give structurally different answers.

The label is unregulated, so verification falls to the buyer. Reference customers running comparable densities are the strongest evidence; an operational tour of a liquid-cooled hall beats any datasheet. Contract terms tell the rest: committed power with delivery dates, density guarantees on defined space, expansion options with timelines, and remedies if delivery slips.

On this platform you can filter facilities by rack density and cooling method, compare what operators actually report, and put your requirements to several of them in one request — their answers on the questions above are the real selection criteria.