Service · high density

GPU colocation in Amsterdam: 20–70 kW racks, liquid cooled

NVIDIA rates an eight-GPU DGX B200 at 14.3 kW, so one node already fills an air-cooled rack. Above roughly 15 kW a rack you are choosing a hall rather than a rack, and the constraint is cooling and reserved power, not floor space.

What you get

A hall that can cool it, and the power booked.

Halls rated for the load

Rear-door heat exchangers or direct-to-chip liquid, in a hall whose power ceiling is not already sold. There are fewer than you would hope.

Cooling

Power reserved before you buy GPUs

We find a hall with the kW free and reserve it quickly. The purchase order follows, not the other way around.

Reserved first

Sized on real draw

Training racks run near nameplate, inference racks do not. We size the commit on the workload instead of on the spec sheet.

Measured

An honest answer about building your own

Your office has neither the feed nor the cooling plant, and Liander expects to connect part of its Amsterdam waiting list only at the end of 2037.

Advice

Specs that matter

Density, cooling, weight, lead time.

High density is a queueing problem as much as an engineering one. Every figure below is about whether the hall can take you this year.

Density: 20–70 kW per rack with liquid cooling is the band we place. NVL72-class racks draw 120 kW and more and need a hall built for it. Most Amsterdam air-cooled halls are provisioned for about 10–15 kW a rack, whatever the rack is rated for.
Cooling: Rear-door heat exchangers or direct-to-chip liquid. Immersion is rare in Amsterdam and rarely on offer.
Where: Digital Realty’s AMS11 in Schiphol-Rijk offers direct liquid cooling next to air; elsewhere in the metro it depends on the hall. Which hall has room changes by the quarter.
Power reservation: We find a hall with the kW free and get the block reserved quickly, so the power is ready before the hardware lands.
Weight: A filled liquid-cooled rack can pass 1,200 kg. Floor loading and the route from the loading bay are real constraints.
Connectivity: Same building, same meet-me room: IX ports, cross-connects and dark fibre to Nikhef.

How we place it

Four steps, from GPU count to install day.

01
Tell us what you needGPU count, node type, kW per rack, cooling you can accept, and the date the hardware lands.
02
We engineer and negotiateWe shortlist halls that can cool it and reserve the power block before the price moves.
03
We fast-track the introductionThe operator gets your density, cooling type and delivery date in one brief, so capacity is checked in days.
04
You contract with the datacenterContract, power commit and cooling terms with the datacenter. We sit in the install.

Read the full process

Who we’ll send elsewhere

Not for one GPU box, and not for building your own room.

One or two PCIe GPU servers fit in a normal rack in a normal hall; you do not need this page and we will place them without ceremony. At the other end, if you are planning to turn an office floor into a GPU room, the numbers are against you: the grid connection, the cooling plant, the fire suppression and the standby generator are a building project of their own, and a new heavy connection in Amsterdam is measured in years, not months. We will say that before you spend anything.

Questions we get

Density, cooling technique, reserving power, mixed racks.

How dense can a rack go here?

20 to 70 kW with liquid cooling is the usual band. NVL72-class racks draw 120 kW and more and need a hall built for direct liquid cooling, so we check each hall’s ceiling. Air-cooled halls stop around 10 to 15 kW no matter what the rack is rated for, and availability decides it more often than physics.

Rear-door or direct-to-chip?

Rear-door heat exchangers work with standard servers and are easier to get. Direct-to-chip goes denser but has to match your node vendor’s loop. In practice the hall decides which one is on offer. The questions to ask each operator about its loop are in Digital Realty vs NorthC.

How early do I reserve power?

Before the hardware order. Tell us the kW and we make it available quickly: we find the hall with room and get the block reserved, so the power is ready when the servers arrive. Why the block has to come first is in Capacity in Amsterdam in 2026.

Can I mix GPU and normal racks?

Yes, and usually you should. Storage and control-plane racks do not belong in a liquid-cooled row at liquid-cooled prices.

Is Amsterdam the right metro for training?

For anything that talks to European users or European data, yes. For a pure training farm where latency is irrelevant and power is the whole cost, we will say honestly that a cheaper metro may win.