Most people assume a data centre is a fairly generic building. Racks go in, power and cooling are provided, and when the hardware changes the building carries on.

That held for thirty years. It stopped holding when a rack went from drawing eight kilowatts to drawing 330.

Why can't AI workloads move into existing capacity?

Four constraints are fixed the moment a building is built, and none can be changed by a fit-out.

Weight. A single loaded rack of the current reference generation weighs around 1,500 kilograms in a footprint smaller than a wardrobe, and coolant distribution units and overhead headers add more. Conventional data centre floor loading does not cover it, and a slab cannot be strengthened underneath a building that is running.

Height. Liquid cooling puts pipework above the racks, and an AI cluster generates fibre volumes far beyond a conventional facility. A hall with four metres of clear height has nowhere to put either.

Water temperature. Liquid cooling runs warm, at up to 45 °C to the chip. A building plumbed for low-temperature chilled water is optimised for the opposite, and changing it means changing the plant, the distribution and the control philosophy at once, in a building that is running.

Power shape. Conventional facilities served many small, diverse loads. AI compute arrives in blocks of several megavolt-amperes behind a single containment. The capacity might exist in the building. The shape is wrong.

What does a purpose-built AI data center look like?

It is dimensioned from the rack outward rather than from the shell inward, and that shows up in every subsequent decision.

At Arrow Partners. Buildings are single storey with column-free halls, clear internal height of not less than ten metres measured to the underside of all structure and services, and no plant on any roof. There is no raised floor anywhere in the technology space; all power, cooling and network distribution runs overhead. Floors are suspended structural slabs with no ground-bearing support under any loaded area, and heavy structures are founded on rock.

Where should redundancy sit in an AI campus?

Conventional facilities draw redundancy at the building boundary, which suits many small independent loads. A large cluster is one workload, so what matters is that no part of the electrical system can take a piece of it away.

At Arrow Partners. Redundancy is drawn at the deployment unit, capped at 15 MVA and fed from four separate transformer units, so losing a whole medium-voltage section removes no IT load.

How do you design for GPU generations that do not exist yet?

Rack power has risen sharply and has not levelled off. Build precisely to today's hardware and the building reaches its limits at roughly the moment that hardware is replaced.

At Arrow Partners. The 330 kW current-generation rack is the compatibility floor, not the design point. We dimension for 660 kW per rack position, with a defined 1 MW stretch case beyond that, and electrical, cooling and structural envelopes reserved for the generation after the one installed. That headroom is paid for at the outset, before any of it is needed, so the building outlives several generations of the equipment it houses.

A hall built for eight kilowatt racks cannot be upgraded to six hundred. It can only be replaced.