I have watched several AI infrastructure programmes stall at the same point, and it is almost never the model, the GPU or the budget. It is the building.

A team does the work properly. They profile the workload, size the cluster, shortlist accelerators, get a quote approved. Then someone finally walks down to the datacenter and asks facilities what the available power per rack actually is. The answer comes back at a fraction of what the proposed platform draws, and the whole design has to be rebuilt — after the business case has already been signed off.

Establish the facility constraint first. Not fourth.

Here is the chain to verify, end to end, before you shortlist anything:

• Utility feed capacity and remaining headroom

• Transformer and switchgear capacity

• UPS capacity, runtime and redundancy model

• Generator capacity and fuel autonomy

• PDU and busway ratings at the rack

• Circuit configuration and phase balancing per rack

Every link is a potential ceiling. The lowest one is your real capacity, and it is rarely the one on the original design drawing.

Then the cooling question, which is really a building question.

Air cooling with proper containment has a hard practical density ceiling. High-density GPU racks go straight through it. Rear-door heat exchangers buy you meaningful headroom and are usually the pragmatic retrofit. Direct-to-chip liquid cooling is required for the highest-density platforms — and that means a facility water loop, CDUs, manifolds, leak detection and a works package with its own schedule and its own budget line.

Liquid cooling is a construction project. If your platform needs it and your building does not have it, that has to be scoped and costed inside the same programme, not discovered during installation.

Three more things that get missed at the same stage:

1. Floor loading. Populated high-density racks are heavy and raised floors have limits.

2. Cable pathway capacity. A rail-optimised fabric produces a volume of fibre that existing trays were never sized for.

3. The delivery path. Service clearance front and rear, and whether the racks physically fit through the route into the room.

What to demand from any vendor, in writing, at bid stage:

• A rack elevation with unit-by-unit placement

• Per-rack power draw at sustained load, not nameplate

• Inlet temperature and airflow requirement

• A signed statement of the cooling method required

Then have your own facilities team confirm feasibility in writing before award. That single step costs a week and routinely saves a quarter.

The wider point. The order in which AI infrastructure decisions get made determines whether the money is well spent. Workload first, sizing second, constraints third, architecture fourth, procurement last. Teams that start at the accelerator and work backwards end up rebuilding the design at least once — usually after the budget is committed and the timeline is public.

If you are early in a build, the readiness checklist we publish free covers this chain and the rest of the pre-RFP gates. Run it with your facilities team in the room. The items you cannot answer are the ones that will cost you.