8 July 2026
IT load vs total facility power: the question that breaks colocation budgets
A 5.2 MW requirement can mean two different numbers with a seven-figure gap between them. Here is how to force the answer out before you quote.
Take a concrete deployment: 256 nodes, each a 10U air-cooled chassis carrying eight NVIDIA HGX B200 GPUs. That is 2,048 GPUs and roughly 5.2 MW. Now go and find somewhere to put it.
The density problem comes first
The megawatt number is the headline, but it is not the constraint. The constraint is what happens inside a single cabinet.
A large chunk of the existing colocation estate was designed around 5 to 10 kW per rack. A lot of the newer estate reaches 15 to 20 kW. An 8-GPU Blackwell node draws vastly more than that, and once you put more than one in a cabinet you are into densities that a conventional raised-floor hall simply cannot cool with air, no matter how much air it moves.
So the first filter is not "do you have 5 MW". It is "what can one cabinet actually sustain, and how are you removing the heat".
Three cooling architectures, three different conversations
- Conventional air. Works up to a point, and that point is lower than most operators claim. Ask for the sustained per-cabinet figure, not the peak, and ask what happens to it when the hall next door fills up.
- Rear-door heat exchangers. A genuine step up, and often the pragmatic answer for air-cooled 8-GPU chassis. But it puts water in the white space, which changes the operator's own risk posture and sometimes their insurance.
- Direct-to-chip liquid. Where the highest-density Blackwell builds are going. This is not a colocation product you can just buy — it needs coolant distribution units, a water-side loop, and a facility that has actually commissioned the plant rather than described it in a brochure.
What to ask before a facility goes on the shortlist
- Sustained usable kW per cabinet, in the specific hall being offered — not the facility average.
- Cooling architecture, with CDU capacity and water-side infrastructure evidenced if liquid is on the table.
- Is the power contracted, or is it an application with the utility? These are not the same thing and the difference can be eighteen months.
- Concurrent maintainability. Tier III or it does not go on the list — a training run that has been going for three weeks does not survive a maintenance window that drops a feed.
- Floor loading. Twenty-odd fully populated GPU cabinets is a serious structural load, and it is the question nobody remembers to ask until the raised floor is being surveyed.
- Delivery and staging. 256 nodes is a lot of pallets, and a loading dock, a lift and a staging area become part of the critical path.
The point
The GPU purchase order is the easy half of a deployment like this. The hard half is finding a facility that can honestly take the load, on the date, at a rate that leaves the deployment economically viable — and then getting 256 nodes through customs, into a hall and into burn-in without losing a fortnight.
If a broker cannot talk fluently about CDUs, ORV3 busbars and rear-door heat exchangers, they cannot qualify a facility for this. They can only forward your email to it.