Buying the GPUs is the easy part. The harder question — the one that stalls more AI projects than budget does — is where you physically put the rack. A current-generation NVIDIA Blackwell AI rack draws well beyond what a normal server room was ever built to power or cool, and the next generation pushes further still. This guide is the facilities-side decision: what an AI rack actually demands, why your existing comms room probably can't host one, and how to choose between retrofitting on-premises and going to colocation.
The numbers that break a normal server room
A single NVIDIA GB200 NVL72 rack draws on the order of 120-132 kW, with each Blackwell GPU dissipating up to ~1,000 W — and next-generation designs are projected toward ~240 kW per rack. For context, a conventional enterprise rack is provisioned for roughly 5-15 kW. You are not adding a server; you are adding the power and heat load of a small building to one floor tile. Most server rooms simply do not have the electrical supply, the distribution, or the cooling headroom — and bolting it on is rarely trivial.
Cooling is the harder half. Air cooling tops out long before these densities, so direct-to-chip (DTC) liquid cooling is now the default for current AI infrastructure — it captures roughly 70-80% of the rack's heat into liquid, with the remaining ~20-30% (memory, NICs, PSUs) still needing air. That means a facility that can deliver coolant to the rack and reject that heat, not just a bigger CRAC unit.
Can you retrofit on-premises?
Sometimes — but go in with eyes open. Hosting even one AI rack on-prem typically means a serious electrical upgrade (supply, PDUs, possibly switchgear), a liquid-cooling loop (CDU, manifolds, leak detection), structural and floor-loading checks, and the fire and safety sign-off that comes with it. For a single rack that can be disproportionately expensive and slow. It can be the right call when data must stay on your premises for sovereignty or latency reasons, or when you're building out enough racks to amortise the facilities work.
When colocation wins
For most organisations deploying one or a handful of AI racks, an AI-ready colocation facility is the faster, cheaper route to live: the power, liquid cooling and density are already there, and you're renting capability rather than rebuilding your building. The trade-offs are data residency (choose a UK facility if that matters), latency to your other systems, and ongoing rack and power cost versus a capital retrofit. The decision usually comes down to scale and data-location constraints: one or two racks with no hard on-prem requirement points to colo; a sustained build-out, or a sovereignty/latency constraint, can justify the on-prem investment.
How to plan it
Start from the facility, not the GPUs. Confirm the per-rack power and cooling your chosen platform needs, model your IT load into kW and heat (our power and cooling calculators help), then test it against what your room can realistically deliver — and price the retrofit against colo for the number of racks you actually plan to run.
Servnet specs AI servers and the supporting power, cooling and networking, and will help you weigh an on-prem retrofit against UK colocation honestly — including whether refurbished previous-generation GPUs get you there for less.