Power and cooling for racks that draw ten times what the hall was built for
The electrical and thermal engineering under AI compute: rack density, distribution, liquid cooling, grid connection and the energy economics.
Second constraint
We work from the grid connection inwards: what the site can import, how the distribution carries it and how the heat leaves the building. The output states the constraints — kW per rack, redundancy tier, and where the hall stops scaling.
Electrical
Legacy halls were built around 5–10 kW racks. AI racks run 40–80 kW and dense liquid designs reach 130 kW, which changes distribution, containment and floor loading entirely.
Density
40–130 kW
Per-rack draw for current AI systems, against 5–10 kW in legacy halls
Overhead
1.1–1.5
PUE range across modern designs; most of the overhead is heat rejection
Thermal
Air is practical to roughly 30 kW a rack. Rear-door exchangers take an existing floor to 40–50 kW; direct-to-chip cold plates carry 70–80% of the heat into liquid.
Supply
In constrained markets the connection queue, not construction, is the longest item on the plan. We start the capacity study early and phase energisation so the first racks earn.
Power
Cooling
Generation and storage
Monitoring
Standards
Who we do this for
2–4 weeks
What a site can support: available capacity, rack density ceiling, cooling options, connection lead time and an energy cost model.
Feasibility report with density ceiling, cooling options, energy cost model
12–40 weeks
Electrical and mechanical design, contractor selection and oversight, grid connection management, commissioning and load testing.
Single-line diagram, thermal design, commissioning results, as-built drawings
Ongoing
Monitoring, PUE and WUE reporting, tariff and demand charge optimisation, and capacity planning against the compute roadmap.
PUE and WUE reporting, tariff review, capacity plan
Sometimes, at reduced density. A hall built for 6–10 kW racks can often reach 15–20 kW with containment work, and 40–50 kW with rear-door heat exchangers if the chilled water plant and floor loading allow. Beyond that you are rebuilding, and a purpose-built or colocated site is usually cheaper and always faster.
If you need fewer than about four dense racks, or your compute plan will not be stable for two years, do not build. Colocation with a provider that already supports liquid cooling gets you running in a fraction of the time with none of the connection queue risk, and you can build later with real load data.
Direct-to-chip for almost everyone. It is supported by mainstream OEM platforms, retrofits into a conventional hall, and removes 70–80% of the heat to liquid. Immersion earns its place where density is extreme or the environment is dusty, but it narrows your hardware options and changes how the kit is serviced.
Send us the constraint you are actually up against — budget, latency, regulator, deadline.
