New AI facilities are being designed for rack densities that would have been unthinkable five years ago, and the electrical infrastructure behind them hasn't caught up. For UK project teams, that gap now decides whether a build reaches commercial operation on time or stands complete and unpowered, stalled until the grid connection arrives.
View the data behind this chart
| Legacy racks… | New AI-ready… | GPU-dense… | |
|---|---|---|---|
| Power draw per rack | kW8 | kW50 | kW250 |
The density shock UK build plans weren't sized for
Rack power draw has multiplied dramatically in a short space of time. Where a typical rack pulled just 5kW to 8kW five years ago, new AI facilities are now built to handle 15kW to 50kW per rack, and GPU-dense configurations push that as high as 100kW to 250kW, according to the source brief from Turtle's Kevin Cruts writing for DataCenterDynamics. Newly built facilities are trending toward an average of 45kW per rack, and roughly 80 percent are incorporating liquid cooling from the outset.
For UK project teams still working from specifications drawn up even two or three years ago, that's not an incremental change — it's a different building. Anyone scoping a new facility or a colocation upgrade needs to understand rack power density before committing to a switchgear order, because the answer determines everything downstream: transformer sizing, distribution paths, and whether AI server cooling solutions need to be liquid-based from day one rather than retrofitted later.
Why equipment lead times, not labour, now set the schedule
The specifics, per the brief, are stark: distribution-class switchgear and transformers for AI-density builds carry lead times of 52 to 78 weeks, and large power transformers can take 128 weeks or more depending on manufacturer and configuration. High-voltage utility connections, on-site substations, and backup distribution routes all need design sign-off and ordering decisions locked in long before civil and structural works are anywhere near finished.
This flips the usual project sequence on its head. It's the delivery of equipment, rather than crew availability or site progress, that now dictates the build timeline. UK teams that leave electrical procurement until after design sign-off routinely land in the same place: the structure finishes on time and then stands empty for well over a year while switchgear is still in transit. Modelling power and cooling needs with power and cooling tools before the design is locked isn't a nice-to-have anymore — it's what actually keeps a programme moving.
The AC-to-DC shift UK teams can't plan around
The bulk of data centre builds are still wired around alternating current. Yet for facilities built around GPU-dense AI clusters, a move toward 800-volt direct current distribution is picking up pace and, per the brief, is fast becoming the default architecture for that category of project. Supporting DC means sourcing entirely new equipment: switchgear built for DC, DC-rated circuit breakers, solid-state transformers, and cabling engineered for the current-carrying demands of dense AI clusters. Busbars are also taking over from conventional cable in these GPU-heavy deployments, since they can handle current loads at densities that cable simply can't.
Most electrical contracting expertise sits firmly in AC territory, and firms capable of designing, specifying and commissioning DC systems remain, in the brief's own phrase, "in short supply." Spotting that skills shortfall early isn't about getting ahead of rivals — it's a basic scheduling necessity, and it matters directly for anyone specifying against hosting Blackwell AI rack power and cooling requirements in a UK colo footprint.
Two electrical architectures on one campus
It's now common for one campus to need both systems running side by side — AC for conventional enterprise and cloud workloads, high-density DC for AI clusters — with each having its own protection scheme, monitoring setup and power route. That dual requirement brings a level of coordination that many facilities teams simply haven't had to handle before.
For UK operators running mixed estates — older enterprise racks sitting alongside new AI clusters — the connection points, handovers and protection coordination between the two systems have to be planned as one from the start, rather than stitched together after each side is specified separately. Contractors taking on this kind of project for the first time discover that it isn't simply a matter of scale; the whole scope of work is fundamentally different from anything in their AC-only history.
What UK project teams should change now
Three practical changes, drawn directly from the brief, translate directly into UK procurement cycles.
Buyers should be running early scenario planning through an AI GPU calculator and specifying against configurators such as the Dell server configurator, HPE server configurator or Lenovo server configurator before final electrical drawings are locked, so power and cooling assumptions match the compute that's actually going to be deployed.
- •Kick off electrical procurement before the design is finalised — switchgear and transformer specifications can be worked up alongside ongoing design development
- •Specify to match the architecture the facility will genuinely operate on (DC-ready kit, high-ampacity busway, power modules built for liquid cooling), rather than defaulting to what was standard on the last project
- •Treat AC and DC as one combined system from the design stage onward, with protection design, monitoring integration and commissioning schedules agreed together by the engineer of record, electrical contractor and equipment suppliers
View the data behind this chart
| Phase | Starts (week) | Duration (weeks) |
|---|---|---|
| Design lock & spec… | 0 | 12 |
| Distribution-class… | 12 | 78 |
| Large power transformer… | 12 | 128 |
The capex case for moving early
The numbers underline just how significant this shift is: the brief puts the data centre switchgear market alone on track to hit $13.6 billion by 2031, expanding at 16 percent a year. That figure reflects genuine build activity rather than forecast optimism — which means UK buyers chasing the same distribution-class kit and DC-capable contractors should expect availability to get tighter, not easier.
For capital planning, that argues for locking equipment specifications and contractor capability early rather than waiting for design certainty. Teams weighing build-versus-colocate decisions, or benchmarking spend against the wider market, should factor procurement timing into any IT procurement services engagement and cross-check assumptions against the AI server cost index before finalising budgets.