Product1 publisher3 min readPublished
Halls wired for 10kW racks need eight times the power before AI clusters land
An opinion piece on DataCenterDynamics puts UK capacity at 2.9GW today against at least 6GW of AI-capable capacity needed by 2030, and makes the case for retrofits. Only 29 percent of operators lean that way.
The Product Desk · Product desk
What happened
- An opinion piece on DataCenterDynamics argues operators should free stranded power, cooling, floor-loading and network capacity inside sites they already run, by upgrading distribution, containment and fiber pathways.
- The piece reports that 29 percent of operators are leaning towards retrofitting to expand AI infrastructure, and that phased upgrades can add capacity in months instead of years.
- Global data center construction inflation averaged around 5.5 percent in 2025 for traditional facilities, while building a new site can take several years.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Grid constraints, planning timelines and supply chain delays sit outside an operator's control, so the one variable it can still move is how much load its existing buildings will carry.
- exposure The work lands in occupied buildings, so a tenant under an SLA absorbs the risk of an aisle-by-aisle upgrade it did not order and cannot schedule.
- cost Compressing commissioning leaves hidden faults, and they surface once the higher-density load is already running.
An operator with empty tiles in an older hall gets asked whether it can take a 30kW rack. The answer sits in the electrical path behind those tiles. Power is often the hard ceiling, according to the DataCenterDynamics piece, and even where extra grid capacity is available, internal switchgear, backup systems and rack-level distribution may need significant rework [13]. A site can have floor space and still lack the power, cooling or structural capacity for AI workloads [12]. Older buildings also hit floor-loading limits, restricted ceiling heights and constrained plant space [13].
Measure the step change per rack. Halls designed for air cooling and classic cloud workloads of 5-10kW per rack are being asked to carry 30-80kW or more [11]. At the top end that is eight times the power into the same footprint, and three times even comparing a 30kW AI rack against a 10kW cloud rack [19]. Heat concentrates with it. Containment, rear-door heat exchangers, modular cooling and liquid cooling each have to be integrated without creating new operational risks, and the design has to leave room for maintenance access, pipework, leak detection and future expansion [14].
The case holds up best on schedule. Global capacity is expected to nearly double by 2030 [1], a new site can take several years to build [7], and phased upgrades can add capacity in months while the facility stays live [10]. Construction is getting dearer meanwhile: global data center construction inflation averaged around 5.5 percent in 2025 for traditional facilities [8]. Hold that rate for three years and the same build costs about 17 percent more [22].
Adoption lags the argument. The piece reports 29 percent of operators leaning towards retrofitting to expand AI infrastructure [9]. That leaves 71 percent doing something else [21], and the article does not say what, or name the survey the figure came from [24].
The UK figures need checking before anyone quotes them upward. Capacity in early 2026 is put at 2.9GW, with at least 6GW of AI-capable capacity required by 2030, described as a threefold increase on today [5][6]. Six divided by 2.9 is 2.07 [20]. For 6GW to be three times today's number, the base would have to be roughly 2GW of AI-capable capacity, not the 2.9GW total [20].
Before any purchase order comes a three-way sort: which constraints are fixed, which can be engineered around, and which make the site unsuitable for AI workloads [18]. Run structure, power, cooling and cabling through those three buckets, hall by hall, and each site ends up with either a sequenced plan or a decision not to try. Retrofits also need power, cooling, structural and network dependencies mapped against live operations first [4]. Without one plan shared by facilities teams, network specialists, contractors and suppliers, operators accept what is available instead of what is required, or make piecemeal upgrades that create future constraints [17].
What to watch
- Publication of the 29 percent retrofit figure with its survey, sample size and region, which would show how firm the operator preference is.
- A UK breakdown separating AI-capable gigawatts from the 2.9GW total, which would settle whether 6GW is a doubling or a tripling.
- Reports from operators running live-hall retrofits on commissioning faults or SLA incidents once the high-density zones come up.