AI Data Center Power, Cooling & Interconnect Limits 2026

Grid connection is a major constraint for UK AI projects

Computer Weekly's analysis puts the UK datacentre pipeline at 173 committed projects in planning, representing 14.6GW of planned capacity, against Computer Weekly's estimate of roughly 1.6GW of existing capacity, concentrated mainly in London and the M4 corridor. That mismatch can delay projects before cooling or interconnect decisions are finalised.

High-voltage connections for large facilities can take several years; industry reporting indicates some projects have reported connection dates in the early 2030s, though the actual energisation date may depend on further transmission, substation and distribution-network works rather than site availability alone. Ofgem's proposed connection commitment fees are intended to remove speculative or non-progressing projects from the queue, but the upfront cost risk could also price out viable developers or push up financing requirements.

The result: industry reporting has described the government's reported ambition of 6GW of AI-capable capacity by 2030 as currently out of reach without materially faster project delivery, which makes schedule risk a core investment variable rather than a planning footnote. For context on how this plays into site selection, our UK data-centre power analysis covers the regional detail.

Why air cooling can't follow GPU density upward

High-density modern GPU racks can exceed 50 to 100 kW per cabinet, while conventional air cooling is often designed around roughly 30 to 40 kW per rack as an indicative range. Push beyond that and GPU firmware can throttle performance when model-specific thermal limits are reached — exactly the outcome operators are paying to avoid.

Liquid cooling solves the heat-transfer problem but introduces a pumping-power cost of its own, which is why operators are shifting toward direct-to-chip cold plates mounted straight on the processor rather than over it. Two-phase immersion cooling uses dielectric fluid that boils at heated components and is then condensed and returned within the system, and can reduce cooling energy use versus traditional air cooling. Microsoft says its newer datacentres use closed-loop systems designed to recirculate water with little or no ongoing water consumption — a costly approach, but a clear signal of where density-driven design is heading.

For UK buyers, the audit point is retrofit cost: structural loading, ceiling height, plant-room capacity, pipe routing and disruption to live operations can all extend delivery timelines well beyond the headline rack spec. Run your target density through a cooling requirements check before assuming an existing hall can take the upgrade, and compare approaches in our AI server cooling guide.

Copper faces increasing limits at high bandwidth and reach — optical interconnects are an alternative for demanding links

Electrical copper interconnects lose signal strength past roughly two metres, with usable reach halving every time bandwidth requirements double. Larger GPU clusters mean longer communication distances, and tightly packed copper also generates electromagnetic interference that demands power-hungry equalisation just to hold a clean signal.

Optical interconnects, which convert data into light pulses, suffer minimal loss over distance and free up airflow in dense racks because the cabling is far thinner. Some vendors are moving further still, to co-packaged optics where photonic circuits sit directly on the GPU package. The catch is cost: optical setups can run up to 7x more than copper today, a premium buyers will have to absorb in the short term as copper faces increasing limits at high bandwidth and reach, rather than this being simply a design choice. See our GPU interconnects primer for how this affects cluster architecture decisions.

Electricity cost and water risk compound the capacity problem

Even where power is available, it may not be competitive. Pure Data Centers' CEO, citing UK government figures, said UK power costs were approximately four times US levels, equating to an estimated £226m annual difference for a 100MW facility — a gap that could steer hyperscaler investment elsewhere. The same executive flagged Ofgem's RIIO network price-control cycles as too slow to match AI infrastructure deployment timetables.

Water is the quieter risk. UK operators aren't generally required to report water use, yet one estimate puts English datacentre consumption at almost 1.9 billion litres a year, with demand peaking during heatwaves — precisely when AI workloads push cooling demand highest and public supply is under most strain. Neither cost nor water constraints show up in a headline grid offer, which is exactly why they need separate due diligence.

North vs south: what moving the site actually trades off

Development is shifting toward Scotland and north-east England as southern markets, including West London, run into grid pressure despite their fibre, cloud and talent density. Moving north can improve land and power availability, but it introduces costs that need pricing rather than assuming away: longer-haul latency to existing cloud regions, thinner carrier density, workforce access and construction supply chains all affect total delivered cost per rack.

Before locking a location, test diverse fibre routes, carrier availability, cloud on-ramps, cross-connect lead times, meet-me-room capacity and the ability to add route diversity without civil works. Pair that with a power density planner exercise so the rack-level numbers and the site-level grid offer actually line up.

The capex checklist before you sign

None of these constraints are individually fatal, but stacked together — grid queue, thermal ceiling, interconnect cost and electricity price — they turn a straightforward GPU deployment into a multi-year infrastructure project. Treat every grid offer, cooling spec and connectivity claim as something to be verified, not assumed.

• •Firm energisation date, not an indicative connection offer
• •Detailed grid-upgrade liabilities and who pays for them
• •Staged power availability versus full nameplate capacity
• •Cooling capacity proven at the intended rack density, not building average
• •Water and discharge constraints, especially for liquid or immersion cooling
• •Fibre-route diversity and expansion rights beyond day-one capacity
• •A tested operating plan for grid, cooling or connectivity failure