Britain's AI infrastructure debate often begins with chips, cloud platforms and planning reform. The government's own AI Growth Zones policy points to a more basic constraint: electricity. Large compute clusters need grid connections measured in hundreds of megawatts, and the process of securing that capacity can take longer than the buildings themselves.
The programme is therefore a useful test of whether the UK can turn AI industrial policy into physical infrastructure. It is easy to designate a growth zone. It is harder to prove that enough power can arrive on a commercially useful timetable.
The 500MW test changes the meaning of a viable site
Guidance for AI Growth Zone applicants says sites should demonstrate access to at least 500MW of power capacity by 2030, either through a secured grid connection or a credible behind-the-meter solution. That threshold is large enough to make power availability a site-selection issue at national scale.
Government has also acknowledged that the connections queue is oversubscribed. Its delivery plan proposes removing speculative demand, reserving capacity for strategic projects and allowing viable developers to consider building high-voltage infrastructure themselves. Those measures are an admission that conventional connection timelines do not fit the pace policymakers want from AI investment.
Regionalisation works only if power and fibre arrive together
Moving capacity away from the South East can ease land and grid pressure, but a regional data-centre strategy cannot be reduced to cheaper property. AI infrastructure also needs resilient fibre, multiple power routes, specialist contractors and customers willing to use the capacity.
The strongest regional locations will therefore be places where surplus or constrained generation can be matched with digital connectivity. The government's plan explicitly notes that some northern and Scottish sites could reduce wider system costs by locating demand closer to generation that would otherwise face transmission constraints.
The commercial risk is counting announced megawatts as delivered capacity
Data-centre announcements are often reported as if every proposed megawatt already exists. They do not. A project moves through land control, connection agreements, planning, financing, construction, energisation and customer ramp-up. Failure at any stage can delay the capacity by years.
For investors, local authorities and suppliers, the useful question is not how many AI campuses have been announced. It is how many have secured power, started construction and established a credible commissioning schedule.
What to watch next
The next evidence should come from named connection agreements, planning decisions, substations and construction timetables rather than another round of headline investment totals. Britain has a plausible case for regional AI infrastructure, but the grid will decide how much of that case becomes real.