The Guardian reports that a datacentre in Loughton, Essex, presented as the UK’s largest AI supercomputer, will miss its planned launch next year because of power supply problems, and could be held back into the mid-2030s.
Key takeaways
- A datacentre project in Loughton, Essex, that the UK government publicly welcomed will not begin operating next year as intended, according to the Guardian.
- The reason given is power supply: the site cannot obtain the electricity it needs on the timetable the project assumed.
- The Guardian reports the delay could stretch into the mid-2030s, which would be a slip of most of a decade rather than a few months.
- Large AI computing sites depend on grid infrastructure that takes years to plan, consent and build, so a computing plan can be sound while the power plan is not.
- The specific cause of the shortfall, the parties responsible and the revised timetable have not been established in the material available here.
A flagship AI datacentre has slipped years behind schedule
The immediate facts are narrow. The Guardian reports that a large datacentre project at Loughton in Essex, announced in 2025 and described at the time as the country’s largest AI supercomputer, will not start operating next year as planned. Power supply problems mean the site faces a long wait before it can come online, and the delay could run into the mid-2030s.
Several things follow from that description, and several do not. A datacentre of this kind is a building full of specialised computing hardware — dense racks of accelerators used to train and run AI models — together with the electrical and cooling systems that keep them working. The building can be finished and the hardware ordered while the electricity to energise it remains unavailable. That is the distinction at the heart of this story: the constraint is reported to be power, not construction alone, and not the computing equipment.
What is not known from the reporting summarised here is how much power the site requires, which connection or generation arrangement fell through, who bears the cost of the delay, and whether a firm replacement date exists. Those details would change how serious the setback looks.
Why this is drawing attention now
The delay matters because of the gap between the announcement and the outcome. A project publicly held up by government as national infrastructure has moved from an imminent start to a possible wait of years. That contrast is what makes it newsworthy: the change is in the timetable, not in any dramatic single event.
It also lands in the middle of a wider argument about whether countries can physically build the AI capacity they have promised. Governments across Europe, North America and Asia have announced compute programmes, national AI facilities and industrial zones intended to attract datacentre investment. Announcements are quick. Substations, transmission lines, transformers and generating capacity are slow. When a headline project stalls on electricity rather than on money or chips, it is read as evidence about the whole category, not just one site.
There is a second reason for attention. AI datacentres have become a live local and political issue in many places, because they consume large amounts of electricity and land while employing relatively few people on site. A delay attributed to power supply gives both supporters and critics of these projects material for their case.
The background a newcomer needs
Modern AI computing is unusually power-hungry per square metre. Racks used for AI training draw far more electricity than traditional web-hosting equipment, and much of that electricity becomes heat that must be removed. A large AI site therefore needs a firm, high-capacity electricity connection, resilient supply arrangements, and cooling infrastructure sized to match.
Getting that connection is a process, not a purchase. A developer applies to connect to the local distribution network or, for very large loads, to the high-voltage transmission network. The network operator assesses whether existing cables, transformers and substations can carry the load, and what reinforcement would be required. Where reinforcement is needed, the work has to be designed, consented, funded and built, and the equipment involved has long manufacturing lead times. In many networks there is also a queue of projects — generation and demand alike — waiting for capacity, and a project’s place in that queue shapes its earliest possible energisation date.
Alternatives exist. A site can build its own generation, arrange a private wire from a nearby plant, or phase its capacity so a smaller first tranche opens early. Each option carries its own consents, costs and delays. None is instant.
Who is affected, and how
The developer and its investors carry the most direct exposure. Capital is committed, land and buildings may be part-built, and hardware bought for a 2027 opening would be several generations old by the mid-2030s. Long delays usually force a redesign rather than a simple pause.
Prospective customers are next. Organisations that expected to rent domestic AI compute — companies, universities, public bodies — must either wait, buy capacity abroad, or use existing commercial cloud providers. For work involving sensitive data or legal requirements about where data is processed, that substitution is not always straightforward.
Government is affected reputationally and practically. A project it publicised as evidence of national AI capability is now an example of delivery risk, and the delay feeds a broader question about whether grid capacity is keeping pace with industrial ambitions.
Locally, effects cut both ways. Construction work and rates income are postponed, as is any local employment attached to operations. Equally, residents and businesses concerned about electricity and land being allocated to a datacentre get a longer reprieve. The balance of local opinion in this case is not established in the reporting available here.
Where informed people disagree
There is genuine disagreement about priorities. Some argue that AI compute is strategic infrastructure and should be fast-tracked for grid connection, because the economic value per unit of electricity is high and the alternative is investment going overseas. Others argue that scarce network capacity should favour housing, electrified transport, heat pumps and manufacturing, and that datacentres should wait or fund their own generation.
Specialists also disagree about diagnosis. One view treats delays like this as a symptom of slow planning and connection regimes that need reform. Another treats them as a predictable consequence of demand forecasts that ran ahead of any physical plan to supply them, in which case the appropriate response is more conservative announcements rather than faster approvals.
A third disagreement concerns the underlying demand. Some analysts expect AI compute needs to keep rising steeply, making every delayed facility a lost opportunity. Others expect efficiency gains in chips and models, and possible consolidation among AI companies, to soften demand — in which case a site opening in the mid-2030s may face a market very different from the one it was designed for.
The practical implications
For anyone tracking AI infrastructure, the useful lesson is to read announcements in terms of power rather than headline computing claims. The questions that predict delivery are unglamorous: does the project hold a firm connection agreement, for what capacity, at what date, and does energising it depend on network reinforcement that has not yet been built?
For developers, the episode reinforces a shift already visible in the sector towards siting decisions driven by available electricity — near existing substations, retired power stations, or new generation — rather than by proximity to customers. It also strengthens the case for phased builds, so that a first block can open while later capacity waits.
For policymakers, the implication is that compute strategy and energy strategy cannot be written separately. Commitments to host large AI facilities imply commitments to deliver grid capacity within the same timeframe, and the second is harder than the first.
For readers, the practical takeaway is caution about dates. Announced opening years for large energy-intensive facilities are forecasts, and the electricity connection is usually the item most likely to move them.
What to watch next
The first thing to watch is whether a revised timetable is published, and whether it is expressed as a firm energisation date or a range. A specific date backed by a connection agreement means something; a range ending in the mid-2030s does not.
Second, watch for the cause being spelled out. Power supply problems can mean a queue position, missing local network capacity, a failed private generation arrangement, or equipment lead times, and each implies a different fix.
Third, watch for a redesign. Projects in this position often re-emerge smaller, phased, or on a different site with better connection prospects. Any of those would signal that the original specification has been abandoned.
Fourth, watch policy. Changes to how large electricity users are prioritised in connection queues, or to planning rules for datacentres and their supporting infrastructure, would affect this project and others like it.
Finally, watch whether other announced AI facilities report similar problems. A single delay is a project story; a pattern is an infrastructure story.
Frequently asked questions
Why is the Loughton AI datacentre delayed?
The Guardian reports that power supply problems are the cause. The site, announced in 2025 and described then as the UK’s largest AI supercomputer, was due to begin operating next year, but cannot obtain the electricity it needs on that timetable and may not come online until the mid-2030s. The precise nature of the power problem is not established in the reporting summarised here.
When will the site actually open?
No confirmed date is available. The reporting indicates the original plan to start operating next year will be missed and that the delay could extend into the mid-2030s, which is a range rather than a commitment. Timetables for large energy-intensive facilities usually firm up only once an electricity connection agreement and any required network reinforcement are in place.
Why do AI datacentres need so much electricity?
AI training and inference run on dense arrays of accelerators that draw far more power per rack than conventional servers, and nearly all of that power ends up as heat that cooling systems must remove. A large AI facility therefore behaves like an industrial electricity consumer, requiring a high-capacity, reliable connection rather than the modest supply an ordinary office building needs.
What is a grid connection queue?
When a project wants to connect to an electricity network, the network operator assesses whether local cables, transformers and substations can carry the load. Where upgrades are needed, the project waits while that work is designed, approved and built, alongside other applicants. A project’s position in that sequence determines the earliest date it can be energised, sometimes years ahead.
Can a datacentre generate its own power instead?
It can, in principle. Options include on-site generation, a private wire to a nearby power plant, or phasing capacity so a smaller first stage opens sooner. In practice each route needs its own consents, funding and construction time, and fuel or grid backup arrangements. Self-supply changes the shape of the delay rather than removing it. Whether any such option applies here is not known.
Does this affect UK AI capability?
It removes an expected source of domestic AI computing capacity from the near-term picture. Organisations that planned to use it must wait, rent capacity from commercial providers, or process work abroad, which can be a complication where data location matters. How significant the loss is depends on the site’s intended capacity and customers, details not established in the material available here.
Sources and further reading
- The Guardian, technology reporting on the delayed Loughton datacentre project, which is the basis for the factual claims here.
- Great Britain’s electricity system and network operators, whose published material explains how demand connection applications, queues and network reinforcement work.
- UK government publications on artificial intelligence and compute strategy, for the policy context in which such projects are announced.
- Industry and academic analyses of datacentre electricity demand and cooling, for background on why AI facilities are power-intensive.
Surfaced from the rss:guardian_tech signal “delayed AI datacentre project”. AI-assisted draft, editorially reviewed.
