AI chips and models are scaling faster than the energy systems that underwrite them. In round figures annual AI growth is around 50%, data centre capacity grows at ~15%, while the national grid adds perhaps 3.5% at best. The disconnect is real and its here now. The UK – like most countries – has spent the past couple of decades building cloud data centres designed for SaaS, web, and enterprise IT. These facilities were never built for the heat, density, loading, growth factors or reliability requirements of modern AI. Yet both government and industry now talk about “6 GW of AI capacity by 2030” as if it’s simply an extension of what we already have.
It isn’t. And that’s where we can get beyond the headlines.
The UK sits on ~1.2 GW of operational data centre capacity today, but only ~100 MW of that is genuinely AI-ready. The rest is legacy cloud. Air-cooled. Low-density. Architected for a completely different era. Retrofitting that footprint for AI is like trying to run an industrial factory out of a Victorian townhouse. You quickly hit physical limits. Other nations are already reorganising their scientific infrastructure around AI-driven discovery. The UK cannot afford to treat this as a conventional data centre expansion problem.
Two 2025 government reports – AI and nuclear
Two major government reports – the AI Opportunities Action Plan January 2025 and the Nuclear Regulatory Review 2025 published earlier this week – both say the same thing in different language: if the UK wants national AI capability, the old model of “build another warehouse near Slough and plug it into the grid” is dead. The grid can’t support it. The planning system can’t support it. And the energy system wasn’t designed for this kind of load.
The UK’s own grid operator – NESO – makes the point bluntly. The national connections queue is now 700 – 800 GW, several times system need. Transmission constraints in Scotland alone have cost the public £3 – 4 billion in curtailment payments in the past few years. Even when we have surplus power, we can’t move it to where people want to build data centres. That’s why data centres in West London are looking at connection dates in the late 2030s.
Nuclear pivot
This is where the newly published Nuclear Regulatory Review becomes more than a policy footnote. The report makes two clear moves that matter for AI:
First, it reframes nuclear not as a monolithic megaproject industry but as a deploy-anywhere technology platform. Small modular reactors (SMRs) can be embedded directly into industrial clusters. They can anchor new digital zones. They can generate heat, power, and resilience for high-density compute campuses. Nuclear goes from “national megaproject” to local enabler of AI ecosystems.
Second, it explicitly positions nuclear as a strategic supply-side solution for digital infrastructure. That’s a quiet but significant pivot. Until now, energy policy and digital policy lived in different universes. This is the first time we see formal alignment around the idea that AI growth requires colocated, dedicated, sovereign power, not just more grid connections.
From a warehouse in Slough to…
This opens the door to a fundamentally different blueprint for the next generation of UK data centres. Today’s AI clusters already demand 100-200 kW per rack, and with 800 kW on the horizon. They need liquid cooling, direct-to-chip systems, 800V DC architectures coming in 2026, dense InfiniBand fabrics, and load factors that look more like industrial smelters than cloud servers. As we move deeper into Blackwell and soon the Rubin cycle, traditional data centres start failing physics. Eventually you can’t push any more heat or electrons through the format we’ve relied on since the 1990s.
Which is why the next wave won’t look like data centres at all.
They’ll look like integrated power-compute cells. Much closer to a Tesla Megapack site than a hyperscale campus. Self-contained modules that combine power, cooling, GPU density, storage, and networking into a single engineered unit. Standalone or clusterable. Near-load or near-generation. Designed around compute first, not real estate. The grid becomes optional, not assumed.
AI Growth Zones and sovereign compute strategy
This shift also breaks one of the UK’s most entrenched structural issues: the concentration of roughly 90% of national datacentre capacity in Slough, West London, and the M4 corridor. Once you remove the hyperscaler requirement to co-locate inside availability zones – and replace it with integrated compute-power systems – the whole map opens up. You can put 200 MW of AI compute in Wales, the North East, the Scottish Central Belt, Teesside, or anywhere with fibre, space, cooling potential, community alignment, and a path to either an SMR or firmed renewables. Even with slow planning cycles, the locations become viable because the architecture is modular, self-contained, and energy-anchored.
The UK’s energy economics make this even sharper. High electricity prices are beyond an inconvenience; they are the primary drag coefficient on national AI capacity. The government’s AI Growth Zones policy acknowledges this by introducing targeted power-price support from 2027, but only in specific regions like Scotland, Cumbria, and the North East. These areas already suffer from heavy grid constraints, so the policy effectively recycles avoided curtailment costs into reduced grid and system charges for qualifying AI facilities. It is not a national removal of energy taxes or distribution fees. It is a location-specific signal designed to redirect large AI clusters toward places where constrained renewable output is regularly wasted.
Useful, yes – but not transformative. And it does not change the underlying physics: AI infrastructure requires firm, low-volatility power, delivered through integrated modules that unify compute, cooling, and local generation. Until the UK aligns price signals, permitting, and physical capacity, AI Growth Zones will help at the margins but will not close the strategic gap.
This is where the two UK Government reports finally converge. The future is clusters, not campuses. Dedicated power. Fast-track planning. Hybrid nuclear-renewable-storage districts. Heat reuse into local communities and greenhouses. Export-controlled sovereign compute for defence and science. And an operator ecosystem that includes the big three hyperscalers, but also the new specialist players – CoreWeave, Nscale, Nebius, and others – who are redefining what “a data centre” even is.
What a 6 GW AI system actually requires by 2030
The UK has a choice. Keep betting on a legacy data centre model with decade-long grid queues and planning battles. Or pivot to a new architecture where compute and generation are designed together and deployed at industrial scale.
If the country wants to get anywhere near 6 GW of AI capacity by 2030, it can’t rely on the grid alone. It needs distributed, engineered compute power. It needs SMRs as anchor assets. It needs AI Growth Zones with dedicated substation capacity and enforced readiness rules. And it needs to accept that the data centre of the future is not a warehouse. It’s a power plant that happens to run intelligence.
The UK’s challenge is not just domestic. Other countries are beginning to treat AI infrastructure as a state-level capability, not a commercial utility. Some are centralising datasets, supercomputers, and research workflows into unified discovery engines. Whoever aligns compute, data, and energy into a single architecture will set the pace of scientific progress for the next century.
The UK can lead this shift if it moves now. If not, the next generation of AI breakthroughs will run somewhere else.
