The Patient Capital: How AI Demand Could Finance Britain's Nuclear Future
US tech giants are signing 20-year nuclear deals to power AI. Britain has the reactors and the demand but loses the financing, for the same reason it loses Bitcoin miners: the developed world's most expensive electricity.
Part 2 of 2: The Opportunity
Britain’s electricity is too expensive for the industry that wants it most. On 9 April 2026, OpenAI paused its flagship Stargate UK data centre, citing “regulation and the cost of energy.” Across the Atlantic, the world’s largest technology companies are doing the opposite, signing multi-decade deals to finance new nuclear power for their AI data centres. It is the same root cause this series identified in Part 1, seen from the other side. Britain has the developed world’s most expensive industrial electricity, around four times US levels. In Part 1, that price drove the Bitcoin miners abroad. Here, it is driving away the AI hyperscalers and the nuclear-financing wave they bring with them.
The American deals are already signed. Microsoft committed to a 20-year power purchase agreement to restart Three Mile Island’s Unit 1 reactor, securing 835 MW of carbon-free baseload for its AI data centres. Amazon bought Talen Energy’s 960 MW Cumulus campus in Pennsylvania for around $650 million, powered directly by the adjacent Susquehanna nuclear station. Google signed a development agreement with Kairos Power for up to 500 MW of small modular reactor capacity, and Meta opened a request for up to 4 GW of new nuclear. Hyperscale AI demand has become the financing event of the decade for new nuclear baseload.
Britain has the ingredients to join it. Rolls-Royce SMR is the chosen UK technology, with Wylfa on Anglesey confirmed in November 2025 as the first site. The government’s AI Growth Zones promise planning reform and faster grid connections. The demand is real and largely unmet: British data-centre developers have lodged more than 70 GW of grid-connection requests against a central forecast a fraction of that size.
Yet a grid constraint blocks Britain from capitalising. Reactors at Wylfa cannot easily move a gigawatt of power to the data centres clustered around London, 250 miles southeast. The American answer is co-location: placing data centres behind the meter at the nuclear plant. Britain’s regulatory framework has no clear pathway for it. This is Part 2 of our energy-independence series: how AI demand could finance Britain’s nuclear future, if Britain fixed the plumbing.
Patient Capital Meets Nuclear Baseload
The hyperscale technology companies reshaping American energy procurement share one characteristic: they need vast quantities of reliable, carbon-free electricity, and they will sign 20-year contracts to get it. That contract length is the whole point. It is the patient capital that makes new nuclear bankable.
Microsoft and Three Mile Island
Constellation Energy’s September 2024 announcement was the first of its kind. The company will invest around $1.6 billion to restart Three Mile Island Unit 1 (the reactor unaffected by the 1979 partial meltdown of Unit 2), rebranding it the Crane Clean Energy Center. Microsoft’s 20-year power purchase agreement is, by Constellation’s own account, the largest single contract the company has ever signed.
The economics hinge on commitment. Microsoft contracts the plant’s entire 835 MW output for two decades, which gives Constellation the revenue certainty to finance the restart: replacement steam generators, turbine upgrades, extensive regulatory approvals. The plant could return to service by 2028, adding 835 MW of carbon-free baseload to Pennsylvania’s grid while Microsoft matches its data-centre consumption with nuclear generation.
Amazon Goes Behind the Meter
Amazon Web Services bought the Cumulus data centre campus from Talen Energy in March 2024 for around $650 million (reported as roughly $350 million on close plus milestone payments). The 960 MW campus sits directly next to Talen’s Susquehanna nuclear station in Luzerne County, Pennsylvania, and takes power through a behind-the-meter arrangement: electricity moves from generator to load without touching the public grid.
That structure sidesteps transmission charges and connection delays entirely. It also drew a regulatory rebuke. In November 2024, the Federal Energy Regulatory Commission rejected, 2 to 1, an amended interconnection agreement that would have raised the co-located load from 300 MW to 480 MW, on the grounds that pulling firm generation off-grid for a single customer could shift costs and reliability risk onto everyone else. The behind-the-meter model works technically. Even in America, it has run into the same regulatory question Britain will have to answer.
Google Bets on Generation IV
Google and Kairos Power signed a Master Plant Development Agreement in October 2024, a framework to deploy up to 500 MW of small modular reactor capacity by 2035, with the first plant targeted for 2030. Kairos uses a fluoride-salt-cooled, high-temperature reactor, operating far hotter than conventional water-cooled designs. The Tennessee Valley Authority later joined the partnership, the first time a US utility has entered a power purchase agreement for electricity from a Generation IV reactor.
The Pattern, and Its Caveats
These are not four identical deals, and the honest reading matters. Only Microsoft’s is a signed, conventional 20-year fixed-price PPA for an operating reactor. Google’s is a development framework, not a single locked contract. Meta’s is a request for proposals for 1 to 4 GW, not yet a contract at all. Amazon’s headline event is a regulatory rejection, not a clean win. What unites them is direction of travel: the largest, most credit-worthy buyers on earth are willing to underwrite firm nuclear for decades.
The reason is the load profile. AI training facilities are multi-billion-pound investments designed to run flat-out for 20 to 30 years. Nuclear plants have high upfront capital costs and very low running costs once built. Conventional electricity markets, priced on short-term flexibility, struggle to finance new nuclear because future revenue is uncertain. A 20-year contract at a fixed price removes that uncertainty and makes the project bankable.
The demand behind these contracts is large and concentrated in America. The IEA’s 2025 Energy and AI report estimates data centres consumed about 415 TWh in 2024 (roughly 1.5% of global electricity) and projects this more than doubles to around 945 TWh by 2030 in its base case. The United States is about 45% of that, close to 180 TWh in 2024 (a figure on which the IEA and EPRI independently agree), and the IEA expects data centres to account for nearly half of all US electricity-demand growth to 2030. EPRI’s high-growth scenario puts US data centres at up to 17% of national electricity by 2030.
Data-centre electricity demand, US vs UK (TWh per year), verified anchor years
- 2024 US AI–nuclear deals signed
- 2026 OpenAI pauses Stargate UK
That demand is financing new nuclear in the United States. Britain has the same demand. It has the same nuclear ambitions. It does not have the same regulatory framework.
Britain’s Pieces on the Board
Britain has assembled most of the technical components to replicate the American model. The framework to connect them does not exist.
Rolls-Royce SMRs at Wylfa
Rolls-Royce SMR was named the UK’s preferred reactor technology in June 2025, a 470 MWe design now in the final stage of the regulator’s Generic Design Assessment, due to complete in August 2026. In November 2025 the government chose Wylfa on Anglesey as the first site, ahead of the rival Oldbury site, with an initial three units totalling around 1.4 GWe. An April 2026 contract funded the pre-construction design work.
Two caveats keep this honest. No British SMR has yet generated a watt: the design has not finished assessment, no final investment decision has been taken, and grid connection in the “mid-2030s” is an ambition rather than a committed date. And the standardised, factory-built model that is meant to make SMRs cheaper than bespoke gigawatt reactors is, so far, a promise rather than a proven cost.
Oldbury in Gloucestershire (bought, with Wylfa, from Hitachi in 2024) remains a government-owned candidate site rather than a confirmed one. Heysham in Lancashire hosts ageing conventional reactors and is a long-standing designated nuclear locality, but it has not been selected for SMRs. The realistic near-term picture is one confirmed SMR site, not a fleet of three.
AI Growth Zones: Planning Reform, Patchy Power
The government launched its AI Opportunities Action Plan in January 2025, creating AI Growth Zones with faster planning and prioritised grid access. By early 2026 five had been designated. Crucially, the policy is candid that power is the binding constraint: ideal sites need “large existing power connections” of 500 MW or more, or proximity to major energy infrastructure including nuclear reactors, and the government calls timely grid connection “the single biggest blocker” to establishing the zones.
There is exactly one place where this policy meets new nuclear. The North Wales zone is described as “uniquely positioned” to leverage the Wylfa SMR, the only explicit link in UK policy between an AI Growth Zone and a reactor. The catch is that the connection runs one way. The North Wales zone reaches toward Wylfa, but the Wylfa SMR announcements themselves make no mention of AI, data centres, or co-location: the reactor is justified on energy security and net zero. Of the five zones, the others are powered by renewables and the grid, not nuclear. The Stargate UK project that OpenAI paused was to sit in the renewable-powered North East zone, not beside a reactor. AI demand as patient capital for British nuclear is, for now, an opportunity visible in the policy, not a plan written into it.
The Grid Problem: Anglesey to London
Wylfa sits on an island off the northwest coast of Wales. The data centres that would buy its power are clustered 250 miles southeast, around London. The grid between them cannot carry gigawatt-scale flows. Anglesey County Council flagged transmission capacity as the key enabler of the region’s “Energy Island” ambitions back in 2009. Sixteen years on, the upgrades needed for multiple gigawatt-class reactors have not been built.
New transmission from Anglesey to the southeast would need planning approval, land acquisition, and years of construction at a cost running into billions. Even if approved today, it would not be operational before the late 2030s, well after the window in which AI demand is driving hyperscalers to sign long-term contracts. The American model bypasses the grid entirely through co-location. Britain’s framework has no pathway for it.
The Regulatory Wall
Behind-the-meter arrangements let a consumer connect directly to a generator without using the public network. In the United States this is established practice for large industrial users and, increasingly, for data centres beside nuclear plants.
How Behind-the-Meter Works
When a data centre operates behind the meter at a nuclear plant, power moves straight from generator to load. There are no transmission charges, no distribution charges, no balancing costs; a long-term contract fixes the price per MWh for decades, and faults on the public network do not reach the co-located facility. The savings are large. UK industrial electricity prices are loaded with transmission, distribution, balancing and policy costs that, stacked together, can rival the wholesale cost of the power itself. Behind-the-meter removes most of them.
Why Britain Blocks It
British regulators treat behind-the-meter arrangements with deep suspicion, and their reasoning is not baseless. If a large consumer leaves the grid, the fixed cost of maintaining the network is spread across fewer remaining users, raising their bills; pulling firm generation out of centralised dispatch can also affect system stability. This is precisely the concern that drove FERC to reject Amazon’s expanded Susquehanna arrangement.
But there is a distinction the rules miss. Those concerns apply to an existing grid-connected generator choosing to go private. They have little bearing on new capacity that would never have been built without the behind-the-meter contract financing it. An SMR built at Wylfa specifically to power a co-located data centre removes nothing from the grid, because without the data centre contract the reactor is not built at all. Applying grid-exit rules to new-build co-location prevents new capacity from being financed in the first place.
Stranded Assets on Anglesey
Building reactors at Wylfa to sell into a constrained grid carries a real stranded-asset risk: spending billions on units that cannot export their full output because the wires south are too thin. Co-location removes that risk. The reactors generate, the adjacent data centre consumes, and the grid bottleneck becomes irrelevant. The North Wales AI Growth Zone already creates a policy home for data-centre development in the region. What is missing is explicit permission to connect the two behind one meter.
Britain’s Warning Sign
OpenAI’s pause of Stargate UK was not subtle. Its flagship UK data centre (a partnership with Nvidia and Nscale at Cobalt Park in North Tyneside, part of a wider £31 billion UK-US technology package) shelved over the cost of energy and regulatory uncertainty. OpenAI says it will proceed when “the right conditions, such as regulation and the cost of energy,” allow. It is paused, not cancelled. The signal is the point: a company spending tens of billions globally looked at British electricity and walked.
The Price Gap
Britain has the highest industrial electricity prices in the developed world. In 2024 it was the most expensive of all the countries reporting to the International Energy Agency, at around 26 pence per kWh, roughly four times US levels and well above the IEA median. This is the same structural fact Part 1 documented: it is why Britain’s Bitcoin miners run their machines in Texas, and it is why AI hyperscalers are wary of building here.
The wholesale gap is structural, not a passing spike. Through the first half of 2025, IEA data put British wholesale power at roughly $115 per MWh against about $48 in the United States and $73 in France. France runs around 70% nuclear and pays markedly less. AI training needs constant power for months at a time, a load profile that suits firm nuclear baseload and sits awkwardly with intermittent renewables backed by gas. The country with abundant, firm, low-cost power wins this investment. Britain, today, is not that country.
The Financing Trap
If British electricity is too dear to attract hyperscale data centres, Britain loses the very mechanism financing new nuclear in America. Microsoft’s 20-year PPA works because the economics work for Microsoft. Make British power uncompetitive over a 20-year horizon and the hyperscalers build elsewhere, taking the patient capital with them. Britain needs new nuclear for energy independence; AI data centres need firm baseload and will pay for it over decades. The grid and the pricing keep the two apart. The demand sits on one side, the technology on the other, separated by a framework designed for a different era.
What Would Actually Work
Britain can copy the American outcome with reforms aimed squarely at nuclear-data-centre co-location.
Distinguish new build from grid exit. An existing generator going private shifts costs and deserves scrutiny. A new generator built for a co-located load creates capacity that would not otherwise exist. An SMR at Wylfa built to power an adjacent data centre belongs firmly in the second category: it removes nothing and shifts nothing. The regulations should say so explicitly.
Keep a grid connection for export. Co-located sites can retain a grid link to sell surplus power when data-centre demand dips below reactor output, adding flexible capacity to the system without charging the data centre full transmission costs for power it never imports.
Unify the planning process. Data-centre approvals run through the AI Growth Zones framework; nuclear runs through a separate National Policy Statement. A co-located project should get a single, joined-up approval that treats the reactor and the data centre as one facility. Two tracks for one site is a recipe for delay.
Tie permission to signed contracts. If Britain wants hyperscalers to sign 20-year PPAs with SMR developers, planning permission for co-located sites could be made contingent on a signed offtake agreement, giving both sides legal certainty and investors confidence.
Designate co-location zones. Wylfa’s first phase is around 1.4 GW of SMR capacity, enough for two or three large data centres behind the meter, with room for further units later; Oldbury could follow if confirmed. Designating these as nuclear-AI co-location zones, with pre-cleared planning and explicit behind-the-meter rules, would tell international investors that Britain intends to compete for this infrastructure rather than watch it leave.
Counterarguments and Genuine Complexity
An evidence-based case has to meet the strongest objections, not the weakest.
Hyperscalers buy renewables, not nuclear. This is the most serious challenge, because it describes Britain’s actual revealed preference. Four of the five AI Growth Zones are powered by wind, solar and batteries, not reactors; the Stargate UK project was bound for a renewable-powered zone. The honest rebuttal is narrow but real: renewables plus storage cannot yet deliver the round-the-clock, >95%-utilisation firm power that AI training campuses demand at scale, which is exactly why America’s largest buyers turned to nuclear after starting with renewables. Britain may follow the same learning curve — or may not, if storage improves faster than reactors are built.
The SMRs do not exist yet. Rolls-Royce SMR has not completed design assessment, has taken no final investment decision, and “mid-2030s” power is an aspiration. Building an AI-siting strategy on unbuilt reactors is speculative. The counter is that long-horizon offtake is precisely what de-risks a final investment decision: patient capital is the cause of the reactor, not a reward for it. But that is an argument, not a guarantee, and Britain should not pretend the reactors are already real.
The price problem is policy-made, so fix the policy instead. If British power is dear because of levies, carbon costs and grid charges rather than physics, the cheaper remedy is reforming those charges, not building reactors to escape them. This is a genuine alternative, and the strongest version of the case for co-located nuclear has to concede that cheaper grid power would help every consumer, not just hyperscalers.
Behind-the-meter undermines the grid. Pulling firm generation out of centralised dispatch can raise costs for remaining users and complicate balancing, the reasoning behind FERC’s rejection of Amazon’s expanded arrangement. The reply is that new-build co-location adds capacity the grid never had, and that retained export connections and standby charges can keep cost allocation fair. The concern is real and must be designed around, not dismissed.
AI demand may be a bubble. Britain’s own connection queue — more than 70 GW of requests against a central forecast a fraction of that size — is itself evidence of heavy attrition: most requested capacity never gets built. If the boom fades, 20-year contracts could strand reactors. This risk is real, but it applies equally to every American deal already signed; the difference is that Microsoft, Amazon and Google are betting their own shareholders’ money on the demand being durable, and Britain is betting nothing at all.
Who Benefits from These Numbers
Every source here has reasons to present data a particular way, and the reader should know them.
The Nuclear Industry Association’s analysis of nuclear-powered data centres comes from a body that exists to advocate for nuclear; its employment and economic-impact figures are best-case advocacy, not independent analysis. Rolls-Royce’s job projections are developer estimates designed to build support. Constellation’s “Crane Clean Energy Center” branding is corporate rebranding, and its $1.6 billion restart figure comes from its own materials. The hyperscaler commitments are commercial: Microsoft, Amazon, Google and Meta need firm power and will pay for it, which is why their move into nuclear is informative precisely because it is self-interested. Government enthusiasm for AI Growth Zones reflects pressure to show growth; the “£100 billion” figure is an aspiration, not committed capital.
The demand projections deserve particular care. The IEA and EPRI figures are scenarios, not forecasts; EPRI’s 17% is its high-growth case, and the IEA’s 945 TWh covers data centres broadly rather than AI alone. We use these sources for their hard data (signed contracts, measured prices, grid figures) and treat their forecasts and recommendations as what they are.
The Window That Closes
AI demand is financing new nuclear baseload in the United States right now, turning projects that markets would not fund into bankable ones through long-term contracts. Britain has the reactor technology in the Rolls-Royce SMR, a confirmed first site at Wylfa, and a policy framework in the AI Growth Zones. What it lacks is regulatory permission to combine them through behind-the-meter co-location.
The alternative is the familiar one. Watch billion-pound data-centre investments leave over electricity costs, as OpenAI’s pause already shows. Keep paying for energy imports, as Part 1 of this series documented. Build reactors at remote sites that cannot move their power to where the demand is.
That demand will be met somewhere: the IEA’s roughly 945 TWh of global data-centre electricity by 2030 will be served by some grid, in some country. Hyperscalers planning 20-year investments need equally long contracts; SMR developers need that revenue certainty to raise finance; co-location solves both at once and routes around the grid bottleneck that would otherwise strand the reactors. The technology exists. The demand exists. The financing model exists, proven in America. The regulatory permission does not.
That is the only thing standing between Britain and the patient capital that could build its nuclear future.
Data Sources & References
UK Government: Nuclear & SMR Programme
- Rolls-Royce SMR selected as UK’s preferred technology (June 2025) (470 MWe; over £2.5bn committed; mid-2030s ambition)
- North Wales / Wylfa SMR + AI Growth Zone confirmation (Nov 2025) (Wylfa first SMR site; North Wales AGZ)
- Wylfa SMR contract kickstarts UK SMR plans (ANS, April 2026) (≈1.4 GWe over 3 units; pre-FID design contract)
- Rolls-Royce SMR contract with UK government (April 2026)
UK Government: AI Growth Zones & Compute
- AI Opportunities Action Plan launch (January 2025)
- Delivering AI Growth Zones policy document (power as siting prerequisite; grid connection the “single biggest blocker”)
UK Demand, Grid & Price
- NESO data-centre demand evidence to Parliament (FES 2025) (7.6 TWh in 2024 → ~20 TWh central by 2030; 72.8 GW of connection requests)
- DESNZ international industrial energy prices (UK highest in the IEA, ~26 p/kWh, ~4× US)
- Britain’s AI hopes and high electricity costs (Reuters / IEA, Aug 2025) (UK ~$115/MWh vs US $48 wholesale, H1 2025)
- Anglesey County Council transmission submission (2009)
- Data centres: planning, sustainability, resilience (House of Commons Library) (UK industrial electricity cost components)
US Hyperscaler Nuclear Deals
- Microsoft-Constellation Three Mile Island PPA (Sept 2024) (20-year PPA; 835 MW; ~$1.6bn restart; 2028 target)
- Microsoft 20-year PPA analysis (Orrick)
- Amazon-Talen Cumulus campus sale (March 2024) (960 MW; ~$650m)
- FERC rejects expanded Susquehanna interconnection (Nov 2024) (2-1; 300→480 MW)
- Google-Kairos Power SMR agreement (Oct 2024) (up to 500 MW by 2035)
- Google-Kairos-TVA partnership (Aug 2025) (first Gen IV reactor PPA with a US utility)
- Meta nuclear RFP (Dec 2024) and scope, 1-4 GW, SMRs included
Demand Projections
- IEA, Energy and AI (2025) (~415 TWh data centres in 2024 → ~945 TWh by 2030; US ~half of US demand growth)
- EPRI data-centre load growth (up to 17% of US electricity by 2030, high-growth scenario)
UK Data-Centre Pause
- OpenAI pauses Stargate UK over regulation and energy cost (Reuters, April 2026)
- Stargate UK identity and the £31bn UK-US package (BBC)
Industry & Behind-the-Meter
- Nuclear Industry Association: Powering the UK Data Boom (Dec 2025) (advocacy; employment and co-location case)
- Behind-the-meter for UK data centres (Data Center Dynamics) (industry commentary on regulatory barriers)
This is Part 2 of a two-part series. Part 1 — “The Buyer of Last Resort” — examined how Bitcoin mining could monetise Britain’s stranded wind and flared gas. Both parts trace one root cause: the developed world’s most expensive industrial electricity.