Nuclear and data centres: beyond the PPA

The commercial models being used to connect nuclear energy with hyperscaler demand are more complex than they appear to be and carry risks that neither side has fully priced.

Between September 2024 and January 2026, the four largest technology companies by data centre scale, Alphabet, Amazon, Meta, and Microsoft, each signed agreements to access nuclear power. In aggregate, those commitments represent up to 10 gigawatts of potential capacity and have been widely described as the moment when the nuclear-hyperscaler relationship moved from aspiration to contract.

The description is broadly accurate. But looking at what has actually been agreed, a more complex picture emerges. The transactions use similar language (power purchase agreement, clean energy contract, long-term deal) to describe arrangements that are structurally very different from one another. One company is buying electricity. Another is buying carbon credits from electricity it will never receive. A third is signing options on capacity from reactors that have not yet been licensed. Each carries different risks, requires different regulatory treatment, and provides a different level of revenue certainty for the nuclear asset.

Understanding the commercial architecture of these deals, what each structure actually is, what it does and does not guarantee, and where the risks sit, is not just a legal question. It is a strategic one. For utilities seeking to monetise existing or restarted nuclear capacity, for technology companies under pressure to demonstrate energy security and sustainability, and for the financing structures that must underpin any new nuclear investment, the commercial model is where ambition either converts into bankable certainty or stalls.

Three structures, not one

The transactions that have defined the nuclear-hyperscaler relationship since 2024 fall into three categories, each of which allocates risk differently.

The first is a conventional power purchase agreement, in which a technology company agrees to buy electricity from a specific nuclear plant over a defined term. Microsoft’s 20-year agreement with Constellation Energy for 835 megawatts from the restarted Three Mile Island Unit 1, signed in September 2024, is the clearest example. Amazon’s 17-year, 1.92 gigawatt agreement with Talen Energy for power from the Susquehanna nuclear plant, announced in June 2025, is another. Under these arrangements, the nuclear operator generates and delivers power, the technology company receives and pays for it, and the PPA price provides the revenue certainty needed to support the capital investment required to restart or extend the plant’s operating life.

The second model is structurally different. Meta’s 20-year agreement with Constellation Energy for 1,121 megawatts from the Clinton Clean Energy Center in Illinois is not a power purchase in the conventional sense. Meta is not buying electricity for its data centres. The plant continues to sell its output to the regional wholesale market. What Meta has purchased is the plant’s clean energy credits or the right to count Clinton’s zero-carbon generation toward Meta’s own renewable energy targets. The commercial logic is coherent: Meta’s long-term commitment provided sufficient economic certainty for Constellation to reverse plans to close the plant and renew its operating licence. But the mechanism is different from a conventional electricity PPA, and its treatment under future clean energy accounting standards is not guaranteed to remain unchanged.

The third model is still earlier-stage. Alphabet’s agreement with Kairos Power, targeting 500 megawatts of advanced reactor capacity by 2035, and Amazon’s $700 million investment in X-energy to support the Xe-100 small modular reactor are structures better understood as options on future capacity than as power purchase agreements. They signal demand and provide commercial validation for reactor developers, but the power they represent depends on technologies that have not yet been deployed at commercial scale. Project finance lenders treat them accordingly.

The distinction has practical consequences. A conventional electricity PPA for an operating plant can be used directly to support project financing; it provides the revenue certainty that lenders require to assess debt serviceability. A clean energy credit purchase provides softer commercial support. A development partnership with an unproven technology provider provides little that a lender can underwrite. Organisations designing nuclear supply arrangements need to understand which of these structures they are entering, and what it does and does not enable.

The grid question

Of the transactions completed to date, the Amazon-Talen Energy sequence is the most instructive. Not because of its size, but because of what happened when the proposed commercial structure was tested against regulatory reality.

The original deal proposed a behind-the-meter arrangement: Amazon’s data centre campus, located adjacent to the Susquehanna nuclear plant in Pennsylvania, would receive power directly from the plant without routing it through the regional transmission grid. From an operational perspective, this model has clear appeal. It eliminates transmission losses, avoids grid congestion, and provides dedicated, uninterrupted supply to a facility whose energy requirements are continuous and predictable.

In November 2024, the Federal Energy Regulatory Commission rejected the interconnection service agreement required to implement the arrangement, voting 2-1. The commission’s concern was not technical. It was commercial: the proposed structure would allow Amazon’s data centres to draw on the Susquehanna plant’s output while avoiding the transmission charges that other grid users pay, potentially shifting up to $140 million in annual costs onto other ratepayers.

The deal was restructured. In June 2025, Amazon and Talen announced a revised arrangement: a 17-year, $18 billion front-of-meter PPA in which Talen sells power to the regional wholesale market and serves Amazon as a licensed retail electricity provider, delivering nuclear-sourced power across the standard grid. Amazon receives the same carbon-free nuclear power. But it receives it as a grid customer, not as a co-located facility.

The FERC decision has become a reference point for any organisation considering behind-the-meter nuclear supply to a data centre. It establishes that dedicated nuclear supply, even where that supply is physically adjacent to the generator, is not treated as a purely private commercial arrangement. Transmission cost allocation is a regulated question, and the answer affects how the economics of the deal are distributed across the wider grid. Organisations designing nuclear supply arrangements for data centres need to engage with the regulatory structure early, and certainly before commercial terms have been agreed.

A mismatch in time horizons

The most significant structural challenge in the nuclear-hyperscaler relationship is one that the headline numbers obscure. The PPAs being signed run for 17 to 20 years. Nuclear power plants, following life extension approvals, are licensed and operated for 60 to 80 years. The gap between those two figures represents a commercial problem that has not been seriously addressed in most of the transactions announced to date.

For a nuclear operator, a 20-year PPA provides revenue certainty for a fraction of the plant’s economic life. The Microsoft-Constellation agreement for Three Mile Island runs from approximately 2027 to 2047. The Susquehanna plant, under its current licence, could operate well into the 2050s. The revenue model for the years after the PPA expires depends on market conditions, regulatory renewal, and the availability of a new offtaker in a market that does not yet exist.

For a technology company, the duration concern runs in the other direction. A 20-year commitment to pay approximately $110 to $115 per megawatt-hour, the estimated fixed price in the Microsoft-Constellation deal, is a substantial long-term financial obligation. In 2027, when the Crane Clean Energy Center returns to service, that price may be highly competitive relative to wholesale market rates for firm, low-carbon power. A decade later, as the energy landscape continues to evolve, the comparison may look different. Fixed-price long-term contracts provide certainty, but that certainty is not symmetric: it protects against price increases, but does not protect against the possibility that better or cheaper alternatives emerge.

The deeper issue is what happens if a data centre’s energy requirements change significantly before the PPA expires. Facilities designed for one generation of AI infrastructure may be partially repurposed, consolidated, or closed before a 20-year nuclear offtake agreement reaches its natural expiry. Unlike most renewable energy contracts, nuclear PPAs cannot simply be reassigned or cancelled without substantial financial consequences. Rigorous deal design addresses this by clarifying termination provisions, force majeure definitions, and the conditions under which the offtake obligation transfers or is released. Most of the deals announced to date have not disclosed this level of contractual detail publicly. The risk exists regardless of whether it has been disclosed.

The financing layer

Behind every nuclear-hyperscaler deal is a financing question that the headline announcement does not address. Nuclear plant restarts, life extensions, and uprates require significant capital investment. That investment needs to be financed. And the terms on which it can be financed depend substantially on the creditworthiness and enforceability of the offtake arrangements underpinning it.

The Three Mile Island restart illustrates the dynamic clearly. Constellation’s investment of approximately $1.6 billion in bringing Unit 1 back to service was supported by a $1 billion loan from the Department of Energy’s Loan Programs Office. The federal facility provided both the financing and a signal to private lenders and equity investors that the project met government bankability criteria. The Microsoft PPA provided the commercial revenue certainty. Both were necessary; neither alone would have been sufficient to move the project forward.

This financing structure (corporate PPA supported by government-backed debt) is likely to remain the template for nuclear restart and life extension projects for the foreseeable future. It reflects the credit reality that, even with an investment-grade corporate offtaker, lenders to nuclear projects require additional risk mitigation that private capital alone does not currently provide. The DOE Loan Programs Office has indicated that nuclear will be its largest single use of funds going forward, with a 2026 budget request including $30 billion in new loan authority oriented primarily toward nuclear and firm generation capacity.

For technology companies negotiating nuclear offtake arrangements, the financing structure is not a background consideration. The terms of a PPA, from its duration and price to termination provisions and step-in rights for lenders, directly affect whether the project can attract the project finance it requires to proceed. A PPA that provides insufficient revenue certainty, or that includes termination provisions that lenders regard as credit risk, may leave the nuclear asset unable to raise the capital it needs. In that scenario, the deal is signed, but the project does not advance. Organisations that understand this dynamic negotiate PPA terms with the financing structure explicitly in view, rather than treating the commercial agreement and the project finance as separate conversations.

A market that is not yet scaling

The headline figures from the nuclear-hyperscaler deals of 2024 to 2026 suggest a transformation in the sector’s commercial dynamics. The underlying numbers suggest something more measured.

The total nuclear capacity represented by hyperscaler agreements amounts to approximately 13 gigawatts. If all of those commitments were fulfilled on the timelines announced, they would generate approximately 102 terawatt-hours of electricity per year. Against the mid-range projections for US data centre electricity demand through 2035, that figure represents less than 20 percent of projected need.

The gap reflects a structural tension. The hyperscalers are spending at extraordinary scale on data centre infrastructure, $443 billion in 2025 and a projected $700 billion or more in 2026. Nuclear commitments, by contrast, represent a fraction of that capital deployment. The reason is partly timing: new nuclear capacity requires years of development, and the AI infrastructure buildout is moving on a quarterly schedule, not a decadal one. But it is also structural: the commercial models for nuclear-hyperscaler supply are not yet standardised or replicable enough to scale at the pace the gap implies.

Closing the gap will require deal structures that are more consistent and replicable, regulatory frameworks that are more clearly established following the Federal Energy Regulatory Commission precedent, and financing structures that depend less on bespoke government support for each individual transaction. None of those conditions yet exist at the scale the opportunity implies. The organisations that contribute to building clearer frameworks for grid interconnection treatment, for PPA bankability criteria, for the contractual provisions that lenders require, will shape the commercial environment in which subsequent deals are negotiated.

Getting the structure right

The nuclear-hyperscaler deals of the past two years have demonstrated that a commercial market for dedicated nuclear supply to data centres is forming. They have also demonstrated that the contractual, regulatory, and financing complexity of these arrangements is not always fully understood by all parties entering them.

For nuclear operators, the priority is revenue clarity across the full plant life, not just the PPA term. A 20-year offtake agreement that covers a plant’s operating costs provides near-term certainty, but a plant that cannot demonstrate a credible commercial model for years 21 and beyond will face increasingly difficult conversations as PPA expiry approaches. Agreements that include extension options, price reset mechanisms, or structured transition provisions for market-based offtake after the initial term strengthen the long-term commercial position and make licence renewal arguments easier to sustain.

For technology companies, the priority is understanding what the commitment actually involves. An energy credit purchase and a conventional power purchase agreement both involve long-term contracts with a nuclear utility, but they represent different obligations, different regulatory exposures, and different accounting treatments. A behind-the-meter arrangement and a front-of-meter retail supply deal carry very different grid interaction requirements. The organisations that engage with these distinctions early, with regulatory counsel, project finance advisors, and energy procurement specialists involved from the outset, are less likely to face the structural renegotiations that the Amazon-Talen sequence required.

For both sides, the lesson from the transactions completed to date is that the commercial model for nuclear-hyperscaler supply is not yet standardised. Each transaction has required significant bespoke structuring. The immediate challenge is not a shortage of intent, but the absence of a commercial and regulatory framework mature enough to scale. Building that framework is as important as signing the next deal.