For most of the last decade, the phrase "AI is going to need nuclear" functioned inside the technology industry the way every good macro thesis does at first, as a line you could say at a dinner in Menlo Park and get nods without being asked to defend it. It was directionally obvious. Training runs were getting bigger, inference footprints were getting denser, and the fundamental physics of cooling racks full of GPUs did not leave much room for creative accounting about where the electrons would come from. What the thesis did not have, through most of 2023 and 2024, was a paper trail. There were letters of intent. There were press releases about studying reactors. There were framework agreements that let a CEO say the word nuclear on an earnings call without actually committing a dollar. What there was not, with any volume, was a signed power purchase agreement attached to steel in the ground.
That changed, and it changed faster than almost anyone inside the utility sector expected. On January 27, 2026, Meta disclosed a trio of nuclear deals that together represent as much as 6.6 gigawatts of firm, carbon free electricity earmarked for its artificial intelligence data centers. Six point six gigawatts is not a marketing number. It is roughly the output of six full size reactors, or a meaningful fraction of the entire installed nuclear fleet of a midsize European country, routed to the private compute needs of a single American advertising company. The deal came seven months after Meta's June 2025 announcement of a twenty year agreement with Constellation Energy to take 1.1 gigawatts of output from the Clinton Clean Energy Center in Illinois starting in 2027, the first contract of its kind to be structured the way renewable PPAs have been structured for a decade, with a fixed off taker, a fixed price, and a fixed term. Between the Clinton deal and the January trio, Meta has put its name on more firm nuclear capacity than most American states will build in the next ten years. Microsoft, Amazon, and Google are not far behind, and the bluesky language on their slide decks is being rewritten in real time into contract language, filed with public utility commissions, and flagged by rating agencies.
What makes this moment worth sitting with is not the headline gigawatt number but what it does to the shape of several adjacent industries simultaneously. Nuclear construction in the United States, by any honest accounting, has been in stasis for forty years. Small modular reactor companies have raised venture capital on PowerPoint for almost a decade without a single unit operating commercially on American soil. State level energy policy has been paralyzed by the inability to decide whether to treat new nuclear as an infrastructure problem or a technology problem. The hyperscalers have walked in with a checkbook, a load curve, and a deadline, and in roughly eighteen months they have broken that paralysis in several places at once. In April 2026 alone we have watched a Republican governor in Indiana sign a small modular reactor tax incentive bill, a venture backed fission company file suit against the federal government alongside the state of Louisiana to overturn a licensing rule, and Taiwan Semiconductor Manufacturing Company report blockbuster quarterly growth that reinforces the demand signal the nuclear race is trying to meet. These are not separate stories. They are one story told in four different dialects.
From Bluesky Thesis to Signed Contract
The speed of the transition is the part that deserves the most attention. Infrastructure of this type does not normally move this fast. A traditional utility scale generation project in the United States measures its development cycle in decades, not quarters. The environmental impact statement alone can run five years. Siting, permitting, interconnection studies, state regulatory approval, financing, and then the actual construction phase are each line items measured in years, and they are typically sequential rather than parallel. The reason the hyperscalers have been able to compress that timeline is that most of them are not, at this stage, trying to build new plants from scratch. The first wave of deals, Meta's Clinton agreement being the canonical example, is structured around existing reactors that have already cleared every regulatory hurdle and just need a credit worthy off taker willing to sign a twenty year commitment. For Constellation Energy, a twenty year contract from Meta is the kind of balance sheet anchor that justifies life extension investments, uprate projects, and the capital spending that keeps a forty year old reactor competitive against a new gas combined cycle plant with zero fuel price volatility.
The second wave, which is where the January 2026 Meta trio begins to live, is structured differently. These deals layer existing reactor output with options on new capacity, including small modular reactors that are not yet operating, not yet licensed, and in some cases not yet fully designed. What Meta is buying there is not kilowatt hours. It is a call option on future firm capacity, priced against the risk that the SMR timeline slips and underwritten by the willingness of a technology company with three hundred billion dollars of cash to eat some of that risk itself. That is a structural innovation. It is the first time in American nuclear history that a private off taker has been willing to put real money into a project before the design has been certified, and it is only happening because the alternative, which is failing to secure firm power for AI training clusters in 2028 and 2029, is worse.
The financial community has noticed. Constellation Energy's share price roughly tripled between the first Meta discussions and the Clinton signing. Vistra and Talen, the other two large merchant generators with nuclear in their portfolios, have seen similar if smaller moves. The old story about nuclear, which was that it was a stranded asset in a world of cheap gas and subsidized renewables, has been rewritten inside eighteen months into a new story about nuclear as the only firm carbon free generation technology that can plausibly scale to multi gigawatt data center campuses. Whether that new story survives contact with the cost overruns and construction delays that have historically plagued American nuclear is the open question that every participant is trying not to talk about in public.
Meta as the Biggest Buyer
Meta has taken a position in this market that is worth describing precisely, because it is quite different from what Microsoft, Amazon, or Google are doing. Microsoft has made loud commitments around the restart of Three Mile Island Unit 1 through its agreement with Constellation, which is a real deal but a smaller one, roughly 835 megawatts. Amazon purchased the Cumulus data center campus adjacent to the Susquehanna nuclear plant in Pennsylvania and built its capacity strategy around that single colocation, a move that was later complicated by a Federal Energy Regulatory Commission ruling on the interconnection arrangement. Google has signed agreements with Kairos Power for small modular reactors and with NV Energy for a project in Nevada, putting it into the SMR pipeline early but in smaller tranches. Meta, by contrast, has signed the largest raw number, 6.6 gigawatts, and has signed it across three separate counterparties, which is the portfolio approach you would expect from a buyer that expects some of the projects to slip and wants optionality if one counterparty runs into trouble.
The six point six number also needs to be read against Meta's published capital expenditure plans. Mark Zuckerberg has publicly committed to capital spending in the range of sixty to seventy billion dollars for 2026, a number that is itself roughly double what it was two years ago, and the majority of that capital expenditure is AI infrastructure. The company's Louisiana data center campus, the Hyperion project, is planned to grow to multi gigawatt scale on its own. Meta is not buying 6.6 gigawatts of nuclear because it likes round numbers. It is buying 6.6 gigawatts because the internal load forecasts for Llama training and inference, combined with the external workload that Meta AI is placing on the same fleet, produce a power demand curve that the company cannot reliably fill from the merchant grid alone in the locations where it has land and fiber. Nuclear is the residual, the thing that gets bought once the renewable and gas options have been exhausted and the shortfall is still large.
That framing matters, because it reveals what the hyperscalers actually believe about the shape of the next five years. They do not believe that inference efficiency gains from better silicon, better compilers, and better model architectures will close the power gap. They believe those gains are real but will be outrun by the workload. They do not believe that demand response or curtailment will be acceptable on training runs, because a training run that gets curtailed becomes a training run that takes longer and costs more at a moment when the competitive advantage of shipping the next model first is measured in billions. And they do not believe that the existing American generation mix, which is majority natural gas with a thick layer of wind and solar on top, can meet the reliability and carbon targets their largest customers are writing into their own sustainability reports. Nuclear is the only resource on the menu that satisfies all three constraints simultaneously, which is why Meta is willing to pay a premium for it and to eat construction risk that a utility would not take.
Small Modular Reactors, Real Money, and a Federal Lawsuit
The small modular reactor story is the part of this moment that has the steepest gap between narrative and delivery. SMRs have been, for most of the last ten years, a reliable way to get an investor meeting in Silicon Valley and a reliable way to disappoint everyone who showed up. NuScale Power, which was the most visible publicly traded SMR developer, saw its flagship Utah Associated Municipal Power Systems project collapse in 2023 when the cost estimate climbed from roughly fifty eight dollars per megawatt hour to eighty nine dollars per megawatt hour and the municipal off takers walked. That failure cast a shadow over the entire sector, and for most of 2024 the consensus view among utility planners was that SMRs were going to be a 2035 story at the earliest, not a 2028 story.
Eighteen months later the consensus has been rewritten, and it has been rewritten by two forces pulling in opposite directions. On one side, the hyperscalers have started writing the kinds of offtake commitments that make SMR projects financeable for the first time. A twenty year fixed price commitment from Meta or Google is functionally equivalent to the municipal bond backstops that used to underwrite new nuclear in the 1970s, except that the credit rating of a hyperscaler is higher than that of a typical municipality and the willingness to absorb schedule risk is, at this moment at least, greater. On the other side, the federal regulatory apparatus that governs SMR licensing has been slower to adapt than the private market would like. The Nuclear Regulatory Commission has a reactor licensing framework that was designed for gigawatt scale light water reactors built over a decade, and it has struggled to adjust that framework to the very different risk profile of a reactor that is ten times smaller, factory built, and designed to be walk away safe.
That tension produced the Deep Fission lawsuit. Earlier this month, Deep Fission, a venture backed reactor company working on a deep borehole small modular design, joined with the state of Louisiana in filing suit against the federal government to overturn a specific NRC rule that the plaintiffs argue is inhibiting the commercial SMR market. The details of the rule are technical and fall outside the scope of this article, but the political signal is not. A private company and a state government have decided that litigation is faster than rulemaking, and they have decided that the economic stakes of being first to market with a commercial SMR are large enough to justify taking the federal government to court. That would have been unthinkable in the NuScale era. It is thinkable now because the hyperscaler demand curve makes the prize worth the fight.
The Indiana development runs on a parallel track. In April 2026, Governor Mike Braun signed legislation providing state tax incentives for energy companies developing small modular reactors in Indiana, framing the bill explicitly around the need to attract data center investment to a state that has been losing ground to Virginia, Ohio, and Texas in that competition. Indiana is not a nuclear state in any traditional sense. It is a coal state that has spent the last fifteen years slowly shifting into natural gas. The decision to put state money behind SMR development is a recognition that the data center economic development game has changed, and that the winners of the next round will be states that can credibly promise firm, carbon free power at multi gigawatt scale within a four year window. Indiana is betting that an SMR incentive structure today buys it a seat at the table when Meta or Google is picking its next Midwest campus in 2028.
The Grid Problem Nobody Wants to Own
Underneath all of this sits a problem that the American electricity grid has never faced before at this scale. The problem is not that data centers use a lot of power, although they do. The problem is that AI data centers use power in a way that is fundamentally different from any other class of large industrial load that utilities and grid operators have learned to plan around. A traditional industrial load, a steel mill or an aluminum smelter, has a relatively predictable duty cycle, a long negotiation period with the utility before it interconnects, and some degree of flexibility on timing. A data center doing AI training has none of those properties. The training workload is high, sustained, and essentially inflexible. Once a model run begins, the operator cannot accept curtailment without destroying weeks of progress and tens of millions of dollars of compute spending. The inference workload is somewhat more flexible, but is growing fast enough that its peak is beginning to rival the training peak, and the two overlap on the same hardware in ways that make demand response negotiations complicated.
The other thing that makes AI load different is the speed at which it arrives. A steel mill announces five years in advance. An AI training campus announces eighteen months in advance, breaks ground in twelve, and wants power in twenty four. Grid operators are being asked to firm up multi gigawatt load commitments on a timeline that is faster than any new generation project can be built, which means the capacity has to come from somewhere that already exists. In most regions, the only somewhere that exists is existing coal plants that were scheduled for retirement, existing gas plants running at higher capacity factors than they were designed for, and existing nuclear plants that now have a bidding war underway for their output. Every new hyperscaler nuclear deal is, in some accounting, a subtraction from the pool of firm capacity available to everyone else on the same grid. The public utility commissions have started to notice, and a new category of regulatory dispute is emerging around whether a merchant generator can sell its entire output to a single technology company or whether some portion of that output must remain available to the broader customer base at regulated rates.
That dispute is the load commitment problem in its sharpest form. It has not been resolved in any jurisdiction yet, and it is going to be resolved differently in different places, and the outcome of those resolutions is going to determine how much of the nuclear capacity Meta, Microsoft, Amazon, and Google have paid for actually ends up in their data centers rather than being clawed back by state commissions acting on behalf of residential ratepayers. The cleanest legal ground is the new build case. If a hyperscaler pays for an entirely new reactor that would not otherwise exist, no state commission has a strong claim to redirect its output. The messier ground is the existing reactor case, which is exactly where most of the January 2026 deal value sits, and which is where the arguments are going to be loudest over the next eighteen months.
State Policy Fragmentation and the New Regional Map
One of the second order effects of the hyperscaler nuclear pivot is that it is producing a new regional map of the United States that does not match the old regional maps anyone in Washington has been working from. The old data center map was Northern Virginia, the Pacific Northwest, and central Texas, with a scattering of smaller hubs. The new map is being drawn by where firm power can be brought online fastest, which is a different question with a different answer. Pennsylvania, Illinois, South Carolina, Georgia, and Tennessee, states with existing nuclear fleets and merchant generation willing to sign long dated contracts, are going to get a disproportionate share of the next AI buildout. Indiana, Ohio, and Louisiana are fighting to join them, with state level incentive structures that are more aggressive than anything the federal government has produced. Virginia, which has been the dominant data center destination for a decade, is losing its advantage because Dominion Energy's ability to firm up multi gigawatt new load on a two year timeline is, at this point, genuinely uncertain.
The Indiana incentive bill and the Louisiana lawsuit are the two most visible examples of a broader pattern, which is that state governments have concluded that the federal permitting and licensing apparatus is too slow to deliver the kind of capacity that hyperscalers will pay a premium to secure, and that the economic development returns on being first are large enough to justify unilateral action. Indiana is using its tax code. Louisiana is using litigation. Texas has been quietly amending its interconnection rules to make it easier for large loads to pair with dedicated generation behind the meter. Each of these is an experiment in regulatory arbitrage, and each of them is going to be watched carefully by every other state economic development office that wants to land the next Meta or Google campus.
The fragmentation has a cost. A unified federal strategy on SMR licensing, grid interconnection, and long term load planning would almost certainly produce more capacity faster than fifty separate state strategies pulling in different directions. But a unified federal strategy is not what this industry is going to get, and the hyperscalers have made the calculated decision that waiting for Washington is more expensive than navigating fifty state level chess boards in parallel. That calculation is visible in the way the deals are being announced. When Meta discloses a nuclear PPA, the first thing the press release mentions is which state the reactor is in and which state level official is being thanked. Federal officials are mentioned further down, if at all. The center of gravity in American nuclear policy has moved to the state capitals, and it is not moving back.
The Utility Perspective and the Question of Who Absorbs the Risk
The view from the utility side of this market is considerably less celebratory than the view from the hyperscaler side. The utilities that own existing nuclear capacity, principally Constellation, Vistra, Talen, Duke, and Southern, are enjoying a once in a generation pricing environment and are using it to fund uprates, life extensions, and in a few cases the serious contemplation of new build projects they would not have taken seriously in 2023. That is good for their shareholders and good for their employees. It is more complicated for their regulators, who are watching firm capacity that was paid for by ratepayers over forty years of rate base recovery migrate into long term contracts with private technology companies at rates that may or may not reflect the historical cost basis.
The other complication is financing new capacity. A merchant generator looking at the numbers for a new reactor, whether a full scale AP1000 or a small modular reactor, is trying to solve a construction risk problem that American nuclear has historically been bad at. Vogtle Units 3 and 4 in Georgia, the most recent large reactors to reach commercial operation in the United States, came in at more than double their original budget and years behind schedule. The lesson of Vogtle is not that nuclear cannot be built. The lesson is that it can be built, but the cost overrun risk is so concentrated on whoever holds the construction contract that financing new projects on a merchant basis is nearly impossible without someone, usually the ratepayer or the taxpayer, agreeing to absorb the tail risk. The hyperscaler deals are novel because they introduce a third possible absorber, the private technology off taker, but nobody yet knows how much tail risk the hyperscalers are actually willing to eat if a project runs three years late and thirty percent over budget.
That uncertainty is the reason most of the current deal flow is still anchored on existing reactors rather than new builds. It is easier to sign a twenty year PPA for output that is already on the grid than it is to sign a PPA for output that depends on a first of a kind construction program. The industry will know the nuclear pivot has fully arrived when the first new build reactor is financed primarily against a hyperscaler offtake contract rather than a regulated utility rate base. That milestone has not been crossed yet. It will be crossed within the next eighteen months or the thesis will begin to lose credibility, and every participant in the market knows it.
Competing Solutions and the Efficiency Wildcard
Nuclear is not the only technology competing to supply AI data centers, and the other options are real enough that any honest analysis has to account for them. Enhanced geothermal, led by companies like Fervo Energy, has made real progress in the last three years and has a credible path to multi gigawatt scale in the western United States over the next decade. Fervo's Cape Station project in Utah is already supplying power to Google under a commercial agreement, and the technology has the advantage of being entirely above board on federal public land with fewer licensing hurdles than nuclear. Natural gas peakers, while facing carbon intensity headwinds in sustainability reporting, remain the default choice for any load that needs power in twenty four months rather than sixty, and several hyperscalers have quietly built or contracted for gas capacity to bridge the gap until nuclear and geothermal arrive. Grid scale battery storage, paired with solar and wind, continues to get cheaper, although the duration problem makes it a supplement rather than a replacement for firm generation in AI applications.
The wildcard in all of this is efficiency. Every gigawatt of nuclear capacity the hyperscalers are racing to contract is a gigawatt they will not need if the chips and the data center infrastructure get meaningfully more efficient between now and 2028. There is reason to believe some of that efficiency is coming. Recent research out of UC San Diego, published earlier this spring, described a new chip design that rethinks how power conversion works inside AI accelerator cards, potentially reducing the losses incurred between the rack level power distribution and the silicon itself. On paper, the improvement is in the range of several percentage points of total data center power consumption. Applied across a hyperscaler fleet measured in tens of gigawatts, several percentage points is hundreds of megawatts, which is the size of a small nuclear reactor. Parallel efficiency gains are being reported at the model level, with architectural improvements, better quantization, and more efficient attention mechanisms reducing the compute needed per training step and per inference token.
None of those efficiency gains are going to make the nuclear race unnecessary. The workload is growing faster than the efficiency curve, and will continue to for at least the next three years. But the efficiency gains are going to make the nuclear race less desperate, and they are going to give the hyperscalers room to be more selective about which projects they anchor with PPAs and which ones they let other buyers fund. The story that TSMC told in its most recent earnings report, which was a story of blockbuster growth driven almost entirely by AI accelerator demand, is the bull case for nuclear. The story that UC San Diego is telling about power conversion efficiency is the bear case. Both stories are true simultaneously, and the industry is going to live inside the tension between them for the rest of the decade.
What to Watch Over the Next Eighteen Months
The first thing to watch is the groundbreaking timing on the first hyperscaler anchored SMR project. As of this writing, no small modular reactor in the United States has poured concrete on a site that is committed to a technology company off taker. That milestone is likely to arrive in late 2026 or early 2027, and whichever project crosses it first, whether it is a Kairos project in Tennessee, a TerraPower project in Wyoming, or something from the X energy pipeline, will set the template for how these deals are structured, how state and federal regulators respond, and how quickly the second wave of projects can move. The first groundbreaking is also the moment at which the hyperscaler construction risk exposure becomes real, and the moment at which the financial community will start pricing in whether these deals can actually deliver.
The second thing to watch is the first AI data center that actually powers up on dedicated nuclear capacity. The Meta Clinton agreement begins delivery in 2027, which makes Clinton the current favorite to be the first reactor supplying an AI training cluster under a long term technology company PPA. If that transition goes smoothly, it will validate the portfolio approach and accelerate the second wave. If it runs into interconnection problems, capacity factor shortfalls, or regulatory clawback attempts from the Illinois Commerce Commission, it will become the case study everyone points to when explaining why the nuclear pivot is harder than the January 2026 press releases made it sound.
The third thing to watch is the first outage. Nuclear plants have outages. It is the nature of the technology. When a reactor trips offline, a normal grid absorbs the loss through reserves and balancing markets, and residential customers do not notice. When a reactor trips offline while supplying a dedicated AI training campus under a PPA that specifies firm delivery, somebody has to decide, in the moment, whether the campus runs on backup generation, curtails its workload, or draws from the broader grid at emergency market prices. The contracts being signed today contain clauses that attempt to allocate that risk, but the clauses have not been tested against a real trip on a real reactor serving a real training run. The first time that happens is going to be educational for everyone involved, and the lessons from it are going to rewrite the next generation of contracts.
The fourth thing to watch is how the hyperscaler balance sheets absorb the capital commitments. Meta, Microsoft, Amazon, and Google are wealthy companies, but the combined capital expenditure trajectories they have announced for AI infrastructure, including power, are without precedent in the history of American corporate capital spending. At some point, either the workloads justify that spending through revenue growth or the market begins asking harder questions about return on invested capital. The nuclear PPAs are the piece of that spending that is hardest to walk back, because they are long dated and contractually binding in a way that chip orders are not. If the AI revenue case weakens at any point in the next five years, the nuclear commitments are going to be the commitments that the market punishes first. The hyperscalers clearly believe that is not going to happen. Whether they are right is the single most important question hanging over this whole pivot, and it is a question that only time and revenue can answer.
What we can say with confidence today is that the nuclear thesis has moved from the dinner party to the signature line. Meta has committed to 6.6 gigawatts. Microsoft has committed to Three Mile Island Unit 1. Google has committed to Kairos and Nevada. Amazon has committed to Susquehanna and is fighting for the interconnection that makes it work. Indiana is writing tax incentives. Louisiana is filing lawsuits. The grid is being rewired around a kind of customer it has never served before, and the technology industry that spent a decade treating power as somebody else's problem has been forced to pick up the pen and sign the contract itself. Whether the reactors actually show up on schedule is a question we will answer over the next several years. That the contracts have been signed is a fact we can state today. The rest of the story follows from that.