In 2008, a generator asking to connect to the PJM grid could expect to be operating in under two years. By 2025 that timeline had stretched past eight, with more than 170 gigawatts of requests stacked in the study process and commercial operation dates for new projects extending into the early 2030s.
That single number explains almost everything that follows. Data center nuclear power is not primarily a story about carbon, or about a technology whose moment arrived, or about executives developing an interest in fission. Data center nuclear power is a story about a queue, and about what a company with capital and a deadline does when the orderly path to electricity takes longer than the useful life of the equipment it wants to plug in.
The response has taken three forms, and they are worth separating because they carry completely different risk profiles. Restart a reactor that already exists and already has an interconnection. Contract with an operating plant and take the output. Or commission a reactor that has not been built, of a type that has never operated commercially in the United States, and hope it arrives before the demand does. The first two are happening. The third is mostly a press release with a date on it.
The queue, and why it is the binding constraint
Grid interconnection is the process by which a new generator or a new large load gets permission to connect, and it exists for a reason that has nothing to do with obstruction. Adding a gigawatt of load or generation at a specific point changes power flows across the whole network, and somebody has to study whether the existing wires can carry the result, what upgrades are required, and who pays for them.
That study process was designed for a world adding generation incrementally against demand that had been essentially flat for twenty years. It was not designed for a queue containing more than 170 gigawatts of requests, and it has degraded accordingly. Roughly 57 gigawatts of PJM projects have completed the study process and hold interconnection agreements while remaining stalled by local opposition, permitting, and equipment lead times, which means completing the queue is necessary and not sufficient.
The demand arriving against that bottleneck is the second half of the problem. PJM’s December 2025 capacity auction cleared 6,623 megawatts short of its reliability target, with data centers responsible for nearly 5,100 megawatts of the demand surge. Capacity prices responded, as auctions do when supply misses a target, and the cost showed up on bills across the thirteen states PJM serves. That is the mechanism by which a corporate procurement decision becomes a line item on a residential bill, and it operates without anybody intending it, which is why the politics of large infrastructure loads tend to arrive after the contracts are signed.
Which produces the incentive structure driving everything that follows. A hyperscaler with capital, hardware on order, and a competitive clock cannot wait eight years. An existing power plant, already interconnected, already licensed, already generating, is the only asset that shortens the timeline, and there are a finite number of them. Every data center nuclear power arrangement announced since 2024 is a claim on that finite set.
The load characteristics make nuclear a genuinely good match, which is worth granting before the criticisms. A hyperscale data center runs at high utilization around the clock with limited seasonal variation, which is close to the ideal customer profile for a baseload plant that wants to run flat out and hates cycling. That is a better technical fit than the same plant selling into a wholesale market where it competes against gas on the margin and gets dispatched around cheaper resources. The firm, around-the-clock generation problem has a short list of solutions, and a constant load is the customer each of them was designed for.
Data center nuclear power via restart, and why Three Mile Island was obvious
Unit 1 at Three Mile Island shut down in 2019 with more than a decade remaining on its NRC license. It closed for economics rather than for any technical problem, in a market where cheap gas had made it unprofitable. It was not the unit involved in the 1979 accident, which was Unit 2, damaged beyond repair and permanently shut.
In September 2024 Constellation announced a twenty-year power purchase agreement with Microsoft to restart it, renamed the Crane Clean Energy Center, for 835 megawatts dedicated to Microsoft’s data center operations. The project is a roughly $1.6 billion capital undertaking supported by a Department of Energy loan of about a billion dollars, and Constellation intends to pursue license renewal extending operation to at least 2054. The restart timeline moved earlier, from 2028 to 2027, and as of early 2026 the site was roughly eighty percent staffed with more than five hundred people working toward fuel load.
The regulatory sequence is the part worth appreciating, because it makes the difficulty concrete. Restarting a plant that entered decommissioning status requires rescinding the exemptions granted because it was decommissioning, submitting an updated decommissioning and fuel management report, amending the operating license and technical specifications, filing a preliminary final safety analysis, submitting revised emergency and physical security plans, training an operator cohort from scratch, and obtaining NRC and FEMA sign-off on emergency planning. Constellation’s filed schedule ran that sequence across 2025 through 2027 with an operational readiness letter in July 2027.
Palisades in Michigan went first and became the first restarted reactor in United States history, which established that the pathway exists. It is a narrow pathway. The number of recently shut reactors that retain licenses, intact equipment, an interconnection, and a willing owner is small, and every restart consumes one of them.
Uprates at operating plants are the quieter version and deserve a mention, since recovering additional output from an existing licensed reactor through equipment upgrades and analytical margin adds capacity without any of the siting, licensing, or interconnection work that dominates every other option. The increments are modest, typically a few percent per unit, and they are the cheapest megawatts in the sector. The critical materials and long-lead components that a refurbishment consumes are the same ones every other grid project is bidding for, which is a constraint that scales badly if the restart list ever gets long, and which the transformer and switchgear shortage has already made visible across the sector.
That scarcity is the thing to hold. Restarts are the fastest and cheapest gigawatts available and there are almost none of them left. The same structure appears in every market where an existing, permitted, already-built asset commands a premium over a new one because the permitting is the scarce input rather than the capital, from pipeline rights-of-way to mineral leases in jurisdictions that no longer issue them.
Contracting with plants that are already running
The second pathway is larger in volume and less interesting to write about, which is why it gets less coverage than it deserves.
Talen Energy sold a 960-megawatt data center campus adjacent to its Susquehanna nuclear plant to Amazon Web Services for $650 million in March 2024, with a co-located power arrangement. In June 2025, after regulatory complications discussed below, Talen entered a front-of-meter power purchase agreement with AWS for up to 1,920 megawatts, running through 2042, with delivery ramping from 840 to 1,200 megawatts in 2029 to 1,680 to 1,920 megawatts in 2032. Talen put the lifetime contract value at roughly eighteen billion dollars in an investor presentation accompanying the announcement.
Nothing new gets built in that transaction. Susquehanna’s two units are licensed through 2042 and 2044, and the AWS contract runs through 2042, which means the arrangement is sized to remaining license life rather than to any new construction. An existing reactor that was selling into the wholesale market now sells to one counterparty under a long contract, at a price that a data center operator will pay and a wholesale market would not.
Which raises the question the market monitor asked. PJM’s independent market monitor has warned that using existing plants to supply data centers has a significant effect on the market, and its president has argued that new data centers should be required to bring their own new generation. The logic is straightforward: if a gigawatt of existing carbon-free generation is redirected from the general grid to one customer, the general grid is a gigawatt short and has to replace it with something, and the something is currently gas. That substitution is the entire content of the additionality argument, and it operates regardless of anybody’s intentions.
That is the additionality problem, and it is the strongest criticism of data center nuclear power as currently practiced. It is also the criticism most likely to be written into policy, because unlike carbon accounting it has a measurable local consequence in the capacity price. A twenty-year contract with an operating reactor delivers carbon-free electricity to the signatory and does not add a carbon-free electron to the system. The uranium and fuel cycle sees no additional demand from a contract that redirects existing output.
Behind the meter, and the fight over who pays for the wires
The Talen-Amazon arrangement produced the most consequential regulatory fight in this subject, and the substance of it is about cost allocation rather than about nuclear power at all.
Behind-the-meter co-location means a large load sits at the plant and takes power directly, without the electricity passing through the transmission system. The attraction is speed: no interconnection queue, no transmission upgrades, no wait. The objection is that the arrangement still benefits from the grid, since a co-located load relies on the transmission system for backup when the plant trips, and on the system’s reliability services generally, while contributing little or nothing to the cost of maintaining it.
In November 2024, FERC rejected on a two-to-one vote an amended interconnection service agreement that would have expanded co-located load at Susquehanna from 300 to 480 megawatts, concluding that the parties had not demonstrated why the deviation from PJM’s standard tariff was justified. American Electric Power and Exelon had argued in opposition that the deal could shift as much as $140 million annually to ratepayers, and that co-located load should not operate as a free rider on a transmission system paid for by transmission customers. FERC declined to rehear the matter in April 2025.
Then in December 2025 FERC issued a unanimous order directing PJM to establish clear rules for co-location, creating new transmission service options including firm and non-firm contract demand and interim network service alongside the traditional front-of-meter arrangement, with compliance deadlines beginning January 2026.
What that order does and does not do is worth stating precisely. It replaces case-by-case rejection with a defined menu, which makes long-term contracts and project financing structurable. It does not eliminate the interconnection queue, does not resolve the operational complexity of running a plant and a data center as one integrated asset, and does not settle the state jurisdictional questions around retail sales. It changes how the risk gets sliced between the grid operator, the generator, and the load. It does not make the electricity appear faster.
Amazon’s eventual solution was to abandon the behind-the-meter structure entirely and sign a front-of-meter PPA that required no FERC approval, which is the practical lesson: when the novel structure is contested, the conventional structure is available and slower and works.
Small modular reactors, and the timeline problem
The announcements are numerous and the operating reactors are not, and separating those two facts is most of the analytical work here.
The pitch for small modular reactors is genuine. Factory fabrication rather than site construction, standardized designs permitting a single design certification across many units, smaller absolute capital outlay per unit, passive safety systems reducing the engineered safeguards required, and siting flexibility including on retired coal plant sites with existing interconnections. If those advantages materialize, SMRs address the structural problem that has made large nuclear construction in the West a serial financial disaster.
The status is that no commercial SMR is operating in the United States. Design certification is a multi-year regulatory process, and the specialty alloys and fuel forms some designs require have supply chains that do not yet exist at commercial scale. First-of-a-kind construction carries the cost overruns that first-of-a-kind construction always carries, and the modular cost advantage depends on volume that only arrives after the first units prove out. Announced commercial operation dates cluster around 2030 and later, which means the earliest units arrive after the demand they are being announced to serve.
The Vogtle precedent is the reference point everybody in this discussion has memorized. The two AP1000 units in Georgia were the first new commercial reactors completed in the United States in decades, they came in years late and billions over budget, and the contractor building them entered bankruptcy mid-construction. That experience is why the SMR pitch leads with factory fabrication and standardization, and it is also why lenders price new nuclear construction the way they do.
Oracle’s stated plan for a gigawatt campus backed by three SMRs is representative of the category: a real intention, a real capital commitment, and no reactor. Treat every SMR date the way the briefing treats any forecast, which is as a document with an author who wanted something, and note specifically that an announced 2030 reactor is doing work in a 2026 investor presentation regardless of whether it is ever built.
The honest assessment is that SMRs are a plausible answer to the 2035 problem and no answer at all to the 2027 problem, and that most of the load being announced today needs power before any of them will exist.
What actually gets built in the meantime
Since restarts are scarce and new nuclear is slow, the marginal electron for near-term data center demand comes from somewhere else, and it is worth naming plainly.
Gas turbines are the answer the market is producing. They can be permitted and built in a fraction of the time, the equipment is available with lead times measured in a few years rather than a decade, and several large campuses have gone forward with on-site generation, in some cases with aeroderivative turbines originally built for other purposes because the utility-scale units were sold out. The equipment lead times across the whole power sector have become their own constraint, and turbine order books now extend years out. That is the direct consequence of the queue: a company that cannot wait for interconnection builds its own generation, and the fastest self-build is combustion.
The consequence for the emissions arithmetic is the part that gets omitted from the nuclear coverage. A nuclear PPA covering a company’s reported consumption is compatible with the marginal grid response to that company’s load being gas, because the reactor was already running. Reported emissions and system emissions diverge, in exactly the way the supply-chain accounting for any traded commodity diverges depending on where the boundary is drawn, and both numbers are defensible under their own conventions. That gap between reported and system emissions is the same scope-boundary problem that governs water accounting at these facilities, and it resolves the same way, which is that the figure you get depends entirely on where somebody drew the line.
Storage and renewables are being contracted heavily and address a different part of the problem, alongside the magnet and motor supply chains that every wind installation depends on, and the minor metals and specialty materials inside the equipment carry constraints of their own that rarely surface in a procurement announcement. Data center load is close to constant, which makes intermittent generation a poor match without substantial firming, and the storage chemistries and grid-scale batteries that would provide that firming carry their own supply constraints. Geothermal has attracted real capital for the same reason nuclear has, since it is firm, carbon-free, and available around the clock where the resource exists, and the operating geothermal fields are a small resource in a small number of locations. The materials constraints on any large firm-generation buildout apply to turbines as much as to reactors. Enhanced geothermal changes that arithmetic if it works at scale, and it sits at roughly the same stage of demonstration as the small modular reactors, with the same relationship between announced dates and operating units.
The fuel question underneath it
A reactor restart requires fuel, and the fuel supply chain is a separate constraint that nuclear announcements rarely mention.
Reactor fuel requires uranium mining, conversion, enrichment, and fabrication, and each step has limited capacity concentrated in a small number of facilities. Western enrichment capacity has been the acute constraint since the imposition of restrictions on Russian supply, and the enrichment and fuel cycle position is a multi-year buildout rather than a switch.
Advanced reactor designs compound it, because many require high-assay low-enriched uranium enriched above the conventional five percent threshold, and commercial capacity for that material is limited. A reactor design that needs a fuel type nobody produces at scale has a supply problem in addition to a licensing problem and a construction problem.
None of that stops the restarts, which run on conventional fuel through established channels. It does constrain how fast the announced advanced-reactor fleet could arrive even if everything else went well, which is a useful discipline against timeline optimism. The enrichment capacity buildout is itself a multi-billion-dollar industrial program with permitting, and it has to be financed against demand that only materializes if the reactors get built, which is the same problem that has held back every attempt to stand up a Western processing industry from scratch, which is the standard chicken-and-egg problem of any new industrial supply chain trying to start from zero, and which the export-control episodes of the past few years demonstrated is not solvable on announcement timescales.
Who pays, and the cost-allocation fight underneath the technology
The regulatory battle over co-location is a proxy for a distributional question, and the numbers involved are large enough that it will not stay technical.
Analysis from Synapse Energy Economics projects PJM consumers paying an extra hundred billion dollars through 2033 as data center demand outruns available supply. The sixty-seven million people PJM serves absorbed an additional $9.4 billion in electricity costs during summer 2025, with a further $1.4 billion locked in for summer 2026.
Whether those increases are attributable to data centers specifically is contested and the mechanism is not. Capacity markets price scarcity, new large loads tighten scarcity, and prices clear higher for everyone in the zone. A data center does not have to be doing anything improper to raise the price its neighbours pay.
Which is why the additionality argument has policy force independent of its climate merits. If a large new load brings its own new generation, the system is not tightened and the price effect is muted. If it contracts with existing generation, the system is tightened and other customers absorb the difference through the capacity market. The market monitor’s proposal that new data centers supply their own new resources is an attempt to internalize that, and it is opposed by essentially everybody who would have to comply with it. The objection is not unreasonable either: requiring a new load to build new generation is an obligation no other class of customer faces, and setting the threshold means an aluminium smelter or a large industrial extraction operation with comparable draw either faces the same requirement or gets an exemption somebody has to justify.
The long history of large industrial loads negotiating rates with utilities is the relevant precedent, and it generally ends with a special tariff class, in the way that aluminium smelters and other large industrial loads negotiated theirs across the twentieth century, which is what several utilities have begun constructing for hyperscalers. AEP Ohio’s structured tariff for large loads is the reference case, and the terms being negotiated in those proceedings, covering minimum take obligations, contract duration, exit fees, and collateral, are effectively an attempt to make a data center behave like a creditworthy long-term industrial customer rather than a load that might leave when the hardware inside it becomes obsolete.
What a restart actually involves
The word restart implies flipping a switch and it involves nothing of the kind, which is why the number of candidate plants is smaller than the number of shut ones.
A reactor that entered decommissioning has been legally and physically reclassified. Fuel was moved, systems were drained and laid up, staff were dispersed, the operating license was amended to a possession-only status, and exemptions from operating requirements were granted precisely because the plant was not going to operate. Reversing that is not maintenance. It is an application to un-retire a licensed nuclear facility, and the NRC has never had a mature process for it because until Palisades nobody had asked.
The physical work is substantial. Steam generators, turbines, transformers, and instrumentation that sat idle for years require inspection and in many cases replacement. Constellation’s roughly $1.6 billion capital figure for Crane is not a reactivation fee; it is a refurbishment program. The long-lead electrical equipment involved competes for the same manufacturing slots as every transmission project and every new generator in the queue.
The staffing problem is the underrated one. Licensed reactor operators require years of training and NRC examination, and a plant that dispersed its operating crew has to rebuild one from a labor pool that has been shrinking for decades as the fleet aged. Crane being roughly eighty percent staffed with over five hundred people well ahead of fuel load is a measure of how much of the timeline is human rather than mechanical.
And the eligibility filter is narrow. A candidate needs an unexpired or renewable license, equipment that was laid up rather than scrapped and not cannibalized for components with their own long supply chains, an intact interconnection, a site not yet released, an owner willing to spend, and a counterparty willing to sign a twenty-year contract. Reactors that were shut for technical reasons, or whose components were sold, or whose sites were partially released, do not qualify. That is why the restart list is measured in single digits rather than dozens.
The claims that do not hold up
An audit, because this subject generates confident assertions on a schedule.
Data center nuclear power amounts to a nuclear renaissance overstates what is happening. Restarts of existing reactors and PPAs with operating plants are real and represent a change in who buys the output rather than an expansion of the fleet. A renaissance would require new construction at scale, which is announced and not underway.
Three Mile Island is being restarted is imprecise in a way that matters. Unit 1 is being restarted. Unit 2, which had the accident, is permanently shut and was never a candidate.
Nuclear PPAs make data centers carbon-free is true under the accounting convention and does not describe the system effect, because redirecting existing carbon-free output does not add any.
SMRs will solve the power problem is a claim about a technology with no United States commercial operating unit, on timelines that arrive after the demand.
Data centers are causing your electricity bill to rise is directionally supported and frequently overstated. Capacity prices rose, data centers contributed substantially to the demand that tightened them, and generation retirements, transmission costs, fuel prices, and weather also contributed.
Behind-the-meter arrangements are a loophole that has been closed misreads the December 2025 order, which created structured options rather than prohibiting the practice.
The interconnection queue is bureaucratic obstruction misses that a substantial share of queued projects are speculative, that studies exist because the physics requires them, and that roughly 57 gigawatts of PJM projects hold agreements and are stalled for reasons having nothing to do with the queue.
Data center nuclear power is too slow to matter is contradicted by the restarts, which are delivering hundreds of megawatts on a three-year timeline that no new-build technology approaches.
What data center nuclear power is actually telling us
Assemble it and data center nuclear power reads less as an energy transition and more as a scarcity auction for a specific asset class.
The scarce thing is not electricity, and it is not nuclear technology. It is firm, carbon-free, already-interconnected generating capacity, and the number of units in that category is fixed and small. Every restart and every long-term PPA removes one from the general pool and assigns it to a single buyer with the balance sheet to pay above market for speed.
That has three consequences worth carrying. The buyers with capital get power first, which is a market outcome and also a distributional one. The general system replaces what was reassigned with the fastest available substitute, which is gas. And the price of capacity rises for everybody sharing the zone, which shows up on bills belonging to people who did not participate in the auction.
None of that is a scandal, and treating it as one obscures the mechanism. It is what happens when a load with an unusual willingness to pay meets a supply that cannot expand on the timescale of the demand, and the pattern is identical wherever a scarce permitted asset meets a buyer in a hurry. The briefing on how AI data centers work runs the physics, the money, and the politics in sequence for exactly this reason: the cooling determined the power draw, the power draw met a queue, the queue produced the workaround, and the workaround has a cost allocation attached that lands on a rate case. Data center nuclear power is the middle term in that sequence rather than a story of its own.
A reactor on the Susquehanna that closed in 2019 because it could not compete against cheap gas is being restarted in 2027 because one customer will pay enough to make it worth doing. Nothing about the reactor changed. What changed is who is standing at the fence with a checkbook, and how badly they need the electricity before 2030.
Leave a Reply