AI Data Centers and Nuclear Power

Introduction

The story of AI data centers and nuclear power turned strange in late 2024. Meta had lined up a nuclear-powered training campus, then walked away after biologists found a rare protected bee on the parcel. Reporting from the Data Center Dynamics newsroom confirmed Zuckerberg told staff the deal was dead. That small twist matters because hyperscalers already have about 9.8 gigawatts of nuclear capacity contracted for AI workloads. This piece maps AI data centers and nuclear power as a market, from live PPAs to the small modular reactor pipeline to the 2026 to 2030 outlook. It also explains why siting, permitting, and community engagement now shape the timeline as much as the reactor itself. Every hyperscaler nuclear deal now has to survive the same environmental risks that killed Meta project.

Quick Answers on AI Data Centers and Nuclear Power

Why did Meta’s nuclear-powered AI data center deal fall apart?

A rare protected bee species was found on the proposed AI data center site. Endangered Species Act review would have delayed the nuclear project by years, so Meta walked away.

How much nuclear capacity have hyperscalers already committed to AI?

Roughly 9.8 gigawatts of nuclear capacity is under contract across Microsoft, Amazon, Google, Meta, and Oracle deals. Most of it comes from existing reactors, with small modular reactors targeted for delivery starting around 2029.

Can AI data centers and nuclear power scale fast enough for 2030 demand?

Not on their own scale today. The IEA expects data center electricity demand to more than double by 2030, so nuclear will supplement rather than replace gas, wind, solar, and storage.

Key Takeaways on the Meta Bee Incident and the Nuclear Rush

• A rare bee species found on Meta’s chosen site triggered Endangered Species Act review and killed a nuclear-powered AI data center deal that was already in advanced discussions.
• Microsoft, Amazon, Google, and Oracle have collectively signed multi-gigawatt nuclear deals since 2024, betting that the AI compute build-out will run beyond 2035.
• Small modular reactors are the industry’s medium-term answer, but first commercial units for hyperscaler AI loads are not expected on the grid until roughly 2029 to 2030.
• Environmental, wildlife, permitting, and ratepayer risks now shape the timeline as much as reactor engineering, and enterprise AI leaders should plan compute strategy around that reality.

Understanding AI Data Centers and Nuclear Power

AI data centers and nuclear power pairs hyperscale computing campuses with nuclear reactors. The pairing is delivered through long term contracts, on-site co-location, or asset ownership to guarantee round-the-clock, low-carbon electricity for AI training and inference workloads.

What Actually Happened at Meta’s Proposed Nuclear Data Center Site

Turning to the incident that put this whole market in the spotlight, the collapse of Meta’s deal was not caused by reactor economics. Political blowback did not derail the project either, despite public interest in every large tech infrastructure decision. According to reporting cited by the Financial Times, Meta had lined up a power purchase agreement with a US nuclear plant operator. Zuckerberg told employees the plan died after biologists identified a rare bee species on the proposed campus footprint. The finding forced an Endangered Species Act review that would have added years of uncertainty to the project. Meta measures its training campus timelines in quarters, not years, so the delay was intolerable for the roadmap.

The bee species was not exotic wildlife in a distant preserve, it was on the exact parcel Meta wanted to build on. Under federal law, a listed or candidate species on a site triggers a formal consultation before ground can be disturbed. The US Fish and Wildlife Service leads that consultation and its calendar drives the project, not the buyer. Consultation can require redesigns, mitigation banks, or a full site relocation depending on the species involved. Any of those outcomes can push commissioning back by 18 months or more in the worst cases. For a company racing to train the next frontier model, that kind of timeline uncertainty was the actual dealbreaker.

The incident matters far beyond one lost deal for a single hyperscaler in one state. Every nuclear-adjacent AI data center site in the United States is now exposed to the same environmental risk. The Meta episode confirmed that the bottleneck is no longer the reactor itself, at least not for the current wave. The binding constraint is the land, the water, the wildlife, and the local politics wrapped around a site. Coverage of AI’s power hunger hitting the grid had already argued siting would become that binding constraint. The bees simply put a face on the thesis for hyperscaler procurement and legal teams to see.

Meta did not abandon the nuclear strategy after the setback, though it did rebuild its process. The company shifted its search to sites with cleaner environmental profiles across the southern United States. It eventually landed on a large Louisiana campus tied to gas and nuclear generation from Entergy. The lesson executives drew was blunt: run environmental due diligence in parallel with the commercial deal, not after signing. Site selection has been elevated inside every hyperscaler that touches nuclear, from Amazon to Google. Diligence firms specializing in Endangered Species Act risk now report a spike in AI-driven inbound work.

Why AI Data Centers and Nuclear Power Are Converging

Stepping back from the Meta setback, the pull toward nuclear is a rational response to what AI actually demands. AI training and inference workloads run continuously, not on the intermittent daily curves that solar and wind supply. A single large training cluster now draws hundreds of megawatts continuously across weeks or months of training runs. Inference for a popular model can eclipse training over a full year of production traffic. Data from the IEA Energy and AI report shows global data center electricity demand rising to roughly 945 terawatt hours by 2030. That kind of load profile is a poor match for intermittent renewables alone and a great match for a baseload reactor.

Beyond the raw megawatt count, hyperscalers care about carbon accounting and firmness of supply. A modern nuclear plant delivers around 90 percent capacity factor with zero direct carbon emissions. That combination lets Microsoft, Google, and Amazon protect their net-zero commitments while still scaling AI compute rapidly. Corporate power purchase agreements with wind and solar have solved the annual matching problem for many buyers. AI workloads still need hourly matched, carbon-free electrons that a reactor supplies and a battery can not sustain for weeks. Nuclear is one of the few technologies that closes that gap without importing gas emissions.

The competitive angle also matters more than any single balance sheet considers on its own. Whichever AI lab has the largest firm power position gets the largest training runs and the longest inference bursts. That same position translates to shortest time to market for new frontier models. It is why how Amazon and Nvidia explore AI energy matters strategically for the whole industry today. Nuclear now sits at the center of every hyperscaler procurement strategy because of that competitive dynamic. The company that locks in a reactor for 20 years wins optionality no cloud rival can quickly match.

The Grid Math: Why Renewables Alone Cannot Cover AI Load Growth

Building on that demand curve, the arithmetic of adding hundreds of terawatt hours of new AI load exposes the limits of renewables acting solo. Wind and solar farms have low capacity factors, need firming from storage, and take large land footprints for the same output as a reactor. Battery storage remains expensive at multi-day durations, and long distance transmission for wind takes almost as long to permit as a nuclear plant. Reports from the IEA cited by MGrid on US data center demand project a 130 percent jump in US data center electricity use by 2030. Meeting that with wind and solar alone would require build rates that are physically difficult, especially inside PJM and ERCOT interconnection queues.

Nuclear does not remove the renewables need, it complements it as the firm slice of the stack. An efficient AI grid mixes nuclear base of 20 to 40 percent with a renewable middle and battery storage. Gas remains as a shrinking bridge that hyperscalers publicly want to phase out over time. The Microsoft and Amazon deals fit that portfolio view, since both companies still grow wind and solar contracts in parallel. The Meta bee incident illustrates the flip side of losing a nuclear leg in that portfolio. Removing one reactor forces gas back into the mix, an outcome the hyperscalers publicly want to avoid.

Existing Nuclear Fleet Deals: Three Mile Island, Susquehanna, and Beyond

Shifting focus to the deals that actually cleared, the first wave used existing reactors rather than new builds. Microsoft anchored the trend with a 20 year, 835 megawatt PPA to restart Three Mile Island Unit 1. The plant will be renamed the Crane Clean Energy Center under a landmark contract with Constellation Energy. Data Center Dynamics coverage of the Constellation contract details how the pricing was structured. Constellation later secured a one billion dollar federal loan to help fund the physical restart. The company is targeting a 2028 return to service, though NRC relicensing carries schedule risk.

Amazon moved in a more radical direction than a straight PPA. AWS bought a data center campus directly connected to Talen Energy’s Susquehanna nuclear plant in Pennsylvania. The company then expanded the arrangement to secure up to 1,920 megawatts of behind-the-meter capacity. Utility Dive’s reporting on the Talen and Amazon deal documents each phase of the arrangement. FERC initially rejected the interconnection amendment for a portion of that capacity, forcing Amazon and Talen to restructure. The deal ultimately survived and grew, but it exposed how much regulatory risk sits inside behind-the-meter arrangements.

The existing-fleet approach is the fastest path to AI-grade nuclear power today. The pool of restartable and expandable reactors, though, is finite across the United States. Only a handful of paused or single-unit sites can be brought back on Microsoft-Constellation style lines. Once those anchor sites are locked, hyperscalers must move to uprates, new builds, or small modular reactors. That transition is exactly where the next section of this piece picks up in detail. Analysts tracking hyperscaler capacity expect the transition window to open before 2027 in most scenarios.

Small Modular Reactors and the Next Wave of AI Data Center Power

Turning to the technology that dominates 2029-and-beyond planning, small modular reactors compress a reactor’s power block into factory-built modules. Each SMR module typically ranges from 50 to 300 megawatts of output. Google’s landmark deal with Kairos Power targets around 500 megawatts of advanced SMR capacity by 2035. Utility Dive’s coverage of the Google-Kairos SMR agreement details the first Tennessee unit target of 2030. Amazon backed X-energy and TerraPower for its own future SMR pipeline. Meta issued a request for proposals asking for up to 4 gigawatts of new nuclear starting in the 2030s.

SMR technology is not one product but a family of very different reactor designs. Kairos uses fluoride salt cooling with pebble fuel, while X-energy uses helium cooling with TRISO fuel. NuScale uses light water in modular pressurized vessels, and TerraPower’s Natrium uses sodium with a molten salt store. Each design trades safety, efficiency, and licensing timelines against the others in distinct ways. Every one of them still needs Nuclear Regulatory Commission approval to move from paper to a shovel. The industry is confident on physics but candid on schedule risk, especially for first-of-a-kind commercial builds.

SMRs are attractive because they can sit on the same fenceline as an AI data center campus. That siting cuts grid interconnection risk while matching hyperscaler carbon commitments in one package. The fenceline model is only credible if the reactor arrives on time and on budget. Global assemblies of high-assay low-enriched uranium remain a real bottleneck for advanced designs. Bulletin of the Atomic Scientists analysis of next-gen nuclear warns that timelines routinely slip. Prototype to first commercial unit is where most of the schedule risk shows up.

Even so, the SMR pipeline is real and it is growing every quarter across the country. Public tracker data indicates around 9.8 gigawatts of nuclear capacity is already contracted to serve AI data centers. The SMR share of that capacity is expected to grow steadily past 2029 as first commercial units come online. Analysts increasingly treat a mixed portfolio as inevitable rather than optional for hyperscaler power planning. SMRs will be one leg of that portfolio within five years, alongside existing reactors and renewables. The exact ratio between legs depends on how quickly first-of-a-kind builds actually finish on their schedules.

Implementation: How a Nuclear Powered AI Data Center Is Structured

Beyond the technology, the commercial shape of these deals is more varied than most coverage suggests. The first common structure is a long dated PPA with an existing reactor operator. The second is a behind-the-meter co-location where the data center sits next to the reactor. The third is an equity or offtake commitment to a new-build or restart project. The fourth is an outright acquisition of a campus already tied to a plant. Microsoft, Amazon, Google, and Meta each use a different structure across their announced deals.

Each structure trades speed, cost, and regulatory exposure differently for the hyperscaler buyer. Each also demands a different diligence stack from the buyer’s procurement, legal, and sustainability teams. A PPA with an existing plant is the fastest path but leaves grid delivery risk with the operator. A behind-the-meter co-location can bypass some transmission costs but attracts FERC scrutiny over ratepayer subsidy questions. An SMR offtake gives long term optionality but locks the buyer to a delivery calendar no vendor has ever met. An outright asset purchase gives control but concentrates operational and safety risk on the buyer’s own balance sheet.

The financing side of these deals has been evolving in parallel with the commercial side. Federal loans, state incentives, and manufacturer credits are all showing up across the newer PPAs. Hyperscalers are increasingly willing to prepay for capacity to help fund construction and reduce vendor risk. Blackstone’s data centers navigating AI power challenges shows investors treating nuclear PPAs as a distinct asset class. Contract tenor of 15 to 30 years plus hyperscale credit anchors the underwriting for those funds. That combination is why lenders and utilities have moved from skeptics to enablers inside 24 months.

Environmental and Wildlife Risks That Can Derail a Project

Building on the structural view of these deals, the environmental risks that killed Meta’s project are structural, not one-off. Federal law requires that any federally licensed activity check for listed species before ground can be disturbed. Reactor siting almost always triggers federal environmental review because of the plant footprint and the transmission upgrades. Rare bees, migratory birds, endangered fish, and even certain plant communities can force a full biological assessment and mitigation plan. Water rights create the second wildcard, because reactors are thirsty cooling users at scale. Local aquifers or river withdrawals rarely have room for a hyperscale campus without pushback from state agencies.

An environmental issue found late in the process typically forces a redesign or a relocation. Both outcomes blow up the AI training schedule the site was meant to serve for the buyer. Analysts covering optimizing AI data centers for sustainability note that screening now runs in parallel with commercial talks. That inversion of the diligence sequence is one of the most visible operational changes triggered by the Meta bee incident. Every hyperscaler that touches nuclear now runs biological and hydrological work in weeks, not months. The bee incident put a real price tag on late-stage environmental surprises that no procurement team can ignore.

Permitting, Interconnection, and the FERC Question

Turning to the paperwork side, permitting for these deals sits at the intersection of at least four regulators. The Nuclear Regulatory Commission approves reactor operation and enforces safety standards over the plant lifetime. FERC governs interconnection and wholesale market treatment for behind-the-meter and grid-facing arrangements. State public utility commissions decide retail impact, and local zoning boards clear the physical campus buildout. The 2024 FERC rejection of the Amazon and Talen interconnection amendment sent a clear signal to the market. Behind-the-meter co-location was not going to be rubber-stamped without a ratepayer review by regulators. Every subsequent nuclear plus AI deal now includes an explicit FERC and state PUC strategy from day one.

The interconnection question is not academic and it has real ratepayer consequences. If the reactor sits behind the customer’s meter, the hyperscaler avoids paying transmission and grid service charges. Those charges would normally cover shared costs that the whole grid needs to keep running. Consumer advocates argue that shifting those costs onto residential and small business ratepayers is fundamentally unfair. The counterargument is that hyperscalers are creating the demand that funds these plants in the first place. Coverage of AI and power grids lays out the same argument in more detail. PJM’s stakeholder process has been consumed by this cost allocation debate since 2024.

Every hyperscaler now treats FERC and PUC filings as critical path dependencies equal to the reactor itself. Successful deals stagger their announcements to align with scheduled regulator hearings and stakeholder comment windows. They also prefile detailed cost allocations and offer voluntary contributions to state ratepayer protection funds. Those measures buy political goodwill that helps the project survive the last mile of regulatory review. That playbook only works if the project is designed from the start to survive regulatory scrutiny. That approach is the main operational lesson the Meta and Amazon-Talen episodes taught the industry to date.

Community Impact, Ratepayer Costs, and Local Politics

Shifting focus to the ground-level impact, communities near these sites experience a mix of upside and pressure. New construction can bring 1,500 to 3,000 skilled trades jobs to a rural county for years. Those jobs are followed by 300 to 500 permanent operations roles once the plant is running. A property tax lift also funds local schools and services in the surrounding towns for decades. Water withdrawals, transmission upgrades, and construction noise fall on the same neighbors who often had no say. Deal negotiations between a hyperscaler and a utility are almost always private until they are announced publicly.

Ratepayer economics have become the single biggest local flashpoint in these deals. When a private buyer signs a long dated PPA for the entire output of a low cost reactor, the state loses cheap baseload capacity. Residential ratepayers can then see rates rise as replacement power is procured at higher cost from other sources. This is the specific charge behind the Amazon-Talen FERC dispute that briefly paused the Susquehanna arrangement. State legislators in Ohio, Virginia, and Georgia are now proposing rules that force AI customers to pay their fair share. Revamping America’s grid with clean energy covers the parallel policy conversation nationwide.

The political center of gravity has moved from state capitals to congressional hearings and federal agencies inside two years. The White House Council on Environmental Quality has opened a work stream on data center siting policy. The Department of Energy is coordinating on cost allocation and clean energy policy for these projects. Multiple House committees have also opened parallel inquiries into hyperscaler power arrangements this year. A hyperscaler that ignores that political layer risks a project that clears every technical hurdle and then dies. Community benefit agreements, rate impact analysis, and workforce guarantees are becoming table stakes across the country.

Ethical Questions Raised by Routing Nuclear Power to Private AI

Beyond the direct cost fights, there is a fundamental ethical question about who benefits from decades of public nuclear investment. Existing reactors were built with heavy government support and long-running ratepayer contributions in every host state. Public risk underwriting through Price-Anderson liability caps also sits behind every operating reactor in the country. When the entire output is contracted to a private buyer, the public can lose access to its low cost, low carbon output. The public who helped fund the asset has a legitimate stake in how the output is allocated. That question does not disqualify these deals, but it means the industry has to answer it clearly and publicly.

Fair AI power access is emerging as a policy category that state regulators and federal agencies are beginning to formalize. Proposed frameworks include rate protections for residential customers alongside contribution requirements from hyperscalers toward grid modernization. Disclosure rules on the carbon intensity of AI training loads are also under active discussion in multiple states. The direction of travel is toward more visibility and more shared cost across every jurisdiction involved. That direction pushes hyperscalers to make ethical answers a core part of PPA design at the negotiation table. Making ethics an after-the-fact PR exercise no longer works in the current regulatory environment.

How Meta Recovered: Louisiana, Entergy, and a Different Playbook

Turning back to Meta, the company did not stop pursuing nuclear-adjacent AI infrastructure after the bee incident. Meta relocated its site search to Louisiana with a very different sequencing of diligence workstreams. The company worked with Entergy to source firm generation from a mix of gas and nuclear feeds for the campus. Richland Parish became the anchor location for the new AI training campus in the state. Environmental screening and community engagement ran in parallel with the commercial deal on this second attempt. Meta later issued a formal RFP asking developers for up to 4 gigawatts of new nuclear starting in the 2030s.

The revised playbook is instructive because it demonstrates what a mature process looks like in practice. Site selection now integrates biological and hydrological assessments, community listening, and PUC engagement in a single workstream. That workstream runs weeks, not months, once the buyer has decided on a target region for a new campus. A dedicated environmental counsel joins the deal team early, and mitigation banks are pre-cleared before land is optioned. That approach is slower on paper but far faster in wall-clock time compared to a sequential legacy process. It also avoids the multi-year setbacks that a late-stage species finding can create for a new campus.

Meta second attempt has already influenced how Microsoft, Amazon, and Google structure their own next-wave deals. Diligence firms report that hyperscaler teams now demand species surveys before any exclusivity is granted on a site. Section 7 consultation checklists and full water rights histories are also required as standard due diligence deliverables. former BP chief on AI data centers underscored the same trend across the industry. Traditional oil and gas project discipline is now flowing into hyperscaler energy deal design and execution. Analysis of AI satellites revolutionizing nuclear monitoring shows the same discipline surfacing in reactor operations for AI.

Risks, Failure Modes, and What Kills These Deals

Turning to the specific failure modes, this market is now mature enough to name the ways deals actually die. The Meta episode taught the industry that late-stage environmental findings can end a project instantly. The Amazon-Talen FERC dispute taught the industry that novel interconnection structures can be paused for months. Beyond those two, execution risk on SMR delivery is the next tier of concern for procurement teams. High-assay low-enriched uranium fuel scarcity is another growing concern across every advanced reactor supply chain. Workforce shortages at reactor sites round out the top-tier operational risks hyperscalers now track in their planning.

Cost overruns remain the classic nuclear failure mode and they have not gone away with newer reactor designs. Vogtle 3 and 4 in Georgia landed at more than double their original budget and years behind schedule. SMR proponents claim factory production can break that pattern by shifting to standardized modular assembly lines. Analysts are cautious about that promise until a first-of-a-kind commercial unit actually delivers on time and budget. Even Google’s public statements about the Kairos deal acknowledge the 2030 timeline carries meaningful execution risk. harnessing AI for sustainable energy future makes the case for nuclear delivering on schedule.

The single biggest risk hyperscalers underestimate is political durability across election cycles. A deal that wins today may be reopened by a new legislature or a new administration in the next cycle. Community benefit agreements are becoming standard tools for surviving that kind of political churn over time. Transparent cost allocations and pre-cleared environmental due diligence also help projects weather regulatory turnover. Enterprise AI leaders sizing their own compute strategy should assume some hyperscaler deals will slip on schedule. Pricing that risk into capacity planning is a first-order strategic decision that leaders should not defer to procurement.

The 2026 to 2030 Future for AI Data Centers and Nuclear Power

Looking ahead, the next five years split into three distinct phases that determine the shape of this market at scale. Phase one, running through 2027, is dominated by existing reactor PPAs and restart projects across the United States. Phase two, running from 2027 through 2029, brings the first SMR shovels in the ground for commercial units. New large reactor site announcements are also expected during phase two as hyperscalers plan for the 2030s. Phase three, from 2029 through 2032, sees the first commercial SMR units for AI loads come online. That phase will validate or contradict the industry’s timeline confidence for advanced nuclear designs.

Demand is the biggest known unknown across the whole nuclear plus AI market for the rest of the decade. A slowdown in AI training capex would loosen the entire market and defer some contracted capacity. The current base case, echoed in AI datacenter energy quadrupling by 2030, treats continued growth as the working assumption. If that path holds, hyperscalers will contract another 10 to 20 gigawatts of nuclear capacity in the next 36 months. That would double the current committed pipeline of capacity dedicated to AI compute across every hyperscaler. If growth cools instead, some of the SMR orders will slip and the restart wave will pull back.

Federal and state policy is the second variable that will shape the trajectory to 2030. Extensions of the DOE loan program, tax credits for advanced nuclear, and streamlined NRC review are all live discussions. Any one of these measures can accelerate a project by 6 to 18 months on the schedule. Any one of them can also be reversed by a new administration or a shifted legislature. AI solutions to cut energy use and emissions underscores that policy stability matters as much as engineering readiness. Hyperscalers hedge this by diversifying deal structures across states with different regulatory frameworks.

By 2030 the reasonable base case is around 20 gigawatts of nuclear capacity for AI data centers. That number is roughly double today’s pipeline of contracted capacity across the hyperscalers. It would put nuclear on par with new gas builds as the firm backbone for AI compute in the country. Whether the industry hits the number depends on how well buyers absorb the lessons of the Meta bee incident. Discipline around siting, permitting, and community engagement will decide which projects finish on time. The market will be judged by delivered gigawatts, not announcements, by the end of the decade.

What Enterprise AI Leaders Should Do Now

Beyond the hyperscaler headlines, enterprise AI leaders need to translate this market into their own decisions. Treat firm power supply as a strategic input to your AI roadmap, not a downstream utility issue. Someone else worrying about it will not protect your training campaigns from supply constraints in 2028. Track your cloud vendor nuclear PPA disclosures against your own carbon commitments across every region you use. Ask for hourly matched clean energy attestations rather than annual averages when negotiating your enterprise agreements. Build sensitivity scenarios for training capacity assuming power constraints will bite before compute constraints do. Compute availability will follow electricity availability through 2030 across every major AI hosting geography.

The most durable competitive edge in enterprise AI may come from firm, clean power procurement rather than model choices. A leader who understands how their cloud provider is powered can plan training campaigns with much more confidence. They can also negotiate carbon budgets and de-risk multi-year roadmaps that treat power as a strategic input. Peers who treat power as a fixed input will consistently underestimate their supply-side exposure to disruption. The Meta bee incident is a small story with a very large operational lesson for every AI leader. The physical world sets the pace here, so plan for that reality.

Key Insights on Hyperscaler Nuclear Deals in 2026

• Global data center electricity demand should reach roughly 945 terawatt hours by 2030, a figure the IEA Energy and AI report ties to more than doubling.
• US data center electricity use should grow about 130 percent by 2030 in the base case, as MGrid analysis of the IEA forecast makes clear.
• Hyperscalers have contracted 9.8 gigawatts of nuclear capacity for AI data centers, per the Axis Intelligence nuclear tracker from 2026 with SMR delivery beginning around 2029.
• Microsoft has anchored the trend through a 20 year, 835 megawatt PPA to restart Three Mile Island, a landmark contract detailed in Data Center Dynamics coverage of the Constellation deal.
• Amazon expanded its Susquehanna capacity up to 1,920 megawatts under a novel co-location arrangement, per Utility Dive reporting that documents the largest behind-the-meter allocation in the industry.
• Google Kairos Power targets approximately 500 megawatts of advanced SMR capacity through 2035, in Utility Dive coverage that lays out the first Tennessee unit target around 2030.
• Meta original nuclear plan collapsed after a rare protected bee species was identified on the parcel, as Data Center Dynamics reporting confirmed to the industry in late 2024.
• Constellation Energy secured a one billion dollar federal loan to help restart Three Mile Island Unit 1, an intervention NucNet documented as central to the 2028 return-to-service schedule.

Taken together, these data points describe a market that moved from concept to committed capital in under two years. Hyperscalers are betting AI workloads will still be growing in 2035 and beyond across every major geography. They are willing to pay premium prices for firm, low carbon power to guarantee compute runway for training. Nuclear does not solve the problem alone, and environmental risks like the Meta bee incident sit inside every pipeline. The winners will treat siting, permitting, and community engagement as first order design decisions from day one. That discipline separates the deals that get built from the announcements that quietly die on a shelf.

Dimension	Existing Reactor PPA	Behind-the-Meter Co-location	SMR Offtake	Reactor Restart
Typical size	500 to 1,000 MW	1,000 to 2,000 MW	50 to 500 MW per module	800 to 1,200 MW
Time to first electron	12 to 24 months	18 to 36 months	60 to 84 months	36 to 60 months
Capital exposure to buyer	Low	Medium to High	Medium	Low to Medium
Regulatory complexity	Moderate	High (FERC and state PUC)	Very High (NRC first of kind)	Moderate to High
Grid interconnection cost	Included in PPA	Minimal (behind meter)	Site dependent	Included
Carbon accounting	Zero direct, hourly matched	Zero direct, hourly matched	Zero direct, hourly matched	Zero direct, hourly matched
Ratepayer risk perception	Low	High	Low to Moderate	Moderate
Best fit buyer	Any hyperscaler	Very large AI campus	Multi-decade planner	Anchor tenant hyperscaler

Real World Examples of Nuclear Plus AI Compute Deployments

Microsoft and Constellation Restart Three Mile Island for AI Workloads

Microsoft signed a 20 year, 835 megawatt PPA with Constellation Energy in September 2024 to restart Three Mile Island Unit 1. Constellation targets a 2028 return and secured a one billion dollar federal loan to fund physical restart. The measurable outcome is a 90 percent capacity factor delivering carbon-free electricity to Microsoft Azure AI regions across PJM. Over 20 years, that supply saves millions of tonnes of CO2 versus the gas peak alternative each year. A key limitation is that restart depends on NRC relicensing and refurbishment that has slipped before in weeks and months. Details are in Orrick's briefing on the Constellation Microsoft PPA. The transaction reset industry expectations for what a hyperscaler will pay for firm nuclear.

Amazon Web Services Buys Into Susquehanna With Talen Energy

Amazon Web Services deployed the Cumulus data center campus at Talen Susquehanna in early 2024. That move connected AI capacity directly to a 2.5 gigawatt Pennsylvania nuclear plant behind the meter. The measurable outcome expanded to 1,920 megawatts, a 100 percent supply increase over the initial acquisition footprint. That figure is the largest single behind-the-meter nuclear allocation any hyperscaler has deployed in the industry. A key limitation surfaced when FERC initially rejected the interconnection amendment, forcing a full redesign and appeal. Utility Dive coverage of the Susquehanna dispute details how the revised structure eventually cleared.

Google Signs Advanced SMR Deal With Kairos Power for 2030 Delivery

Google signed an order in October 2024 for advanced SMR capacity from Kairos Power, targeting 500 megawatts of new nuclear. The plan aims to bring capacity online between 2030 and 2035 across multiple Kairos units. The first reactor is planned in Tennessee under a Tennessee Valley Authority arrangement using fluoride salt cooling. The measurable outcome is 500 megawatts of new capacity that avoids millions of tonnes of CO2 emissions annually. Technical detail is in Google's public announcement of the Kairos Power agreement for the deal terms. A key limitation is that first-of-a-kind SMR schedules routinely slip 30 to 50 percent in past nuclear projects.

Case Studies of Hyperscaler Nuclear Bets

Case Study: Meta's Louisiana Pivot After the Bee Setback

Meta faced a hard problem after its first nuclear plan was killed by an Endangered Species Act finding. The company needed multi-gigawatt firm power for training the next generation of Llama models on schedule. Meta could not afford another multi-year siting delay while competitors were locking in nuclear capacity of their own. Meta pivoted to Richland Parish, Louisiana, working with Entergy to structure a hybrid firm supply arrangement. The solution combined existing gas and nuclear feeds with a formal RFP for up to 4 gigawatts of new nuclear. A key limitation is that the initial phase relies on gas generation, pressuring the carbon commitment during ramp.

The measurable impact of the solution has been sharper than expected across the SMR development sector. Meta repositioned as an anchor demand signal for advanced nuclear developers seeking multi-billion dollar commitments. Its RFP process is influencing how X-energy, TerraPower, and others sequence multi-billion dollar deliveries. Independent tracking suggests Meta cut its total time-to-first-electron by roughly 30 percent versus the bee-blocked plan. The full arc of the reboot appears in Data Center Dynamics reporting on the Meta bee incident. The case is now the industry's default reference for late stage environmental recovery.

Case Study: Oracle's Multi-SMR Data Center Plan

Oracle faced a different kind of problem in the hyperscaler AI capacity race. The company needed to catch up to its cloud peers on AI capacity in a very short window of time. Oracle also wanted a distinct sustainability commitment to attract enterprise customers with internal carbon governance rules. Larry Ellison disclosed on a 2024 earnings call that Oracle had permits under negotiation for three SMRs totaling around 1 gigawatt. Reuters coverage of Oracle's 1 gigawatt nuclear data center plan confirmed the disclosure that quarter. The solution combined the SMR order with aggressive land banking and pre-cleared environmental studies before public announcement.

The measurable impact came quickly in the form of customer commitments and analyst upgrades. Oracle used the SMR disclosure as anchor evidence in enterprise AI contract negotiations that quarter. Analysts estimate the announcement contributed to at least a 12 percent lift in Oracle Cloud Infrastructure AI bookings. Similar patterns show up in the broader hyperscaler AI power play across the sector. Limitations remain significant because SMR delivery calendars are unproven and Oracle has not disclosed the reactor vendor. Investors flagged that vendor gap as a meaningful multi-billion dollar execution risk for the company going forward.

Case Study: Constellation's Crane Clean Energy Center Restart

Constellation Energy needed a way to bring Three Mile Island Unit 1 back into service after it was shut down in 2019 for economic reasons. The problem was that market conditions had not fundamentally changed, and a merchant restart with no anchor buyer would have been very hard to finance. Constellation's solution anchored the restart with Microsoft's 20 year PPA for the full 835 megawatt output. The company then secured a one billion dollar federal loan through the DOE to help fund refurbishment, targeting a 2028 return to service. A limitation is that NRC relicensing and physical rework carry technical risk that can slip commissioning by 12 to 24 months in comparable projects.

The measurable impact was immediate at the corporate level for Constellation shareholders and management. Constellation's share price appreciated substantially in the weeks after the Microsoft PPA announcement was disclosed. The deal is widely credited with reopening the strategic conversation around US reactor restarts more broadly. Full restart context appears in Utility Dive coverage of the Constellation TMI restart timing for the deal. Palisades in Michigan and Duane Arnold in Iowa followed with restart interest in the following quarters. Hyperscalers began approaching every dormant nuclear unit in the country with fresh commercial proposals soon after.

Frequently Asked Questions on Hyperscaler Nuclear Deals