Nuclear power for AI data centers is moving from long-range ambition into specific restart and contracting decisions. On October 16, 2026, the Associated Press reported on efforts to revive Palisades, Crane and Duane Arnold, with rising electricity demand and technology-company power purchases helping drive the projects. A restart plan, however, is not electricity already delivered to a campus.
For a developer or buyer, the central question is now two-sided: can power reach the compute on time, and can the compute reach its users and other infrastructure? A site can be rich in generation potential but poor in usable network connections. Nuclear supply, transmission, fiber, cooling and capital must be evaluated on one delivery schedule.
What reactor restarts change for AI infrastructure
Restarts deserve a separate lane from new large reactors and small modular reactors. They concern previously operating assets, but still require plant-specific refurbishment, approvals, operating readiness and commercial arrangements. An existing turbine hall does not establish that a closed plant can return, or when it will be ready.
The Department of Energy’s September 2026 Duane Arnold financing announcement describes a loan of up to $1.9 billion for the 615-megawatt Iowa plant. It also explicitly makes restart subject to required Nuclear Regulatory Commission licensing approvals. Financial close and regulatory permission are different milestones; neither should be described as proof that the facility is already generating.
DOE separately documents financing for the Crane restart project. Buyers should compare the project’s current regulatory and equipment milestones with the date their own first phase must energize. Do not convert a supplier’s target into a guaranteed data center opening date.
This is a bounded opportunity, not an unlimited inventory of reusable reactors. Each candidate has a different shutdown history, equipment condition, workforce requirement and development path. A useful procurement question is: which remaining milestone could still move our service date, who controls it, and what is our alternative if it slips? Put those responsibilities beside the allocation of data center growth costs, rather than treating public financing or a customer contract as the whole answer.
Cumulus and Susquehanna: adjacency does not eliminate the grid
The Cumulus-developed campus next to Susquehanna is useful because its commercial history shows how a power arrangement can evolve. Talen’s September 2025 quarterly filing records AWS’s March 2024 purchase of campus assets for $650 million and a June 2025 revision to the power agreement. The revised agreement contemplated up to 1,920 MW of front-of-the-meter supply through 2042. Those are contract terms, not a claim that the full quantity was delivered immediately.
In its June 2025 announcement, Talen described a transition in which Susquehanna supplies the PJM grid, Talen acts as retail supplier, and PPL handles transmission and delivery. The announcement tied the transition to transmission reconfiguration. The practical lesson is to identify the current delivery structure and its dependencies, not assume that physical adjacency means a permanently isolated power system.
Separate three questions in any similar proposal: who has a contractual claim on supply, which physical facilities deliver it, and who pays for network work or service changes? An answer to one does not settle the other two. The same discipline applies to promises of additional reactor output or future SMRs.
Discuss safety and public concerns precisely
The NRC’s Three Mile Island backgrounder identifies the 1979 partial meltdown as an accident at Unit 2. It reports no detectable health effects from the small radioactive releases and describes major changes to training, emergency planning and regulatory oversight. That finding concerns a particular accident; it is not evidence that all nuclear projects have zero risk.
Nor should it be used to dismiss Chernobyl. The World Health Organization documents serious radiation-related health consequences from that separate disaster. Reactor design, accident conditions, operating controls and local emergency arrangements matter. Communities are entitled to clear answers about safety, waste, water, costs and accountability, not a comparison that erases their concerns.
Firm power is valuable, but uptime is a system property
The appeal of nuclear supply is sustained output. DOE’s generation-capacity explainer reports a U.S. nuclear fleet capacity factor above 92% in 2024. Capacity factor measures output over time relative to maximum possible output; it is not a promise that a particular data center will never lose power. Refueling, maintenance, transmission and the campus’s own electrical architecture remain relevant.
Nuclear is not the only pathway worth comparing. DOE’s geothermal and data centers overview describes geothermal’s steady generation potential and applications in cooling. Its usefulness still depends on the site, resource, development work and delivery economics. Compare credible projects against the same service-date and resilience requirements rather than choosing a technology label first.
Power-rich / Network-poor diligence matrix
Use this planning framework to expose dependencies before committing to a site. It is a set of diligence questions, not a score derived from measured campus performance.
| Dimension | Evidence to request | Warning sign |
|---|---|---|
| Power delivery | Approved supply structure, energization milestones, outage responsibilities and bridge plan. | A generation announcement is being treated as an executable delivery date. |
| Fiber reach | Surveyed routes, construction scope, rights of way, service dates and physically diverse entrances. | Two supplier names are offered without proof of separate physical paths. |
| Interconnection | Available carriers, exchange and cloud access, capacity commitments and expansion terms. | A nearby fiber line is being treated as ready-to-use connectivity. |
| Workload fit | Measured latency and throughput needs for training, storage, replication and inference. | One generic latency claim is used for every workload. |
| Cooling and land | Site control, heat-rejection design, water assumptions, permits and phased build requirements. | Electrical capacity is presented as the only site constraint. |
| Capital and accountability | A funded dependency schedule with named owners and consequences for delay. | Power and network teams each assume the other will absorb the gap. |
Apply the matrix to actual bids. For network choices, compare data center interconnect solutions and the underlying interconnect options. Require route, capacity and service evidence before calling a location connected.
Bring power to the network, or the network to power?
Test both approaches before selecting land. Starting near established interconnection may simplify network procurement while leaving a difficult electricity-delivery problem. Starting near generation may improve the power proposition while adding fiber construction, carrier access and operational dependencies. Neither location earns a pass on the missing half.
For a proposed training campus, ask which data transfers can be scheduled and what bandwidth the complete workflow needs. For inference, test the response time experienced by the intended users, including the application and upstream services. These are workload-specific engineering questions; a distance estimate or a carrier map is not a substitute for a measured design.
A private fiber path also does not by itself establish sovereign or isolated operation. Review where data is stored and processed, who operates each service, administrative access and the dependencies that remain outside the campus. Give those requirements to the same team responsible for network design and commercial commitments.
The durable strategy balances power, land, cooling, fiber, interconnection, compute and capital. Percepture’s AI infrastructure strategy guide provides broader context, while the AI corridor discussion connects the campus decision to network geography. For infrastructure marketing teams, a data center marketing plan should explain those concrete dependencies instead of promising readiness from a reactor headline.
Can nuclear energy support an AI data center?
Yes, but the path matters. Existing nuclear generation may be relevant through utility supply or a commercial agreement. Uprates and new reactors require additional development work. SMRs remain a longer-horizon option whose value depends on licensing, financing, fuel, construction and site-specific execution. Nuclear and SMR power for AI data centers should therefore be evaluated as several distinct strategies, not one product.
The decision in brief
Start with the required service date
A power option that arrives after the planned compute capacity has little near-term value, even if its long-run economics look attractive.
Separate existing assets from proposed assets
Existing plants, uprates, new large reactors and SMRs carry different development paths. Combining them under nuclear and SMR power for AI data centers hides the differences buyers need to see.
Test the full delivery chain
Generation is only one part of the decision. Transmission, interconnection, land, permits, contracts, financing and operational responsibility can change the result.
Use a delivery question instead of a headline question
For buyers, nuclear and SMR power for AI data centers is not a simple yes-or-no choice. A developer may announce an agreement, a reactor company may announce a design, or a data center operator may name a future energy goal. None of those events alone establishes when power will reach a particular campus.
A sound review of nuclear and SMR power for AI data centers starts with a service date and works backward. Ask when the first phase must energize, how much capacity it needs, whether demand will rise in stages and what bridge supply is acceptable. Then test each energy option against that schedule.
Keep a long-term supply strategy separate from a near-term interconnection problem. A future reactor may still have strategic value. It should not be credited with solving an earlier power gap unless the development path supports that conclusion.
The U.S. Department of Energy identifies both construction time and first-of-a-kind expense among the challenges associated with nuclear-powered data centers. Its overview is a useful starting point, but each project still needs its own commercial, regulatory and site review: DOE: Advantages and Challenges of Nuclear-Powered Data Centers.
Evaluate four different nuclear pathways
The phrase nuclear and SMR power for AI data centers can refer to several arrangements. Buyers should identify the actual pathway before comparing claims, schedules or prices.
1. Supply connected to existing nuclear generation
An operating nuclear plant is different from a proposed reactor. A data center may explore utility service, a power purchase agreement or another commercial structure associated with existing generation. The practical review still has to address deliverability, grid conditions, contract terms and whether the claimed supply aligns with the data center’s load profile.
Existing generation can make nuclear and SMR power for AI data centers feel like a near-term category, but the reactor’s operating status does not make every surrounding transmission or commercial issue immediate. The buyer should obtain a dated map of approvals, agreements and infrastructure work rather than relying on the age of the generating asset.
2. Uprates and changes at existing sites
An uprate seeks more output from an operating plant. It may avoid some elements of a new greenfield project, yet it is not merely a paper transaction. The buyer should examine the proposed scope, regulatory path, equipment work, outage coordination, transmission requirements and allocation of the added output.
For nuclear and SMR power for AI data centers, an uprate belongs in its own schedule lane. It should not be grouped with power already available, and it should not be judged by the same assumptions as a new reactor design.
3. New large nuclear reactors
A new large reactor can represent a substantial long-term generation strategy. It also creates a wide decision surface: site control, licensing, engineering, construction, financing, fuel, grid connection and operating responsibility.
That makes new large reactors a strategic rather than shorthand answer to nuclear and SMR power for AI data centers. A credible proposal should show who controls each development step, which milestones have been completed and what happens to the data center schedule if one milestone moves.
4. Small modular reactors
An SMR is a smaller reactor concept intended to use modular design and deployment approaches. The label does not remove project-specific questions. Buyers still need to examine the design’s regulatory status, supply chain, fuel plan, site requirements, financing structure, construction plan and operating model.
SMRs may become part of nuclear and SMR power for AI data centers, but a vendor target should not be converted into a buyer commitment date without a milestone review. Use the U.S. Nuclear Regulatory Commission’s public material to identify the applicable regulatory path and records for the design or project under consideration: U.S. Nuclear Regulatory Commission.
Scenario matrix for data center buyers
The clearest way to compare nuclear and SMR power for AI data centers is to separate asset status, commercial access and delivery dependencies.
| Pathway | First buyer question | Key dependencies | Planning role |
|---|---|---|---|
| Existing nuclear supply | Can contracted power be delivered to this load? | Utility structure, transmission, interconnection and contract terms | Potential near- or medium-horizon supply option |
| Uprate | What work and approvals precede added output? | Regulatory review, equipment, outage planning and grid delivery | Incremental capacity tied to an operating site |
| New large reactor | Who owns the complete development schedule? | Site, licensing, engineering, construction, fuel and financing | Long-horizon generation strategy |
| SMR project | Which milestones are complete rather than forecast? | Design status, licensing, supply chain, fuel, capital and operations | Future option requiring milestone-based review |
What could help before 2030, and what belongs later?
A useful timeline for nuclear and SMR power for AI data centers should use evidence gates rather than a single forecast. Use the date labels below as planning buckets rather than promises.
Near-horizon planning
For nuclear and SMR power for AI data centers, focus first on operating generation, available utility service, executable contracts and transmission or interconnection work that can be mapped to the campus schedule. Record which party controls each dependency and whether an alternate supply path exists.
Pre-2030 development cases
Place an option in this bucket only when project records support the required sequence. That sequence may include commercial agreements, regulatory activity, engineering, equipment procurement, construction and grid work. A press release without a complete dependency map is not enough for a capital plan.
Later-horizon options
New reactors and less mature deployment plans may still deserve land, partnership or portfolio analysis. Use milestone gates when reviewing those options. Do not let an attractive long-range concept replace the bridge-power plan needed for earlier phases.
Continuous review
Schedules change. Set review points tied to regulatory actions, contract execution, financing, major equipment and construction progress. The decision should become more committed only as the evidence becomes more concrete.
Put financing beside engineering
The economics of nuclear and SMR power for AI data centers cannot be assessed from an energy-price headline alone. The buyer needs to understand who funds development, who carries delay or performance exposure, when payments begin and what obligations survive if the data center schedule changes.
At minimum, the commercial team should ask:
- Is the arrangement utility service, a power purchase agreement, a development investment or another structure?
- Which party pays for transmission, interconnection and site-specific infrastructure?
- Are payments linked to milestones, availability, delivered energy or reserved capacity?
- What security, credit support or long-term commitment is expected?
- Who bears the cost if the generation project and data center campus move on different schedules?
- Does the structure preserve a practical bridge or exit path?
These questions belong beside the broader capital plan. Percepture’s guide to data center financing structures provides additional context for matching funding sources and project risks.
Check the site, not just the reactor
A plan for nuclear and SMR power for AI data centers can look strong while the site remains unready. Review land control, water strategy, transmission access, interconnection, construction logistics, local approvals, emergency planning responsibilities and the relationship between generation and campus phasing.
The review should also identify operating boundaries. Determine who owns the generating asset, who operates it, who supplies the data center, who manages outages and what backup architecture supports the load. These are separate questions, and the answer may involve several organizations.
Regional electricity information can help frame the market context, but it does not replace project records. The U.S. Energy Information Administration provides public energy data that teams can use as an input to regional research: U.S. Energy Information Administration.
Keep milestone language distinct from commitments
Marketing leaders reviewing nuclear and SMR power for AI data centers can keep public language aligned with the project’s stated maturity. A memorandum, application, site study, contract, construction start and operating asset should be described as distinct events.
Build communications around three layers:
- Completed fact: what has occurred and where the supporting public record can be found.
- Current plan: what the parties intend to do, including the dependencies that remain.
- Forward scenario: what may become possible if defined milestones are completed.
This structure gives executives a way to discuss strategy without presenting a forecast as delivered capacity. It also gives communications, finance, legal, development and operations teams a shared vocabulary for reviewing announcements.
For organizations managing a long buying cycle across investors, communities, partners and customers, an omnichannel marketing strategy can coordinate the message while keeping each audience’s information needs distinct.
Buyer readiness scorecard
- Load: Required capacity and energization phases are documented.
- Pathway: The proposal identifies an operating plant, uprate, new large reactor or SMR.
- Milestones: Completed actions are separated from targets and assumptions.
- Delivery: Transmission and interconnection dependencies have named owners.
- Regulation: The applicable licensing path and public records are identified.
- Commercial terms: Payment triggers, commitments and risk allocation are understood.
- Site: Land, infrastructure, approvals and operating boundaries have been reviewed.
- Fallback: Bridge power and schedule-change options are documented.
- Communications: Public claims match the maturity of the underlying project.
Use a weak result in any category to identify the next question that must be answered before treating the power plan as committed.
Make the decision in stages
A disciplined buyer does not have to choose between dismissing nuclear energy and accepting every forecast. Start with the campus service date. Classify the nuclear pathway. Map the regulatory, physical and commercial dependencies. Then increase commitment as evidence clears each gate.
A staged review of nuclear and SMR power for AI data centers protects the near-term data center schedule while preserving long-term options. It also gives executives a more accurate way to compare existing generation, uprates, new large reactors and SMRs without flattening them into one headline.
The final decision is project-specific. It should be based on dated records, accountable owners, executable contracts and a fallback plan—not on enthusiasm or skepticism alone.
About Bob Generale
Bob Generale is President of Percepture, where his role includes executive strategy. For technical infrastructure coverage, he approaches the subject as a strategist and editor rather than presenting himself as a nuclear engineer.
This guide is designed to help business and marketing leaders ask better questions, separate project milestones from promotional language and coordinate complex infrastructure communications. Connect with Bob on LinkedIn.
