Opportunities and Challenges of Nuclear-powered AI Data Centers

Introduction

The recent development and the explosive growth of AI could prove to be the defining story of the 21^st century. AI is becoming used for everything that we buy, sell and do, as it enhances, accelerates and replaces simple and complex tasks across all sectors of the economy. This ubiquity is fueling the demand for huge new Data Centers. These Data Centers will require large quantities of power to be delivered very quickly - within a few years. Whilst most current Data Centers consume only a few 10-100 MW, the planned AI Data Centers, built with modules of 100-200 MWe, when complete can be rated at a GW, or more. The AI power demand problem has two parts: connecting to the Grid which is proving to be very slow and providing new dedicated clean power. 

The AI Hyperscalers – Meta, Microsoft, Google, Amazon etc.- are seeking over 100 GW of electrical power in the US alone by 2030 [Goldman Sachs (2025)], which will account for 50% of the power growth up to 2030 and will be more than 10% of total US demand in 2030 [IAE (2025)]. Beyond 2030, proposal for AI Data Center power in 2035 are up to 200 GWe. AI Data Center investments by 2030 are estimated to be $700 bn in US and globally $6.7 tr [McKinsey (2025)]. 

Many current AI Data Centers and most of those being planned are powered by fossil fuels, usually open cycle gas turbines because local Grid supplies are not available. The exceptions are planned connections to existing nuclear power plants that have been closed or are planned to be closed. For example, Microsoft have signed a 20-year Power Purchase Agreement with Constellation Energy to restart Three-Mile Island Unit 1 [Investing (2026)]. Amazon have acquired a Data Center campus directly adjacent to Talen Energy's Susquehanna nuclear plant in Pennsylvania with PPA for power excluding a connection to the Grid. [Ibid].  

Why is nuclear being considered for AI Data Centers?

There are over 400 power reactors in 30 countries, generating about 10% of world electricity. These reactors have over 18,000 reactor years of operating history. For many years the issues of public acceptability and the disposal of nuclear waste have dogged the industry. Now the positive features of nuclear power are being appreciated widely and being championed by the technology sector. These are:

• Unmatched Reliability: Nuclear plants boast capacity factors exceeding 90%,
• High Energy Density: Nuclear offers the highest power-to-land ratio, allowing operators to generate significant energy without requiring massive amounts of land for renewables plus their inevitable very large battery stores.
• Stable costs: Energy costs largely unaffected by changes in fuel cost.
• Decarbonization: In the US Hyperscalers need massive amounts of new electricity without violating corporate net-zero pledges. This issue is even more important outside the US.

Nuclear industry today

There is a need to recognize the reality of nuclear industry in the West after a 30-year gap in construction. A small number of new power projects have been completed using the latest designs which employ well-proven light-water reactor technology. All have fared badly. 

The two AP1000 Westinghouse reactors at Vogtle in Georgia were completed in 2024 after more than 14 years, $22bn (150%) over budget [Wikipedia Vogtle (2026]]. French EPR reactors (1,650 MW each) built in France, Finland and UK are even more over budget and many years late. There will be improvements in cost and schedule as others of the same designs are built. Also, similar designs built in China together with earlier GW-sized designs in Korea and UAE [Wikipedia Barakah (2026)] have shown that constructors with current know-how and experienced teams of technicians can deliver a mature large reactor at much lower levels of cost. There are new US efforts to build large reactors for AI, for example: Fermi America [WNN (2025a)] plans for an 11-GW Data Center Park in Amarillo, Texas that includes large nuclear plus gas power supplies. The US Government is supporting these types of initiative right now with $17.5bn funding for long-lead items [DOE (2026)] for large reactors – AP1000.

Nevertheless, at least in West where energy projects depend on private funding, the construction cost and scale of these large GW size reactors are: 

• Too big to fund – $10-15bn per GWe reactor.
• Too slow to construct for power market needs – 8-12 years.
• Too expensive to be competitive - $120-170 /MWh.
• Require too much skilled and experience site labour – 40,000 man-years per GWe.

How will this poor record of construction performance be changed? Replication of large reactor designs such as AP1000 and EPR should lead to some improvement. They are so large, so complicated. They are designed to be constructed almost wholly on-site, where productivity is low, such that improvements will be limited, making them not suitable for the AI Data Center requirement.  

Small Modular Reactors

Many smaller reactors (power rating less than 500 MWe – micro-reactors are units smaller than 50 MWe) are being designed to address these issues. Many use modular design concepts either to provide power in increments, or they use modular methods of off-site construction – hence the name Small or Advanced Modular Reactors. Some (SMRs) use the tried and tested light-water reactor technology.  Advanced designs use of different types of liquid and gas coolants. These are the Advanced Modular Reactors (AMRs). Both types are now being selected by the AI companies to provide their future sources of power

These new designs have been stimulated and funded mostly by private capital, which has transformed the number and choice of reactor designs being investigated. There are over 120 SMR and AMR projects [WNA SMR Tracker (2026}] being developed or pursued globally. These include every type of established and new technology and there are projects in every part of the world, except Australia, which has a long-term ban on nuclear development. Many of these SMR projects are in the US and Europe, with some of leading SMRs being demonstrated now in China.

Advanced design projects that started in the US earlier have been accelerated by the Presidential Executive Orders of May 2025 [WNN (2025b)], which mandated simpler safety and planning regulations, together with investment by the US DoE promoting early demonstration of new designs. This is in three stages: 

• First, critical experiments in 2026 (4 micro-reactor designs achieved the July, 4 target),
• Second, prototype AMR power units in 2028on DoE land and with DoE licencing, 
• Third, the roll-out of commercial units, in volume, in early 2030s.

Also, some near-term commercial projects (Kairos - Natrium and X-energy with Dow) have benefited from matching funding for construction from US DoE and $17.5bn has been pledged by DoE towards long lead items for commercial power plants in US.

New technologies

How are these new SMR and AMR designs different and why will they be cheaper and more attractive for AI Data Centers?

Production engineering for SMRs

Our work at Cambridge [Roulstone, Lloyd & Lyons (2020)] focused on improving the way in which SMR nuclear construction is done rather than new reactor technology. Light water reactors were used for the analysis. The smaller size of SMRs together with an increased numbers of units made a different construction and supply chain strategy possible. The economy of ‘multiples’, as used in all manufacturing sectors, was set against the economies of ‘scale’, which is common in conventional nuclear design where bigger is better.

We found SMRs built conventionally, will be just as, or more expensive than large (GWe) units. The economies of scale make smaller designs more expensive. Making use of production engineering strategies that promote improved productivity and shorter timescales, it is possible to construct early production units at capital costs below those of a large reactor such as a next-of-a-kind Westinghouse 1.1 GW PWR. Further, with a sizeable order book, a continuous schedule of builds, production learning and lower financing costs during construction because of the shorter build schedule, energy costs could well be competitive with other low-carbon sources of energy. Our production engineering approach employed three broad strategies

• Standardization – Designing the plant once, setting the manufacturing and construction strategy once and building the supply chain once, and then repeating these aspects for each project. This not something that is at all practiced in the nuclear industry nuclear (except in Korea). Each project is considered one at a time with different detailed designs, different equipment and different contractors.
  
  The lack of standardization leads to production waste, and it impedes the other strategies which depend on repetition and regularity. 

• Design for modular construction and site assembly – using either on-site factories or off-site factories. This is not usual in nuclear which employs what is called ‘stick-built’ approaches. Basic materials are delivered to site where they are fashioned into sub-units and joined up with specialised equipment such as vessels, controls and turbines. In this model, the majority of the cost of the power plant are site labour and overheads [OECD (2015)]. Site workforces are very large 10,000 or more. Productivity is low because of the lack of specialised tooling, poor access and the difficulty of organising such a complex site, made worse by the very high and bureaucratic quality standards of the industry. Good practice in on-site modular design has been demonstrated by Hitachi with their BWRs in Japan. Nevertheless, production in factories has better conditions, better access, better tools and equipment, together with better systems for control of quality and documentation. All these changes substantially improve workforce productivity and hence reduce cost.

• Production learning – improved productivity by transferring learning from one production unit to the next. This approach is widespread in all other industrial sectors and is one of the main sources of progressive cost reduction throughout the 20th century, whether in consumer goods, automotive or capital goods, some as complex as a nuclear power plant.

Sector wide data on production learning shows that cost improvement from results from production learning - doubling volumes reduces labour costs by 15-20%, a factor that is constant for every doubling of volume. Only in the nuclear sector are the production learning rates less than 5%. This is due to two factors: firstly, the continuous changes in designs and in supply chain between projects; secondly, the long gaps between reactor projects, during which improvement lessons and practices are forgotten. While there will be limits to the amount of fabrication and construction that be transferred to factories, analysis of the transport and economic issues indicates that up to 60% of on-site construction work can be transferred to factories, with on-site manufacture remaining as an option, if less desirable.

There are many credible designers of light water reactors but only a small number of vendors have active SMR projects - which include:

• RR-SMR (PWR 470 MW) chosen for the UK nuclear program with 3 units at Wylfa in Wales for a start but likely to be a dozen across the country, also 6 units in Czech Republic and 3 units in Sweden with others possible in Holland etc
• GE-Hitachi (BWRX-300) has announced a project of 4 units in Ontario, Canada with further projects likely in Saskatchewan, Tennessee and Illinois, and 6 units in Poland and potentially others in Slovakia.
• Holtec (SMR-300) 2 units in Michigan, US at Pallisades, Michigan, also 4 units announced for AI Data Center at Cottan in Notts, England.
• NuScale (4-12 x77 MW modules). NuScale’s incremental modular power plant design was once was the leader with the first NRC safety clearance and first project announced in Idaho, since cancelled- perhaps it has had a cost problem compared with low-cost gas and wind generation, which is also holding-up nuclear projects more broadly in US. A 6-module project is proceeding in Romania and there is the potential for working with developers Standard Nuclear and Entra1 in the US. Further opportunities for NuScale in S Korea are for nuclear-powered AI Data Centers.

The key issue for all SMR vendors is building the volume and sequence of the order book to make the overall program affordable and to allow flow of manufacture, the key to progressive cost reduction. The conditions for volume production of these SMRs which is an order book of about 20 units, look likely to be fulfilled for some vendors.

Advanced reactor technology

There are many Advanced Modular Reactor (AMR, sometimes called Generation IV) design concepts with differing objectives and unique features. AMRs seek advantage over conventional light water-cooled reactors through either better use fuel and/or with simpler more cost-effective designs. AMRs can be classified by the type of coolant: gas - carbon dioxide or helium: liquid metal – sodium or lead/bismuth; molten salt – mixtures of fluoride, chloride or nitrate salts. 

Most AMRs designs have moderators to slow neutrons so that they operate in a manner similar to water-cooled reactors, though they require higher levels of fuel enrichments in the range 10-15% (High Assay Low Enrichment Uranium – HALEU fuel) compared with 5% for existing reactors. Liquid metal-cooled and some molten salt-cooled designs don’t have moderators which allows them to convert uranium or thorium fuel into fissile elements for future use in reactors i.e. breeding. These fast neutron systems require even higher levels enrichment (15-20%).

Each type of AMR has been built and operated before, but have differing levels of maturity:

• Gas-cooled – the high temperature gas-cooled designs being proposed build on both extensive experience with AGRs and more importantly on international high temperature R&D programs in UK, Germany, US, Japan and China. They use their low power density and a wholly ceramic fuel (TRISO) to make their safety case much simpler and avoid complex dedicated safety systems. 
  
  They offer high temperatures, 750^oC or more, making them efficient in generating power and capable of providing process heat. The most recent unit is HTR-PM in China which has demonstrated some of the key safety features. Though the experience is only ~50 reactor years of prototypes these designs are the most mature of the AMRs. Current commercial projects include: X-energy building 4 80 MW units for Dow in Texas and plans for multiple units for Amazon, in both Washington State in US and Teesside in UK.

• Liquid metal-cooled -   There is long history of this concept starting in the mid-1960s   driven by the aim of breeding new fuel. Research and prototype units were built in Russia, UK, US, France and Japan almost all of which used a liquid sodium-based coolant. They were backed by large R&D programs both for the reactor and reprocessing to close the fuel cycle. Though both sodium and lead are proposed now, the leading projects are in the US making use of the designs and fuel technology from the EBRII program at Idaho.
  
  Because liquid metals boil at high temperatures they can operate at their normal outlet temperatures of more than 500^oC without pressurisation. No expensive high-pressure vessels and piping systems are required. Also, their high-power density means the reactors are relatively small. Both features make them cheaper. 
  
  Operating experience from R&D programs is substantial - ~50o reactor years, most of which more recently has been in Russia. After the US fast reactor program was closed 30 years ago the US fast reactor experience was carried forward by GE in the PRISM design. The Natrium reactor (325 MW) being built at Kemmerer in Wyoming uses both EBRII and PRISM experience. Natrium is led by Bill Gates and is seeking other sites in both the US and the UK. The start-up OKLO Aurora (75 MWe) is closer to the EBRII size and technology. It has a very ambitious development schedule and a program of build in volume aimed at AI Data Centers, with plans for a prototype unit on DoE land at INL, operating in 2028.

• Molten salt-cooled – Chloride and Fluoride salts are molten above about 350^oC (Nitrates have lower melting points) and then flow and cooled like water. They have boiling points above 600^oC and therefore can operate at low pressure. Molten salts have been proven as long-term cooling media in solar storage applications. Small molten salt reactors were operated at Oak Ridge in US in 1960s but the program was shut down to focus on light water reactors. Therefore, the operating experience of molten salt design is less than other AMRs. Also, there is little commonality between the designs being developed with no agreement of: the choice of salt, whether the fuel is dissolved in the salt, or is contained in metal pins or ceramic spheres, reactor physics - whether fast or thermal designs.
  
  The most only operating test unit is TMSR in China, a 2 MWt thorium/uranium fuelled conversion reactor. Leading MSR projects include the Canadian IMSR in N Carolina, TerraPower Molten Chloride Reactor Experient at INL, Moltex SSR in Canada and Kairos Hermes at INL. Other MSR power concepts are being pursued in France, UK, Denmark and US - together with much smaller micro-reactor designs for military use. The commercialisation of MSRs looks likely to be later in time than other AMRs.

Though there are many different SMR/AMR technologies they are all part of the revolution in the nuclear industry to reduce costs and accelerate delivery, by changing from one project at a time to sequential project-like delivery of program of units.

AI Data Centers

AI Data Centers have high (up to 1 GW per site) and variable power demands. They require reliable, clean and dedicated supplies not dependant on the Grid. The options for providing these are limited to gas turbines, or nuclear power. While current Data Center projects are using gas turbines, this will become less acceptable in the future and the Hyperscalers wish to be seen as climate sensitive. SMRs are closer in power to the requirement of AI Data Centers - 100-200 MW at a time, and they could be delivered much more quickly and potentially at lower cost. The question is how quickly? US Government and INL are seeking to use this opportunity to press for new urgency in the nuclear industry and to facilitate SMRs and AMRs meeting the AI demands for power. 

Whilst, SMRs could provide the power needs of AI, there are many questions (see Washington Post) to be answered before nuclear power and Data Centers are successfully paired, and before this can be done at the scale and speed required. Some of these questions are common to all projects, some are industry wide, others depend on the specific design of nuclear reactor. Others relate to the process of rejuvenation of the nuclear industry, finding the increases in manpower and skills to design, build and operate these reactors.

“Some plans to quickly cash in on AI’s insatiable appetite for electricity are confronting the realities of - physics, supply chains, energy markets and labor shortages” 

Washington Post April 26

A high level of review of the main issues facing nuclear-powered AI data Centers covers: finance, engineering and safety, fuel supply, construction and supply chain, manning and operations - see Annex A. It describes each issue, provides some context and gives some suggestions about possible solutions. A few examples will illustrate the range of issues.

• Short term power variability - is a key issue. Data Center loads are made up of processing, cooling and domestic, with the first making up a majority and being the most variable. There is little data on AI Data Center load profiles. INL analysis (2025) shows the range and variability of Data Center demands for a Large Language Model in learning mode, which appears to be the most demanding activity.
• While mean demand varies quite slowly, minute by minute power changes are large ~50% of max demand. How will this rapid variability be met? Is this power profile typical?

We know that the speeds of response of all types of fossil fuel and of nuclear power generators are in range of 5-10% of rated power per minute - much slower than variation in AI demand which could be much more 100% per second. The speed of response of nuclear reactors are:

• For large PWRs ~3-5% of rated power per minute, with SMRs rates up to 10% per minute. 
• BWRs are probably less able to change power as quickly, especially with the natural circulation design being proposed.
• Though the information is not readily available for gas-cooled reactors, it is likely their allowed power ramp rates will be lower, especially for low power levels.
• Fast reactors are somewhat less stable than PWRs but information on their power ramping capability is limited - similarly for molten salt reactors.
• Frequent power changes would require reactivity compensation in most cases.

To develop engineered solutions there needs to be better information about demand profiles of total data center. It is recognised that these issues will have been experienced by the nuclear AI data centers being built close to existing nuclear plants in Pennsylvania. Nevertheless, whether power is provided from fossil fuels or nuclear, the power system will require short-term (less than 15 mins) fluctuations to be covered by more responsive battery and inverter systems. Their control response versus the Data Center demand is to be verified by modelling etc. Battery power systems can be sized to compensate for short-term power variations and where the specific design of SMR or AMRs has a slower response capability, larger batteries capacities and inverter control systems will be required. 

Further issues to be considered: Frequency response of control scheme where the plant is Behind-the-Meter or not connected to the Grid and the need for stabilisers/spinning inertia to stabilise the system. Also, the reactor system should plan for the effects of reactivity compensation, resulting from frequent power changes.

• HALEU fuel availability - AMRs use HALEU fuel (enrichment high than regular commercial fuel 5%, but lower than 20%) which is not available outside Russia as commercial enrichment facilities in the West were closed. In the near-term AMR projects are depending on the release in the US of military stocks of enriched uranium, which are limited. The technology for uranium enrichment is understood and will require investment for rapid growth in the enrichment demands in both US and Europe, together with new and/or much expanded fuel production facilities. 
  
  Many AMRs use ceramic TRISO fuel, which consists of hundreds of thousands of millimetre-sized fuel particles each coated with silicon carbide and graphite to provide fission product containment. TRISO fuel is difficult to make consistently to the required standard. It has been made only at lab-scale and at very high cost – 10 times LWR fuel cost. 
  
  The alternative to enriched uranium of using plutonium, sourced from excess military stocks is being explored in the US. This fuel route will not be without its problems in terms of safeguards, fabrication and handling etc. 
  
  New investment is required to both process improvement and production volume capabilities, some which is already committed - in the US [Nuclear News (2025)] [Framatome (2025)] and for HALEU in UK [HMG (2024)].

• How to ensure nuclear supply chains are ready? - In both US and Western Europe nuclear supply chains are very weak and provide limited choice as a direct result of the lack of nuclear orders for last 30 years. Suppliers have either merged or have left the sector. Few remain and those that remain are not necessarily in the financial and business health to adequately respond to the demands of rapid growth and the novelty of AMRs. 
  
  Reactor vendors are moving to re-build supply chains for the relatively small number of current active projects. There are a few existing production suppliers for LWR fuel, vessel and pumps that are applicable to SMRs, but the capacity of suppliers needs to be expanded many times over. The nuclear industry should invest in vessel, pump and valve manufacture, securing supply chains, by either collaborative upgrading of facilities, a program of acquisition, or upgrading of non-nuclear but related suppliers to enable them to meet nuclear standards.
  
  Reactor instrumentation and control is a critical and specialised part of the nuclear supply chain, which can be addressed by a variety of means – investing in current suppliers, transferring nuclear knowledge to allied aerospace suppliers, or creating new suppliers from scratch.
  
  Perhaps more difficult to address is the new workforce, with the skills and the know-how to build and operate the new reactors. Countries that have demonstrated the best nuclear construction have built the workforce and their skills over several projects and many years. Also, there are specific skills such as nuclear welding that are crucial. These cannot be acquired at a drop of hat. It takes many years to acquire the capability to produce nuclear grade welds that pass the rigorous quality checks, first time.
  
  Both needs the broader dearth of a skilled workforce and the need for specialised skills point towards the potential productivity gains from factory manufacture, with their better tools, better systems, retention of skills and a proper flow of work. Nevertheless, the huge volume of technicians, planners, QC inspectors to be trained will be a tough challenge for the industry.

Conclusions

AI Data Center developments provide a huge opportunity for nuclear power, as large as the global build out of nuclear in the 1970 and 1980s. SMRs with either proven light water technology, or the emerging advanced technologies could provide the power for much of this new demand. The financial, technical and industrial issues to couple nuclear power plants to Data Centers are many and will require a broad whole systems approach. 

AI Data Centers represent a welcome but daunting challenge for an industry weakened by 30 years of low demand and with an aging workforce. These challenges can be overcome by strategic and systematic analysis across the different disciplines and a series of clear action plans to close the gaps. 

This challenge is being driven along both by the need for large amounts of clean energy and by President Trump’s Executive Orders of 2025 which provided both the impetus and the framework for rapid development and deployment of new nuclear technologies.

SMRs and AMRs are the vehicle for this transformation and rejuvenation of the industry. SMRs are different from large reactors in both scale and strategy. They are fundamentally about modernising the industry: From one project at a time, to sequential product-like delivery. 

How will the nuclear industry measure up to these challenges and capture the opportunity? It will be successful only by taking a whole system approach - not by focusing on one issue over another – not just the reactor, or the power profile, or the fuel, or the supply chain, or the workforce, or finance – but all of these viewed as a system.

References

1. [1]Goldman Sachs (2025). AI to drive 165% increase in data center power demand by 2030
2. [2]Feb 4, 2025. https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030 
3. [3]IEA (2025). Energy and AI, IEA, Paris https://www.iea.org/reports/energy-and-ai 
4. [4]McKinsey (2025). Data Center demands May 20 2025. https://www.mckinsey.com/featured-insights/week-in-charts/data-center-demands 
5. [5]Investing (2026). Nuclear Power Is the Only Real Answer to AI Electricity Demand May 28, 2026.
6. [6]Wikipedia Vogtle Electric Generating Plant (2026). https://en.wikipedia.org/wiki/Vogtle_Electric_Generating_Plant  
7. [7]Wikipedia. Barakah Nuclear Power Plant (2026). https://en.wikipedia.org/wiki/Barakah_nuclear_power_plant 
8. [8]WNN (2025a). Fermi America, Hyundai E&C team up for Texan reactors. WNN Jul 31 2025. https://www.world-nuclear-news.org/articles/fermi-america-hyundai-ec-team-up-for-texan-reactors 
9. [9]DOE (2026). Department of Energy Announces American Nuclear Supply Chain Loans. US DoE. Jun 23 2026. https://www.energy.gov/articles/department-energy-announces-american-nuclear-supply-chain-loans#:~:text=The%20%2417.5%20billion%20American%20Nuclear,by%20up%20to%20three%20years. 
10. [10]WNA (2026) World Nuclear Association SMR Global Tracker
11. [11]WNN (2025b). US companies welcome Executive Orders. World Nuclear News. 30 May 2025.
12. [12]Roulstone, Lloyd & Lyons (2020).  Expanding nuclear’s contribution to Climate Change with SMRs. Nuclear Future Journal 633. Sep/Oct 2020. https://www.researchgate.net/publication/345362136_Expanding_Nuclear's_Contribution_to_Climate_Change_with_SMRs
13. [13]OECD (2015). Nuclear New Build: Insights into Financing and Project Management. OECD- Nuclear Development Report 7195. 2015.
14. [14]INL (2025) Characterizing Large Loads. A Taxonomy to Support Large Load Integration. INL/RPT-25-88356. 
15. [15]Nuclear News (2025). Work advanced on X-energy TRISO facility. https://www.ans.org/news/2025-08-11/article-7269/work-advances-on-xenergys-triso-fuel-fabrication-facility/ ANS April 11 2025.
16. [16]Framatome (2025). Framatome and Standard Nuclear Announce a U.S.-based TRISO Nuclear Fuel Production Joint Venture. https://www.framatome.com/medias/framatome-and-standard-nuclear-announce-a-us-based-triso-nuclear-fuel-production-joint-venture/ 
17. [17]UK Gov (2024). UK invests in high-tech nuclear fuel to push Putin out of global energy market. £300 million UK investment to support domestic production of fuel required to power next-generation nuclear reactors. 7 Jan 2024. https://www.gov.uk/government/news/uk-invests-in-high-tech-nuclear-fuel-to-push-putin-out-of-global-energy-market

Annex A

Key questions nuclear power AI data Centers to be resolved.