Small Modular Reactors for AI Data Centres: How They Work

What is a small modular reactor?

A small modular reactor is a nuclear fission reactor designed around smaller units and modular construction. The International Atomic Energy Agency describes SMRs as having electrical generating capacity of up to 300 megawatts per unit. Modular refers to systems or components that can be assembled in factories and transported for installation. [1]

The size limit applies per unit, not necessarily to the combined output of a plant with several units. The definition also does not make every advanced nuclear reactor an SMR: an advanced design can exceed this usual electrical-capacity threshold. [1] [6]

Smaller units could suit some sites or grids that cannot accommodate a much larger reactor. But a smaller reactor still needs appropriate infrastructure and approval. Modularity is an approach to construction, not a promise that a complete power station arrives ready to operate without site work. [1]

MWe, MWth and MWh measure different things

MWe means megawatts of electrical output. Thermal power, often written MWth, measures heat rather than the electricity eventually produced. Comparing those ratings directly can misrepresent what is available to a data centre.

MW describes power at a particular rate; MWh describes energy delivered over time. A fleet target of 500 MW is therefore not 500 MWh of annual electricity. Likewise, a temporary storage-assisted peak is not the same as the reactor’s continuing output. The Natrium example below illustrates that distinction. [4] [6]

How an SMR turns nuclear heat into electricity

An SMR uses the same basic principle as other fission power reactors: nuclear reactions release heat, which the power plant converts into electricity. The IAEA describes the familiar route through a coolant, steam, a turbine and an electrical generator. Reactor designs differ, so the equipment arrangement is not identical in every SMR. [1] [2]

The reactor does not supply heat directly to a GPU. The data centre receives electrical power, and its electrical infrastructure delivers usable power to computing equipment. Servers, networks, storage and cooling all contribute to the facility’s electricity demand. [3]

Fission is also different from fusion. SMRs discussed here use fission; a promising fusion experiment is not evidence that a fusion power station can deliver dependable electricity to an AI campus. [2]

Why cooling still matters after shutdown

Shutting down a nuclear chain reaction does not immediately remove every source of heat. The U.S. Nuclear Regulatory Commission defines decay heat as heat from the radioactive decay of fission products after shutdown. Heat removal therefore remains a design and safety concern. [7]

Many proposed SMR designs emphasise passive systems and inherent safety characteristics. The IAEA describes approaches relying on physical effects such as natural circulation and gravity for particular functions. That does not mean all SMRs work the same way or that a reactor needs no trained personnel, oversight or safety assessment. [1]

The power plant’s heat-removal requirements and the data centre’s cooling requirements are separate. A nuclear electricity source does not make heat from servers disappear. The campus still needs environmental control and equipment suited to its computing load. [3]

What factory-built modules could change

Factory assembly aims to make manufacturing more repeatable and reduce some on-site construction. Smaller units can also allow a project to add capacity incrementally rather than committing to one much larger unit at the start. These are the design advantages described by the IAEA. [1]

Repeat production is the intended benefit, not an established saving for every first project. Site work, transport, integration and approval still matter. The same IAEA explainer says economic competitiveness needs to be proved in practice once projects are deployed. [1]

Why AI data centres are looking for more power

The IEA’s Key Questions on Energy and AI, published on 16 April 2026, updates the demand picture. It puts total data-centre electricity consumption at about 485 terawatt-hours in 2025 and projects roughly 950 TWh in 2030, around 3% of global electricity demand by that date. These are all-data-centre figures, not AI-only consumption. The 2030 number is a projection, not electricity already used. [3]

The report says AI-focused consumption is growing faster than the overall data-centre total. It also emphasises changing model uses, efficiency and infrastructure bottlenecks. More demanding AI tasks can increase electricity needs, while better hardware and software can reduce energy per task. Neither trend alone fixes the future total. [3]

A campus needs more than an annual energy quantity. Electricity has to be available in the required location, while transformers, connections and other equipment must support its load. A large amount of proposed generating capacity elsewhere does not automatically solve a local connection constraint. [3]

Firm generation is useful—but it is not guaranteed uptime

Nuclear power can provide low-carbon generation without depending on sunshine or wind at the time of generation. That makes it an option alongside other sources in a system serving sustained electricity demand. The IAEA also discusses nuclear-and-renewables hybrid systems rather than treating the technologies as mutually exclusive. [1] [2]

A data centre’s load is not necessarily flat just because it operates around the clock. The IEA’s 2026 report describes large, rapid power swings from AI training and model use and highlights storage for reliable supply. Energy quantity, available capacity and response to a changing load are different requirements. [3]

A reactor’s nameplate rating does not prove uninterrupted supply to a campus. The supply arrangement must account for unavailable generation and the facility’s reliability needs. Announcing a nuclear agreement is not a reason to assume the data centre’s other power infrastructure is unnecessary. [3] [4]

Low-carbon does not mean zero lifecycle impact

The IAEA describes nuclear electricity as low-carbon because plants practically do not produce carbon dioxide during operation. That is a statement about generating electricity, not a claim that fuel supply, construction or every related activity has no environmental impact. [2]

Nuclear power also produces radioactive waste and spent fuel that require appropriate management. Calling a supply source low-carbon cannot substitute for examining those obligations, the specific site or the electricity actually delivered. [2]

How an SMR could supply a data centre

One possible arrangement is grid-connected generation, with electricity supplied under a contract. Another concept is generation near a campus with a dedicated connection. The location and commercial arrangement need to be established for each project rather than inferred from the term “SMR”. [1] [4] [5]

The Kairos Hermes 2 announcement is explicitly a planned TVA-grid supply example. It is not evidence of a private wire leading directly to a Google campus. A supply agreement, the physical connection and the electricity reaching a facility are related but separate facts. [5]

Nearby generation may alter a project’s supply options, but it does not remove siting, infrastructure or licensing work. It should not be described as a guaranteed shortcut around connection queues. The IEA’s account of tightening energy-equipment and grid bottlenecks shows why the complete supply chain matters. [3]

SMRs versus large nuclear reactors

Both use nuclear fission. The main differences concern unit size, construction approach and the project’s intended application. Smaller does not automatically mean safer, cheaper per unit of electricity or ready sooner. [1] [2]

The comparison describes broad approaches, not identical features across every design. A large plant can also contain more than one unit, and an SMR site may eventually combine several modules.

What the verified project announcements actually show

Three dated sources give a more useful picture than an unsourced list of future reactors. They describe an agreement, a construction-stage milestone and a regulatory action. None of these sources reports the discussed plants already delivering electricity to data centres. This is a statement about these examples—not a worldwide claim that no such supply arrangement exists. [4] [5] [6]

Google and Kairos: a fleet plan, not one operating reactor

On 14 October 2024, Google and Kairos announced a development agreement for planned electrical generating capacity totalling 500 MW by 2035, with the first deployment targeted for 2030. The 500 MW figure concerns the planned fleet, not one module and not capacity already delivered. [4]

Those dates are targets from a commercial announcement. An agreement is not equivalent to completed construction or permission to operate every planned plant. It also does not establish Google–Kairos power delivery in 2027. [4]

Hermes 2: a dated groundbreaking milestone

Kairos’s 17 April 2026 release announced groundbreaking for the Hermes 2 Demonstration Plant in Oak Ridge, Tennessee. It described planned electricity supply of up to 50 MW to the Tennessee Valley Authority grid, supporting decarbonisation of Google data centres in Tennessee and Alabama. [5]

The release states that Hermes 2 has a construction permit. Groundbreaking and a construction permit are not evidence of an operating plant. These are the milestones recorded in that April announcement; the article does not assume that they are the project’s last possible update before November. [5]

Natrium: construction approval and a capacity distinction

The NRC’s 4 March 2026 release says the Commission authorised staff to issue a construction permit for Kemmerer Power Station Unit 1 in Wyoming. It expected issuance soon and stated that a separate operating-licence application and approval would be required before operation. On this source, the precise March milestone is authorisation to issue—not proof that the permit was issued that same day. [6]

The release describes a 345 MWe plant with storage that can temporarily raise electrical output to 500 MWe. The 500 MWe figure is storage-assisted output, not a continuing reactor rating; this source does not specify the duration of that peak. Nor is either electrical figure the reactor core’s thermal rating. [6]

At 345 MWe, Natrium is above the IAEA’s usual 300 MWe-per-unit SMR threshold. It belongs in the discussion as an advanced sodium-cooled reactor, not as an example silently squeezed into that definition. The NRC release also does not establish an AI data-centre supply deal for it. [1] [6]

What can prevent a promising design from delivering power?

A reactor project needs a deliverable design, supply chain, appropriate fuel and completed regulatory steps. A commercial target is not evidence that all of those conditions have been met. The NRC’s separation of construction approval from operating approval is a concrete example. [6]

Factory production may support repeatable construction, but competitiveness must be demonstrated by delivered projects. The cost of a module alone is not the same as the cost of a working electricity supply arrangement. The IAEA’s own qualifications matter alongside its description of potential modular benefits. [1]

Waste management remains part of the nuclear fuel cycle. Site infrastructure and heat removal also remain necessary. An AI campus must fit its computing and cooling requirements to the available electricity system, not assume that a smaller reactor removes the rest of the facility’s engineering constraints. [2] [3] [7]

The project examples are U.S.-specific. Their agreements and NRC actions do not prove an SMR is available for an Indian campus or establish permission to construct one in India. No India-specific deployment or regulatory approval is asserted here.

Potential uses beyond AI data centres

The IAEA discusses electricity generation, heating, desalination, industrial steam and hybrid energy systems among SMR applications. A suitable project could serve industrial or remote users as well as computing facilities. These are possible applications, not evidence that every design can supply every type of heat or product. [1]

An application needs the right conditions, infrastructure and demonstrated design capability. A district-heating project, for example, has different integration needs from supplying electricity to servers. The role of an SMR should be judged against the specific use rather than a generic list of industries. [1]

What to watch in 2027

Watch for dated construction progress and the next regulatory milestone on specific projects. Distinguish site preparation, reactor construction and authorisation to operate. That prevents a groundbreaking announcement from being presented as electricity already on the grid. [5] [6]

Fuel and equipment delivery, demonstrated performance and repeat-project results will also matter. Modular construction is intended to improve delivery, but the evidence needs to show what a complete project achieved. The energy system’s capacity to supply new data centres remains constrained by more than reactor technology alone. [1] [3]

The Google–Kairos agreement’s first deployment target is 2030, not 2027. The sources do not establish a universal deadline for SMR availability or prove that SMRs will solve all near-term AI power shortages. Keep the 2027 outlook focused on milestones rather than guaranteed electricity or adoption. [4]

Frequently asked questions

Are SMRs the same as fusion reactors?

No. The SMRs discussed here use nuclear fission to create heat. Fusion is a different process, and fusion power-generation technology remains a separate development field. [1] [2]

Are smaller reactors automatically safer?

No automatic conclusion follows from size alone. Many proposed SMR designs emphasise passive safety features, but actual safety claims depend on the design and regulatory evaluation. The NRC example is a specific approval process, not a universal comparison proving every SMR safer than every large reactor. [1] [6]

Is a 500 MW announcement one reactor’s output?

Not necessarily. The Google–Kairos 500 MW announcement is a fleet plan. Natrium’s 500 MWe figure is a temporary storage-assisted peak, while its stated plant rating is 345 MWe. Read the unit, system boundary and project status before comparing numbers. [4] [6]

Will the discussed projects power AI campuses in 2027?

The reviewed sources do not demonstrate that. Google–Kairos’s agreement targets its first deployment in 2030, and the other dated sources concern construction-stage or approval milestones. Targets can change; none should be treated as confirmed electricity delivery. [4] [5] [6]