COMPUTE BESIDE THE REACTOR
A Distributed Nuclear Architecture for the Artificial Intelligence Era
DOI: to be assigned
John Swygert
September 2, 2026
Abstract
The rapid expansion of artificial intelligence is creating a new category of electrical demand: extremely large, geographically concentrated, continuously operating computational loads requiring exceptional power reliability. Much of the present debate treats this demand primarily as a burden on the electrical grid. This paper proposes a different approach. Rather than repeatedly attaching enormous new AI data centers to an electrical system that was not designed for them, future large-scale AI facilities should, where technically and economically appropriate, be preferentially sited alongside new nuclear generating capacity designed with those computational loads in mind.
The proposal is not to place unrelated industrial, commercial, or residential development around nuclear facilities. It is a specific infrastructure pairing: large computational loads with compact, firm nuclear generation.
A distributed network of smaller nuclear plants, including small modular reactors where appropriate, could simultaneously provide dependable power for AI computation, reduce the transmission burden associated with remotely generated electricity, create additional generation capacity, and increase electrical redundancy. Multiple reactor modules could further reduce single-unit dependency by allowing generation to continue when individual modules are undergoing maintenance or experiencing an outage.
The central proposition is straightforward: if society intends to build enormous concentrations of computation, it should consider building appropriately scaled generation beside them.
1. Introduction
Artificial intelligence is changing the physical architecture of computing.
AI is frequently discussed as though it were primarily software. At industrial scale, however, artificial intelligence is also physical infrastructure. Models require processors. Processors require data centers. Data centers require electricity, cooling, substations, transmission capacity, backup systems, and extraordinary levels of reliability.
Consequently, the expansion of AI is increasingly an energy-infrastructure problem.
The conventional development sequence is often approximately:
Build or propose a massive data center → request enormous amounts of electricity from the existing grid → expand transmission and generation sufficiently to accommodate the new load.
That sequence deserves reconsideration.
For sufficiently large computational facilities, the better question may be:
Why are we separating the location of enormous continuous electrical demand from the construction of enormous continuous electrical generation?
Nuclear generation and large AI data centers possess unusually complementary operating characteristics.
A nuclear plant is capable of producing substantial quantities of electricity continuously.
A large AI data center can consume substantial quantities of electricity continuously.
Instead of treating these developments as independent infrastructure projects, they can be planned as a complementary pair.
2. The AI Electricity Problem Is Also a Siting Problem
An enormous new electrical load does not merely require sufficient national generating capacity.
It requires sufficient capacity at the appropriate place, at the appropriate time, through infrastructure capable of delivering it.
This distinction is critical.
A region can theoretically possess adequate total generation while lacking the transmission, substations, transformers, or local generation necessary to serve another extremely large concentrated load.
AI data centers therefore create at least two different infrastructure requirements:
generation of the electricity; and
delivery of that electricity to the computational facility.
Locating generation relatively close to the load does not eliminate the need for a robust electrical grid. It can, however, change the scale of the transmission problem.
Instead of asking an existing grid to transport every additional megawatt across substantial distances, part or most of the computational demand could be supplied by generation developed specifically in conjunction with the facility.
The grid would remain important for redundancy, balancing, imports, exports, maintenance conditions, and emergencies.
But the data center would no longer necessarily begin as another enormous uncompensated demand placed upon existing regional electrical infrastructure.
3. Why Nuclear Generation Fits AI Computation
AI facilities require characteristics that nuclear generation is unusually capable of providing.
Large computational systems benefit from:
continuous electricity;
high availability;
predictable generation;
substantial power density;
long-duration operation;
relatively small fuel-volume requirements;
independence from daily weather conditions.
Nuclear generation possesses these characteristics.
This does not mean that every data center requires nuclear power or that nuclear power should replace every other generating technology.
The proposition concerns the largest and most continuously demanding computational facilities.
If an AI installation requires hundreds of megawatts—or eventually gigawatt-scale power—the facility begins to resemble an industrial electrical load sufficiently large to justify generation planning specifically around it.
At that point, locating generation and computation independently can become unnecessarily inefficient.
4. The Case for Smaller Nuclear Plants
Commercial nuclear development historically favored very large generating stations.
There are obvious economies associated with scale, but scale also creates concentration.
One enormous reactor represents one enormous generating unit.
A distributed system of smaller reactors creates a different architecture.
The existence of nuclear-powered submarines and other naval vessels demonstrates an important engineering principle: a useful nuclear generating system does not inherently have to occupy the scale associated with the traditional multi-gigawatt civilian nuclear station.
Naval reactors cannot simply be transplanted into civilian electrical infrastructure. Their economics, fuels, operating requirements, security arrangements, regulatory environment, and engineering objectives are different.
Nevertheless, their existence demonstrates something conceptually important:
Nuclear power can be compact.
Civilian small modular reactor development attempts to exploit that broader principle using systems designed specifically for commercial electricity generation.
For AI infrastructure, this could be particularly valuable.
Instead of building one enormous nuclear station intended to serve a vast geographic region, smaller plants could be distributed according to major concentrations of computational demand.
5. Generation Should Follow Concentrated Compute
The proposal can be summarized as a change in infrastructure sequence.
Conventional approach
AI demand → grid connection → transmission expansion → additional generation
Proposed approach
AI demand + dedicated nuclear generation → integrated electrical connection → grid
The difference is subtle but fundamental.
Under the second architecture, generation is not merely something the broader electrical system must eventually provide.
It becomes part of the original computational infrastructure plan.
A proposed AI facility requiring extraordinary amounts of electricity could therefore trigger consideration of corresponding new generating capacity.
The larger the computational load, the stronger the argument becomes.
A modest data center does not justify a nuclear facility.
A computational complex requiring power comparable to that consumed by a substantial city presents a very different infrastructure question.
6. Modular Generation and Redundancy
Perhaps the most interesting advantage of smaller reactors is not simply their physical size.
It is modularity.
Consider conceptually a computational complex whose maximum electrical requirement approaches the output of a large conventional reactor.
One architecture would supply that demand using one enormous generating unit.
Another could use several smaller generating modules.
The total generating capacity might be comparable, but the failure characteristics would not be.
With one generating unit:
one reactor offline → dedicated nuclear generation falls dramatically or completely.
With multiple modules:
one module offline → remaining modules continue generating.
This creates graceful degradation rather than an all-or-nothing condition.
Scheduled maintenance could potentially be staggered.
Individual modules could be serviced while others continue operating.
Additional modules could potentially be added as computational demand expands, where reactor design and licensing permit.
The generating architecture could therefore become more closely matched to the modular nature of the computational infrastructure itself.
AI servers are modular.
Data halls are modular.
Cooling systems can be modular.
Electrical distribution can be modular.
There is a strong conceptual argument for making generation modular as well.
7. Geographic Distribution Creates Another Layer of Resilience
Modularity provides redundancy within a generating site.
Distribution provides redundancy across the larger electrical system.
A nation dependent upon relatively few enormous generating centers and major transmission corridors necessarily concentrates infrastructure risk.
A larger number of geographically distributed generating facilities changes that topology.
The failure or maintenance of any individual plant represents a smaller fraction of total generation.
Major computational capacity could similarly be distributed among multiple AI-nuclear locations.
This creates two overlapping networks:
a distributed computational network
and
a distributed generating network.
Neither eliminates the need for the national and regional grids.
Instead, the grid connects facilities that possess substantially more local generating capability than conventional data centers.
That architecture may prove particularly valuable as computation becomes increasingly important to economic activity, scientific research, communications, national infrastructure, and government operations.
8. The Grid Should Remain Part of the Architecture
Colocation should not mean isolation.
An AI facility located alongside nuclear generation should generally remain connected to the broader electrical grid.
That connection provides several advantages.
When local generation exceeds computational demand, electricity may potentially be exported.
When reactors undergo maintenance, electricity can potentially be imported.
If computational demand temporarily decreases, nuclear output need not necessarily be wasted.
If an electrical fault separates part of the local system, alternative pathways may remain available.
The architecture therefore becomes:
Nuclear generation ⇄ AI data center ⇄ electrical grid
rather than:
Electrical grid → AI data center
The distinction is important.
The first architecture creates multiple pathways.
The second makes the computational facility overwhelmingly dependent upon whatever generation and transmission capacity already exists upstream.
9. Cooling and Water Must Be Designed as Systems
Both nuclear facilities and high-density computational facilities reject substantial quantities of heat.
That fact has sometimes produced overly simplistic discussions about water consumption.
Water should instead be treated through explicit mass and energy accounting.
The relevant questions include:
What quantity is withdrawn?
What quantity is circulated?
What quantity is discharged?
What quantity evaporates?
What quantity requires replacement?
What is the source?
What is the discharge temperature?
What happens during drought conditions?
What cooling technologies are employed?
What happens during equipment failure?
What is the local watershed capacity?
Water that evaporates has not been destroyed. It has changed phase and entered the hydrologic cycle.
Nevertheless, its temporary removal from a particular local liquid-water supply can matter, particularly in water-stressed regions.
Therefore neither of the simplistic claims—
“the facility uses enormous amounts of water”
nor
“the facility uses essentially no water”
—is sufficient without defining the system boundary.
A colocated nuclear-computational facility should permit cooling infrastructure to be considered during initial site selection rather than being treated as an afterthought.
The objective should be to optimize power generation, computational cooling, heat rejection, water availability, and local environmental constraints simultaneously.
10. Interconnection Does Not Mean Elimination of Safety Boundaries
Colocation should not be confused with physically merging every system.
Nuclear safety requires strict boundaries.
Reactor cooling systems, safety systems, emergency power systems, control networks, and other safety-critical components must remain appropriately isolated from ordinary commercial computing infrastructure.
But isolation of safety-critical components does not prevent infrastructure coordination.
Electrical interconnection, secondary thermal systems, heat exchangers, cooling infrastructure, substations, transmission equipment, and other systems can be designed around the combined requirements of the generating and computational facilities while preserving necessary nuclear safety boundaries.
The principle is therefore:
integrated infrastructure with appropriate engineering isolation.
Those ideas are complementary, not contradictory.
11. AI Could Become an Anchor Customer for New Nuclear Generation
The relationship also works in the opposite direction.
Nuclear power requires enormous initial capital investment.
One of the economic attractions of pairing new generation with major computational facilities is the presence of a large, predictable electricity customer.
AI developers need dependable power.
Nuclear developers need dependable electricity purchasers.
That creates the possibility of long-term power agreements and coordinated development.
Instead of AI expansion merely increasing electricity prices or competing for existing generation, some portion of future AI expansion could become a financial mechanism supporting construction of additional generating capacity.
This distinction is important.
If a new data center requires 500 megawatts and simply attaches itself to an already constrained grid, it has created 500 megawatts of additional demand.
If development of that facility helps finance substantial new firm generating capacity, the system-level consequence can be very different.
AI demand can potentially become an incentive to build infrastructure rather than merely an additional consumer of existing infrastructure.
12. Scaling Generation With Computation
AI infrastructure rarely appears at its ultimate scale on its first day of operation.
Computational campuses expand.
New data halls are constructed.
Processor densities increase.
Models become larger.
Demand evolves.
A modular generating strategy offers the possibility of matching generation expansion more closely to computational expansion.
Conceptually:
Compute Phase 1 → Generation Module A
Compute Phase 2 → Generation Module B
Compute Phase 3 → Generation Module C
The actual engineering and licensing process would be substantially more complicated, but the planning principle is valuable.
Generation growth and computational growth need not be treated as unrelated processes.
They can be coordinated from the beginning.
13. The Siting Principle
This paper does not propose that every nuclear facility should become an AI campus.
Nor does it propose that every AI data center should have a nuclear reactor.
Instead, it proposes a planning preference for circumstances in which both projects are already justified.
When governments, utilities, reactor developers, and technology companies consider:
a new very-large-scale AI facility
and
new nuclear generating capacity,
they should explicitly evaluate whether colocating those investments provides greater system efficiency and resilience than developing them independently.
For the largest computational facilities, this evaluation should become routine.
The relevant comparison should include:
generation cost;
transmission requirements;
grid-upgrade requirements;
reliability;
reactor capacity;
computational demand;
cooling requirements;
water availability;
site security;
emergency planning;
land requirements;
expansion capability;
redundancy;
construction schedules;
and total system cost.
The correct decision will differ by location.
But failing even to evaluate the combined architecture would increasingly make little sense.
14. A Different Way to Think About AI Energy Demand
AI electricity demand is generally presented as something society must somehow accommodate.
That framing is incomplete.
Large new loads can also justify large new investments.
Railroads created demand for steel.
Electrification created demand for generating stations.
Automobiles created demand for petroleum infrastructure and highways.
Large-scale computation may similarly create demand for a new generation of electrical infrastructure.
The policy objective should therefore not simply be:
Find enough electricity for AI.
It should be:
Use the emergence of enormous computational demand to justify building a more capable, distributed, redundant electrical system.
Nuclear generation is particularly suited to providing the firm component of that architecture.
15. The Broader Resilience Argument
The strongest argument for distributed nuclear-computational infrastructure may ultimately be redundancy.
Modern societies increasingly depend upon computation and electricity simultaneously.
If computational infrastructure becomes more concentrated while its electrical supply remains vulnerable to constrained transmission and concentrated generation, two forms of infrastructure dependence compound one another.
A distributed architecture moves in the opposite direction.
Multiple generating locations.
Multiple computational locations.
Multiple reactor modules.
Multiple grid connections.
Multiple electrical pathways.
The objective is not to make failure impossible.
No infrastructure architecture can accomplish that.
The objective is to make individual failures smaller, more containable, and more recoverable.
A resilient system should degrade in pieces rather than collapse as a whole.
16. Policy Recommendation
National and regional infrastructure planning should establish a formal AI–Nuclear Colocation Assessment for proposed computational facilities above an appropriate electrical-demand threshold.
Such an assessment should not mandate nuclear power.
It should determine whether new nuclear generation located at or near the proposed computational facility would outperform conventional grid supply when total system requirements are considered.
The analysis should include:
generation + transmission + grid upgrades + reliability + cooling + water + redundancy + expansion.
For sufficiently large facilities, the comparison should also examine multiple smaller reactor modules against a single larger generating unit.
Where the combined architecture produces superior lifecycle economics, resilience, and environmental performance, permitting and infrastructure planning should allow the two projects to proceed as coordinated developments while maintaining all required nuclear safety and security standards.
Conclusion
Artificial intelligence is creating electrical loads of a magnitude that forces reconsideration of how computational infrastructure is built.
The conventional approach—construct enormous concentrations of computation and then determine how an existing electrical system can supply them—may not remain the most rational architecture.
A simpler principle deserves serious consideration:
Put major new generation near major new demand.
For the largest AI facilities, nuclear power offers an unusually strong match.
Smaller and modular nuclear generating systems could add another advantage by distributing generation geographically and dividing capacity among multiple units. The result could be an architecture in which computational expansion finances additional firm generation, transmission burdens are reduced, individual outages become less consequential, and both electricity and computation become more geographically resilient.
The proposal does not require surrounding nuclear plants with unrelated development. It does not require every data center to become nuclear-powered. It does not require abandoning the electrical grid or other forms of generation.
It requires recognizing an unusually complementary infrastructure relationship.
AI needs enormous quantities of dependable electricity.
Nuclear power can provide enormous quantities of dependable electricity.
When both are going to be built, building them together may make considerably more sense than building them apart.
The artificial intelligence era will require new computing infrastructure.
It will also require new energy infrastructure.
There is no compelling reason those two planning problems must remain separate.
John Swygert
Ivory Tower Publishing
TSTOEAO.com
IvoryTowerJournal.com
SecretarySuite.com
Copyright © John Swygert 2026
