Chapter 3: The Modular Handoff

SMRs as the AI Battery

The most consequential technology deal of 2025 was not an AI model release, a chip announcement, or a software acquisition. It was a twenty-year Power Purchase Agreement between Microsoft and Constellation Energy to restart a nuclear reactor that had been dormant since 2019. Three Mile Island Unit 1 — the reactor that shares a name with America’s most infamous nuclear accident, though Unit 1 was not the reactor involved in the 1979 partial meltdown of Unit 2 — is scheduled to resume operations in 2027, delivering 837 megawatts of carbon-free, baseload electricity directly to Microsoft’s Azure inference infrastructure under the “Crane Clean Energy Center” branding.

This single deal tells you everything you need to know about the Thermodynamic Wall. The most valuable software company on Earth looked at the future of AI, calculated the energy requirements, evaluated every available generation technology — solar, wind, natural gas, geothermal, battery storage, hydrogen fuel cells — and concluded that the answer was a fifty-year-old nuclear reactor. Not a trendy renewable. Not a venture-backed energy startup. A pressurized water reactor designed in the 1960s, built in the 1970s, and operated without incident for decades before being retired for economic reasons. Nuclear — the one energy source that delivers continuous, weather-independent, carbon-free power at the density and reliability that inference clusters demand.

Microsoft is not alone in this conclusion. It is simply the most visible participant in what the PredictionOracle calls the Modular Handoff: the transfer of AI energy supply from the legacy utility grid to dedicated, sovereign, nuclear-powered generation — and specifically, to the next generation of Small Modular Reactors (SMRs) that are being designed, licensed, and built explicitly to serve as the electrical substrate of the Synthesis World. The Handoff represents the most significant shift in nuclear energy’s role since the Navy’s nuclear propulsion program in the 1950s: from centralized utility generation to distributed, customer-dedicated industrial power.

The Nuclear Logic

Why No Other Energy Source Fits the Inference Profile

The inference profile of a modern AI data center has four non-negotiable characteristics, and nuclear is the only generation technology that satisfies all four simultaneously. This is not an argument from ideology or environmentalism. It is an argument from physics and economics — the same kind of argument that led the US Navy to adopt nuclear propulsion for its submarine fleet in the 1950s, and for the same fundamental reasons: when you need continuous, high-density, fuel-independent power in a self-contained system, nuclear is the only option that does not require constant resupply.

Baseload Continuity: An inference cluster operates 24 hours a day, 365 days a year, without seasonal variation or time-of-day cycling. The reasoning kernel does not sleep, does not observe weekends, and does not reduce output during cloudy weather. Solar generation drops to zero every night and varies by latitude, season, and cloud cover — achieving capacity factors of only 20% to 25% in most temperate climates. Wind generation is intermittent and unpredictable on timescales shorter than a week, with capacity factors typically between 25% and 45%. Natural gas can provide baseload, but it is subject to fuel price volatility (Henry Hub natural gas prices have swung between $1.50 and $9.00 per MMBtu in the past five years) and pipeline infrastructure constraints that create geographic dependencies.

Nuclear reactors operate at capacity factors above 90% — the highest of any generation technology, per the US Energy Information Administration’s historical data — because their fuel cycle is measured in years rather than tankfuls, and their output is independent of weather, time of day, or commodity markets. A uranium fuel assembly loaded into a reactor core provides 18 to 24 months of continuous generation before refueling is required. For advanced SMR designs like Oklo’s Aurora, the refueling interval extends to 10 to 20 years. No other energy source offers this degree of operational continuity.

Density: The physical footprint of a nuclear reactor relative to its output is orders of magnitude smaller than any renewable alternative. A 300-megawatt SMR occupies a site of approximately 35 to 50 acres, including security perimeter, cooling infrastructure, and administrative facilities. To generate the same 300 megawatts from solar would require approximately 1,500 to 2,000 acres of panel coverage — an area larger than most municipal airports — plus battery storage for nighttime operation that would add hundreds of additional acres. Wind generation at 300 megawatts requires approximately 30,000 acres when accounting for turbine spacing requirements.

For an Energy Island that needs to co-locate generation with compute infrastructure on a single campus, the density advantage of nuclear is not marginal. It is categorical. A solar-powered Energy Island requires a land acquisition measured in square miles. A nuclear-powered Energy Island can fit its generation, cooling, and compute facilities within a single industrial campus.

Dispatchability: Nuclear reactors can operate in load-following mode, adjusting their output within a range of approximately 50% to 100% of rated capacity to match the demand profile of the inference cluster. This capability, demonstrated extensively by the French nuclear fleet (which has operated in load-following mode for decades to accommodate variable demand), is critical because training runs and large-scale inference bursts create demand spikes that exceed the facility’s average load by 30% to 50%.

A generation source that cannot ramp to meet these spikes forces the operator to either curtail the workload (sacrificing inference revenue) or maintain expensive backup generation (increasing capital cost and operational complexity). Nuclear can serve as both the baseload and the surge provider, eliminating the need for a hybrid generation stack that combines multiple fuel sources, each with its own maintenance requirements, fuel supply chains, and regulatory obligations.

Sovereignty: A nuclear reactor fueled with enriched uranium has a refueling cycle of 18 to 24 months for conventional pressurized water reactor (PWR) designs and up to 20 years for some advanced SMR designs. This means that once fueled, the reactor operates independently of external supply chains for years at a time — a degree of energy independence that no other technology can match.

Compared to natural gas (which requires continuous pipeline delivery and is subject to geopolitical disruption, as Europe discovered after the 2022 Nord Stream incidents), solar (which requires continuous panel replacement, inverter maintenance, and battery cycling), or grid power (which requires continuous cooperation with a utility whose priorities may diverge from yours), nuclear offers the deepest form of energy sovereignty available to a non-state actor. The fuel is loaded, the reactor operates, and the external world becomes irrelevant to your energy supply for years at a time.

The SMR Generation

From Legacy Reactors to Modular AI Batteries

The legacy nuclear fleet — the large, gigawatt-scale pressurized water reactors built between 1960 and 1990 — was designed for a different era. These reactors were conceived as centralized power plants serving entire metropolitan regions through the utility grid. They were built to last fifty years, licensed through a regulatory process that assumed a multi-decade construction timeline, and operated by utilities whose business model was based on selling kilowatt-hours to residential and commercial customers at regulated rates. The regulatory framework that governed them — the NRC’s 10 CFR Part 50 — was designed for an era when a new reactor application arrived once per decade and the review process could afford to be exhaustive rather than efficient.

The SMR generation inverts every one of these assumptions. Small Modular Reactors are designed not as regional power plants but as dedicated energy sources for specific industrial customers — and in 2026, the most important industrial customer is the AI data center. Where legacy reactors were designed to be large (economies of scale), SMRs are designed to be manufacturable (economies of production). Where legacy reactors were custom-engineered on-site, SMRs are factory-fabricated and transported to the deployment location as assembled modules. Where legacy reactors required staffs of hundreds, next-generation SMRs target autonomous or minimal-staff operation.

Oklo (Aurora Powerhouse): Oklo’s Aurora reactor is a fast-spectrum, metallic-fueled fission reactor designed for unattended operation — meaning it requires no on-site operators during normal conditions, relying instead on remote monitoring and AI-driven autonomous control systems. The current design scales from 15 to 50 megawatts electric (with thermal output suitable for district heating or industrial process heat), with modular units that can be combined in arrays to serve larger loads.

Oklo has begun site preparation activities at Idaho National Laboratory, with its first Aurora unit targeting late 2027 for initial operation following receipt of a combined license from the NRC. The company’s CEO, Jacob DeWitte, has stated that the Aurora design is intended to be “the simplest nuclear power plant ever built” — a deliberate departure from the engineering complexity of legacy designs. The strategic significance of Oklo’s deals — 1.2 gigawatts committed to Meta, 12 gigawatts committed to Switch — is not the wattage. It is the operating model: these are not utility contracts negotiated through a public utility commission. They are dedicated, sovereign power agreements in which the AI operator and the reactor operator are functionally fused into a single entity. The Handoff is literal.

NuScale (VOYGR): NuScale’s 77-megawatt electric (250 MWt) NuScale Power Module is the first — and as of early 2026, still the only — SMR design to receive full Standard Design Approval from the US Nuclear Regulatory Commission (granted in January 2023 under 10 CFR Part 52). The VOYGR platform deploys these modules in arrays of 4, 6, or 12, enabling a single plant to scale from 308 MWe to 924 MWe depending on customer requirements.

Standard Power has contracted for 24 NuScale modules across two dedicated AI data center facilities in Ohio and Pennsylvania, with deployment targeted for 2029-2030. The NuScale design emphasizes passive safety systems — the reactor shuts down and cools itself through natural convection without operator intervention, external power, or mechanical pumps in the event of an anomaly, a design feature known as “walk-away safe.” Internationally, the RoPower project in Romania (a joint venture between NuScale and Romania’s Nuclearelectrica) is approaching final investment decision in 2026-2027, with plans for a 462-MWe six-module plant at the Doicești site.

X-energy (Xe-100): Amazon’s partnership with X-energy represents the most direct expression of the “utility company” thesis articulated in this volume’s Preface. AWS has committed capital through its Climate Pledge Fund and entered a co-investment arrangement to accelerate the licensing of X-energy’s Xe-100 high-temperature gas-cooled reactor (HTGR), which uses TRISO (tristructural isotropic) fuel particles — uranium kernels encased in multiple layers of carbon and ceramic that can withstand temperatures above 1,600°C, making a fuel meltdown physically impossible.

The Cascade Nuclear Energy Center near Richland, Washington — located adjacent to the Hanford Nuclear Reservation, the birthplace of America’s plutonium production program — targets up to 960 megawatts across multiple SMR units, with the first phase (320 MW from four Xe-100 reactors) aimed at operations by the end of the decade. The NRC pre-application review is underway, with regulatory clearance anticipated by the end of 2026. AWS is not buying electricity from a nuclear utility. It is financing the construction of its own nuclear fleet — a strategic move that would have been unthinkable for a technology company five years ago and that is now the defining competitive dynamic of the hyperscale industry.

The Boomer-Millennial Synthesis

Why Nuclear Is the Perfect Handoff technology

The SMR renaissance is the purest expression of the Synthesis principle established in Book 1: the fusion of Inert Assets (Boomer-era engineering knowledge) with Active Logic (Millennial-era AI-driven optimization). No other technology sector illustrates this dynamic as clearly, because no other sector possesses such a vast reservoir of proven but underleveraged institutional knowledge.

Nuclear engineering is the quintessential Boomer achievement. The physics of controlled fission was established in the 1940s and 1950s by Fermi, Wigner, Weinberg, and their colleagues at the Manhattan Project’s successor institutions. The reactor designs were proven through decades of naval and civilian operation — the US Navy has operated nuclear-powered vessels since 1955, accumulating over 6,800 reactor-years of operating experience without a single reactor accident. The materials science, the thermal hydraulics, the neutronics calculations, the regulatory frameworks, and the specialized engineering workforce — all are part of a massive historical investment that is being unlocked by AI.

AI-driven design optimization, predictive maintenance based on synthetic data, and simulation-based licensing are compressing the development cycle for SMRs. We are witnessing a synthesis where the physical resilience of legacy engineering meets the agility of digital logic.

External Citations

1. NRC — Advanced Reactors (SMR) Overview: The U.S. Nuclear Regulatory Commission’s official page on advanced reactor and small modular reactor design reviews, covering the licensing frameworks for Oklo, NuScale, and X-energy referenced throughout the SMR Generation section. https://www.nrc.gov/reactors/new-reactors/smr.html
2. X-energy — Xe-100 Reactor: X-energy’s official product page for the Xe-100 high-temperature gas-cooled reactor, providing technical specifications for the TRISO fuel, 80 MWe per module output, and the Cascade Nuclear Energy Center deployment details covering Amazon/AWS’s nuclear partnership. https://x-energy.com/reactors/xe-100
3. EIA — U.S. Nuclear Industry: The U.S. Energy Information Administration’s overview of the American nuclear fleet, documenting capacity factors above 90%, operating reactor counts, and recent construction activity including the SMR pipeline, providing the factual baseline for the nuclear logic arguments in this chapter. https://www.eia.gov/energyexplained/nuclear/us-nuclear-industry.php

Previous: ← Chapter 2 (book 2) | Navigation (book 2) | Next: Chapter 4 (book 2) — The Copper Veto →