Executive summary. Generative artificial intelligence (AI) has collided with the physical grid and forced a structural shift in tech infrastructure. Intermittent renewables and lithium-ion storage struggle to economically firm gigawatt-scale data centers that target 99.999% uptime. The response has been an accelerating, privately funded renaissance in nuclear fission. Hyperscalers are moving from signing conventional power purchase agreements (PPAs) on public grids to backing small modular reactors (SMRs) for captive, islanded compute. But this atomic pivot introduces new bottlenecks: conversion and enrichment capacity. In particular, high-assay low-enriched uranium (HALEU) is constrained, and the geopolitical race for AI capability is increasingly linked to nuclear fuel security. Over the next decade, as fission scales through the 2030s, hyperscaler capital will also intensify the endgame: commercial fusion.

For two decades, nuclear energy looked like a stagnant industry: slow permitting, cost overruns, and political fragility. It was treated as legacy baseload, not a growth platform. That paradigm fractured the moment a hyperscaler discovered it could not build a frontier training site because the local utility needed years to deliver interconnection and transmission upgrades.

The category error of the early 2020s was assuming the AI buildout could be powered entirely by wind, solar, and lithium-ion storage. Renewables remain the engine of broad grid expansion. But data centers do not follow the sun. They are flat, unrelenting, 24/7, high density loads. If you try to manufacture firm power purely from intermittent generation, you pay for layers: overbuild, storage duration, transmission, curtailment, and complex reliability engineering. When you strip away the marketing and look at grid physics, one technology is uniquely dense, weather-independent, and zero carbon at the point of generation: nuclear.

1. The SMR renaissance. Hyperscalers are increasingly unwilling to let public grid queues dictate compute roadmaps. The first adaptation wave focused on legacy nuclear output: co-locating near existing large plants, or extending and uprating current fleets. That buys time, but it is finite. The scalable solution is the SMR: smaller standardized reactor modules, typically tens to a few hundred megawatts, designed to move construction from bespoke site megaprojects toward repeatable manufacturing.

Co-locate clusters of SMRs adjacent to data centers and a new architecture appears: islanded compute with firm electrons at the fence line. You bypass interconnection queues, reduce transmission exposure, and treat power as an owned asset instead of an input. This is why the market is shifting from generic clean energy procurement to nuclear-specific execution. Meta’s recently announced set of agreements, framed as unlocking up to 6.6 gigawatts of capacity by 2035, is strategic capacity reservation in the only firm clean technology that can plausibly scale to the load profile of frontier AI. (Meta)

Timelines still matter. Meaningful direct-wired SMR deployments for large data center loads are more plausibly early 2030s than late 2020s. That means legacy restarts and co-location serve as the bridge. The Crane Clean Energy Center restart, expected in 2027 at 835 megawatts, is emblematic of the bridge strategy: take an existing site, accelerate return to service, and lock in clean firm output. (Constellation)

Momentum is also visible in advanced nuclear contracting. Google’s arrangement with Kairos Power is best understood as a master agreement that enables up to a 500 megawatt SMR fleet, rather than a single project. The initial announcement dates to October 2024, and subsequent execution updates signal that the development stack is maturing toward early 2030s first units. (Kairos Power)

But capital formation is happening now. Hyperscalers are underwriting order books for advanced nuclear developers, effectively doing what governments struggled to do: create durable demand signals that can finance factories, supply chains, and standardized deployment.

2. The fuel bottleneck: is there enough. If compute requires nuclear, and nuclear requires fuel, does the AI wave run out of atoms?

The short answer is no. Uranium as a raw resource is broadly available, including in politically stable jurisdictions. The constraint is industrial capacity: conversion, enrichment, and fabrication.

Many advanced SMR designs rely on HALEU. Higher enrichment supports smaller cores, longer cycles, and different reactor physics choices. For decades, much of the commercial enrichment and HALEU supply chain effectively sat outside the West. The result is that the near term bottleneck is enrichment throughput and the ability to produce the right product specification at scale.

This is why the DOE’s 2.7 billion enrichment awards matter. They are a forced reindustrialization of the nuclear fuel cycle, with major task orders to rebuild domestic enrichment services and jumpstart HALEU supply chains. (DOE)

Even with aggressive funding, the gap is real. Enrichment constraints in the 2026 to 2028 window can still delay first wave deployments of next generation reactors, especially those depending on HALEU. The fuel narrative is often framed as scarcity of uranium rock. The more precise framing is scarcity of fuel cycle capacity and qualified production lines.

3. The geopolitics of atomic AI. AI capability is the economic and military high ground of the century. If the training frontier requires nuclear, then fuel security becomes industrial policy. You cannot dominate compute while outsourcing the inputs that make firm power possible. Expect tighter coupling between hyperscalers and national security establishments, and aggressive incentives for friend-shoring the full fuel cycle: mining, conversion, enrichment, fabrication, and transport security.

Countries with vertically integrated nuclear export capability gain leverage. They can offer bundled infrastructure: data centers plus reactors plus fuel services. That package turns compute into a geopolitical export product. It also raises pressure on non-proliferation regimes, because the economic incentive to expand nuclear materials capability will rise with the strategic value of AI.

4. The fusion horizon. Even if SMRs scale, fission carries long run constraints: waste, security, and regulatory load impose ceilings. But intelligence is an infinite demand sink. Lower compute cost and society will consume more of it. That is why hyperscaler capital, after funding fission scale out through the 2030s, will increasingly target the final frontier: commercial fusion.

Fusion remains longer horizon. Expect meaningful grid contribution post 2040. But it is no longer a purely public science bet. It is becoming a private capital race, and AI itself is a force multiplier in simulation, materials discovery, and control systems.

The investable logic. The atomic pivot reshapes the energy infrastructure trade.

The long book sits in the physical supply chain: Tier one uranium producers in allied jurisdictions; conversion and enrichment buildout; specialized nuclear certified engineering, procurement, and construction (EPC); and advanced nuclear vendors whose designs align with manufacturable deployment. Hyperscalers that successfully execute captive nuclear strategies structurally decouple compute cost from public energy inflation, turning power into a moat.

The short book sits where firm power is unpriced until it is suddenly priced: unhedged data center operators exposed to strained public grids; software layers with weak pricing power and rising compute cost of goods; and renewable developers without firming attachments, who cannot capture the firmness premium demanded by high availability AI loads.

The physical reality remains unbroken. Models get larger. Racks get hotter. Grids get tighter. The next era of software dominance is increasingly underwritten by the atom.