A tiny reactor built for AI just achieved nuclear criticality

A compact test reactor designed to power artificial-intelligence computing has reached the moment that separates a machine on paper from a working nuclear plant: a self-sustaining chain reaction. The startup behind it, Austin-based Aalo Atomics, hit that milestone at a federal laboratory in Idaho over the July 4 holiday and has since tied the achievement to a commercial plan to run an AI data center on the same technology. The pairing of a physics milestone with a computing customer is what has kept the small reactor in the news through the start of August 2026.

Reaching criticality is a defined technical threshold rather than a marketing claim. It is the point at which a reactor’s chain reaction becomes self-sustaining, with each generation of neutron-splitting events triggering the next without outside help. Crossing it demonstrates that a design’s core physics work as intended and clears the way for the testing that must precede any real power production.

What Aalo achieved and when

The company reached criticality with its Aalo-X Critical Test Reactor in the early morning hours of July 4, 2026, at Idaho National Laboratory. The reactor went from groundbreaking in September 2025 to a sustained chain reaction in under eight months, an unusually fast build for a nuclear project, according to a report in Power Magazine. The speed reflects both a deliberately simple design and a testing pathway set up to move quickly.

The reactor is small by the standards of the industry. The company’s commercial design, the Aalo Pod, is a 50-megawatt-electric plant assembled from five independent 10-megawatt microreactors driving a single turbine, cooled by sodium and air rather than water. Dispensing with water cooling is a design choice aimed squarely at data centers, which increasingly land in dry regions where a large water draw draws local opposition.

The federal program behind the deadline

Aalo’s July 4 milestone did not happen on its own schedule alone. It was the fourth reactor to reach criticality under a government initiative that had set that holiday as an explicit target, a coordination effort the Department of Energy has documented through the specialists who helped shepherd the reactors to the milestone, as described by the agency itself. The program named a slate of advanced-reactor projects in the summer of 2025 and challenged at least three of them to reach criticality within roughly a year.

Four cleared the bar by the deadline. Alongside Aalo-X, the reactors that reached criticality were Antares Nuclear’s Mark-0, Valar Atomics’ Ward250, and Deployable Energy’s Unity, per World Nuclear News. Grouping several first-of-a-kind designs under a single deadline was a deliberate bet that competition and a shared testing site could compress timelines that normally stretch across many years.

Why AI is driving the push

The urgency behind these small reactors traces almost entirely to the electricity appetite of AI. Training and running large models demands vast, steady power, and data-center growth has begun to strain regional grids and lengthen the wait for new transmission. Nuclear appeals as an answer because it supplies large amounts of around-the-clock power without direct carbon emissions, and a compact reactor can, in principle, sit on the same industrial parcel as the servers it feeds instead of competing for a spot in a years-long grid queue.

Aalo has moved to make that link concrete. The company has agreed to pair its reactor with a modular data center from Crusoe, an AI-infrastructure firm, in a proof-of-concept deployment planned for the Idaho lab in 2027, as reported by Tech Times. That plan would test whether a small reactor can follow the erratic, fast-swinging electrical load that AI training clusters impose, a demand profile very different from the smooth baseload a traditional plant is built to serve.

What criticality does and does not prove

A sustained chain reaction is a genuine milestone, but it is an early one. Criticality shows that a reactor’s core can hold a stable reaction; it does not by itself demonstrate that the plant can generate electricity reliably for years, ramp its output to chase a data center’s swings, or operate safely and economically at commercial scale. Each of those steps requires its own testing, and each has tripped up advanced-reactor efforts before.

The sodium-cooling approach that lets Aalo’s design skip water also carries baggage. Sodium-cooled reactors have a long research history but a mixed operating record, and handling a coolant that reacts violently with air and water introduces engineering demands that regulators will examine closely as any design moves from a test unit toward a licensed commercial plant. Selling reactor power to a private data center, rather than running an experiment on federal ground, triggers a more demanding regulatory review than the test-reactor phase.

Even so, the milestone marks real forward motion for a technology that has spent decades closer to promise than to hardware. A working chain reaction, a fast build, a federal deadline met, and a named computing customer together move the idea of reactor-powered AI from concept toward demonstration. The decisive evidence will come from the 2027 pairing at the Idaho lab, where a single small reactor and a single data center will show whether compact nuclear can actually keep pace with the jagged, growing hunger of artificial intelligence, or expose the reasons it has never been tried at scale before. For now, the physics have been proven; the far harder questions of cost, reliability, and regulation remain ahead.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview