Two energy and computing companies have agreed to try something the power industry has talked about for years but never built: an artificial-intelligence data center wired directly to a small nuclear reactor rather than the public grid. The plan, announced at the end of July 2026, pairs a Texas reactor developer with an infrastructure firm that specializes in dropping modular computing halls next to their own power sources. Their stated goal is a working proof of concept before the decade’s midpoint.
The arrangement reflects a squeeze that has become the defining constraint of the AI build-out. Training clusters draw enormous, spiky loads that strain regional grids and can wait years for new transmission, so operators have begun hunting for dedicated generation they can site alongside the servers. Nuclear appeals because it delivers large amounts of steady, around-the-clock power without direct carbon emissions, and a compact reactor can, in theory, be planted in the same industrial yard as the machines it feeds.
Who is behind the deal
The partnership brings together Aalo Atomics, an Austin-based reactor startup, and Crusoe, a company that builds and operates AI-optimized data centers close to their energy supply. In a joint announcement, the two described their aim as deploying what they call the first nuclear-powered “AI factory,” a phrase the industry now uses for a data center built specifically to churn out AI training and inference at scale.
The companies framed the agreement as a commercial partnership rather than a research memorandum, with a defined hardware roadmap and a demonstration site already identified. According to the formal release issued July 30, the near-term work centers on a single modular computing unit powered by a single reactor, structured so that lessons from the first pairing can be copied across many more.
How the demonstration would work
The first step, planned for 2027, would place one of Crusoe’s Spark modular data centers at Idaho National Laboratory and run it on power drawn directly from an Aalo reactor. The demonstration is meant to test a specific and technically awkward question: whether a compact reactor can follow the erratic, fast-changing electrical demand that AI training clusters impose, rather than the smooth baseload profile a traditional plant is designed to serve.
The reactor at the center of the plan is the Aalo Pod, described in coverage of the deal as a 50-megawatt-electric power plant assembled from five independent 10-megawatt sodium-cooled microreactors, all driving a single turbine, per a report in Power Magazine. Cooling the reactor with sodium and air rather than water is a deliberate choice, because it frees the design from needing a large local water source, a growing concern as data centers cluster in dry regions.
Why the timing lines up
The partnership did not appear out of nowhere. Aalo had recently reached a technical milestone at the Idaho site, and the company’s decision to move straight from a test reactor into a commercial computing pairing signals confidence that its design can scale. Trade coverage of the agreement noted that the choice of Idaho National Laboratory keeps the first deployment on federal ground already set up for advanced-reactor work, which shortens the regulatory and siting path compared with a greenfield commercial location, according to an account from Data Center Dynamics.
The broader logic is economic as much as technical. Grid interconnection queues in several regions now stretch for years, and utilities in AI hot spots have warned that new large loads could raise electricity prices for existing customers. A reactor sited on the same parcel as the servers sidesteps the transmission bottleneck entirely, at least in principle, and lets an operator promise developers a fixed, self-contained power supply instead of a spot on a waiting list.
The gap between plan and plant
A demonstration is not a fleet, and the roadmap the companies laid out makes the distance clear. Beyond the 2027 proof of concept, the pair said they intend to deploy Aalo Pods and larger 50-megawatt “extra-modular” reactor plants at Crusoe data centers by the end of 2029, a timeline that assumes the first pairing works and that regulators clear the follow-on units. Advanced-reactor projects have a long history of slipping past their original dates, and pairing a first-of-a-kind reactor with a first-of-a-kind load profile compounds the uncertainty.
There are also open questions the announcement did not settle. Licensing a commercial reactor to sell power to a private data center, rather than to run as a laboratory experiment, involves a separate and more demanding regulatory review. The economics of building bespoke reactors for individual computing campuses remain unproven against the alternative of simply buying grid power or natural-gas generation, and the sodium-cooled design the plan relies on has a mixed operating record in earlier reactor programs that will draw scrutiny.
What the deal does establish is momentum. A concrete site, a named reactor, a partner willing to put its computing hardware on the line, and a dated demonstration together move the idea of reactor-adjacent AI from slide decks toward a physical test. Whether the pairing proves that a small reactor can chase the jagged appetite of an AI cluster, or exposes why that has never been done before, the 2027 run at the Idaho lab should supply the first real evidence either way. Until then, the ambition outruns the operating history, and the industry will be watching a single modular unit in the high desert for the answer.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
- Certain older cars are stolen far more often than any brand-new model
- A handful of car transmissions are so tough mechanics say they almost never fail
- Toyota’s refreshed electric SUV now promises up to 314 miles between charges
- Nissan just halted sales of 168,149 vehicles over labels that overstate what they can carry
