Valar Atomics ran an Nvidia AI chip on nuclear power, a US first for next-generation reactors

Isaiah Taylor, chief executive of Valar Atomics, made a precise claim: his company powered the first AI chip ever run directly from a nuclear reactor, and hosted a website on that chip. On July 2, 2026, independent reporting from Bloomberg, Fox13, Tom’s Hardware, Interesting Engineering, and Heise confirmed the event. At Valar Atomics’ Utah facility, the company’s Ward 250 reactor supplied power to an Nvidia Blackwell AI chip. The coverage described it as a US first for a next-generation nuclear reactor powering an AI chip directly.

Taylor’s framing reaches further than the external record strictly confirms. The corroborated claim is a US first, in the category of next-generation reactors. Whether no similar connection has occurred anywhere else in the world is not something the available reporting establishes. That distinction matters for accuracy, though it does not diminish what the Utah demonstration actually represents: a functioning nuclear plant, built and operated by a startup, delivering power to the class of chip now at the center of the global AI buildout.

The choice of an Nvidia Blackwell chip is not incidental. Blackwell is Nvidia’s current flagship AI accelerator architecture, the hardware that hyperscalers and frontier AI labs are deploying at scale in 2026. Connecting one directly to a reactor output, rather than drawing from a grid that might include nuclear somewhere upstream, is the specific claim Valar Atomics is making. The website hosted on that chip is a small detail with a clear rhetorical purpose: it demonstrates that the system was live and functional, not a staged photo opportunity with a chip sitting near a reactor.

We powered uh the first ever AI chip powered by a nuclear reactor. Isaiah Taylor

The energy context surrounding this demonstration has been building for years. Large AI training runs and inference workloads at scale demand consistent, high-density power that renewable sources alone cannot yet guarantee around the clock. Data center developers have begun signing agreements with nuclear operators for exactly this reason, but those arrangements draw from existing grid infrastructure. What Valar Atomics demonstrated is a shorter path: a purpose-built reactor, a startup’s own facility, a direct connection to AI compute hardware.

The Ward 250 designation refers to the reactor’s design. Valar Atomics is among a cohort of companies attempting to commercialize small modular or microreactor designs outside the decades-long licensing and construction timelines associated with conventional nuclear plants. Demonstrating an actual power delivery milestone, confirmed by multiple outlets, moves the company’s claims from the speculative category into the operational one. That transition is significant for investors, regulators, and potential customers trying to assess which nuclear startups are building real systems versus refining slide decks.

What Taylor’s company has shown is that the infrastructure side of the AI energy problem has at least one new entrant capable of delivering working hardware. The policy, regulatory, and scaling questions that surround next-generation nuclear remain unresolved, and a single chip demonstration at a Utah facility is not a grid solution. But the corroborated record of July 2, 2026 establishes that the connection between new nuclear and AI compute is no longer purely theoretical, at least in the United States. The harder work of scaling that connection to the power levels the industry actually needs comes next.