Valar Atomics' Ward 250 MW small modular reactor powers Nvidia Blackwell data center, proving nuclear viability for AI infrastructure.
On July 1, 2026, California-based nuclear startup Valar Atomics staged a live demonstration in Emery County, Utah, connecting its Ward 250 test reactor to an Nvidia RTX Spark desktop PC built on Nvidia's Blackwell architecture. The reactor's electricity, converted from thermal energy via a thermal-electric generator, briefly powered a temporary website—marking the first time a US advanced reactor has supplied power to run an AI chip. The Ward 250 had reached criticality on June 18, 2026, and by the demonstration was producing approximately 100 kilowatts of thermal output.
The reactor itself is a High-Temperature Gas-Cooled Reactor using TRISO-coated fuel particles and helium coolant. Despite its name, the "250" designation refers to a program model rather than a capacity rating; the unit was built for an initial test-power target of around 100 kilowatts thermal, with a design capable of scaling toward roughly 5 megawatts electric. At the July 1 demonstration, Valar CEO Isaiah Taylor confirmed the core was producing about 100 kilowatts of thermal energy.
Valar and Nvidia announced a joint feasibility study exploring designs for a 30-megawatt computing facility employing closed-loop, water-free cooling—eliminating dependence on local municipal water supplies, a critical constraint for hyperscale AI data centers.
For AI infrastructure teams, the significance lies not in the demonstration's wattage. A single test reactor producing roughly 100 kilowatts thermal is trivial next to hyperscale power demand. Rather, the demonstrated pairing of an advanced reactor with closed-loop, water-free GPU cooling—which Valar and Nvidia say they will study at 30-megawatt scale—constitutes a concrete signal that on-site nuclear generation for AI compute is advancing from theoretical to physically demonstrated, years ahead of any commercial licensing.
The demonstration combines three distinct subsystems relevant to data-center planning: an HTGR heat source, a thermal-to-electric conversion chain, and Nvidia's closed-loop liquid cooling for GPUs. Integrating them raises systems-engineering and site-operations questions beyond purely algorithmic challenges. For teams modeling long-term power and water constraints on AI infrastructure, this represents an early but concrete data point that on-site advanced-reactor generation paired with water-free cooling is technically feasible at small scale—not that it is commercially available. Open questions include whether follow-up tests scale steady electrical output beyond the demonstration's kilowatt-level, any site selection or finalized designs for the 30-megawatt facility, regulatory filings or NRC licensing milestones, and further technical disclosure on the heat-to-electric conversion chain.