## Has a Microreactor Actually Powered an AI Computer Yet?

Yes. Valar Atomics' Ward-250 microreactor in Utah has powered an NVIDIA DGX Spark AI computer using nuclear-generated electricity — the first publicly reported demonstration of a [High Temperature Gas-Cooled Reactor](https://smrintel.com/glossary/htgr) directly supplying an AI compute load. The DOE confirmed on June 18, 2026 that the Ward-250 had achieved a self-sustaining nuclear chain reaction — [criticality](https://smrintel.com/glossary/criticality) — at the Utah facility. The subsequent demonstration ran the reactor at approximately 37% of rated capacity, producing 100 kW of thermal energy, which was then converted to electricity and directed to the NVIDIA device. The DGX Spark unit, built around NVIDIA's GB10 Grace Blackwell chip, draws approximately 240 W — a modest but deliberate proof-of-concept load. Valar Atomics is separately targeting a 30 MW commercial project in Utah, though NRC licensing for that scale remains unresolved.

This is an engineering milestone worth tracking, but the gap between a 100 kWth test and a licensed, grid-connected commercial facility is vast. The commercial pathway depends entirely on NRC authorization that has not yet been granted.

---

## What Is the Ward-250 and How Does It Work?

The Ward-250 belongs to the [High Temperature Gas-Cooled Reactor](https://smrintel.com/glossary/htgr) family — a reactor class with a decades-long pedigree stretching from the German THTR-300 to China's HTR-PM now operating at Shidaowan. The design uses [TRISO](https://smrintel.com/glossary/haleu) fuel — tristructural isotropic fuel particles encased in multiple ceramic and carbon layers — a graphite moderator, and helium as the primary heat transfer medium.

TRISO fuel's defining characteristic is its inherent retention of fission products at high temperatures, which underpins the passive safety case for HTGRs. The ceramic coating layers are designed to maintain integrity even under accident conditions without active intervention — a critical selling point for [behind-the-meter generation](https://smrintel.com/glossary/behind-the-meter) applications where round-the-clock human operation may be limited.

During the Utah demonstration, the reactor ran at roughly 100 kWth — approximately 37% of its rated thermal output, per the source. Thermal energy was converted to electricity through a converter before reaching the NVIDIA load. The source does not specify the conversion technology or the electrical output figure, so efficiency numbers cannot be calculated from available data.

---

## The NVIDIA Angle: Why This Partnership Matters for AI Infrastructure

NVIDIA's participation signals that at least one major compute hardware vendor is willing to put its brand alongside nuclear microreactor demonstrations. The DGX Spark unit used in the test is a compact AI workstation — not a hyperscale rack — which makes it an appropriate proof-of-concept load for a sub-megawatt reactor.

The arithmetic illustrates the scaling challenge directly: the source notes that powering a large data center operating at 30 MW would require over a hundred thousand DGX Spark-class devices. That framing underscores why Valar Atomics' own roadmap points toward a 30 MW facility rather than remaining at kilowatt scale. For hyperscale AI operators, microreactors only become relevant at tens of megawatts of continuous, licensed output — not at demonstration wattage.

NVIDIA is simultaneously developing liquid cooling systems designed to operate at inlet temperatures up to 45°C, which the source says enables near-zero water consumption through dry cooling towers. That thermal compatibility with HTGR outlet temperatures is architecturally significant: high-temperature gas reactors can in principle supply heat directly usable by advanced cooling loops, potentially improving overall system efficiency compared to coupling a reactor to legacy water-cooled data center infrastructure.

---

## The Regulatory Gap Nobody Should Ignore

The single biggest constraint on Valar Atomics' commercial ambitions is NRC licensing. The source explicitly flags that "many questions remain open regarding the commercial licensing and industrial-scale expansion of the project" and identifies NRC authorization as the key condition for scaling.

The DOE confirmation of criticality on June 18 is a physics milestone — it confirms the reactor design can sustain a chain reaction. It is not a commercial operating license, a construction permit for a 30 MW facility, or even a completed safety review. The NRC's process for novel microreactor designs — particularly under the evolving Part 53 framework — involves design certification, site licensing, and environmental review steps that collectively take years even for well-resourced applicants.

Other microreactor developers including [Oklo Inc.](https://smrintel.com/companies/oklo) and [Kairos Power](https://smrintel.com/companies/kairos-power) have been navigating NRC processes for years and have not yet reached commercial operation. Valar Atomics is at an earlier stage. The Utah demonstration is a [FOAK](https://smrintel.com/glossary/foak) engineering proof point, not a commercial product.

The nuclear-AI data center thesis is structurally sound — AI compute demands high-capacity-factor, location-flexible power that nuclear can theoretically provide. But the timeline between a 100 kWth test in 2026 and licensed megawatt-scale nuclear supply to a commercial data center is measured in years, not months.

---

## Industry Trajectory

The Valar Atomics demonstration adds a data point to an emerging pattern: nuclear startups are increasingly positioning microreactors as dedicated power sources for AI compute rather than grid-connected baseload generators. This [behind-the-meter](https://smrintel.com/glossary/behind-the-meter) model sidesteps some grid interconnection complexity but introduces a different regulatory challenge — siting a nuclear facility on or adjacent to a private data center campus.

HTGR technology with TRISO fuel has a credible safety and operational basis, which is part of why multiple developers have converged on it for early microreactor applications. The passive safety characteristics reduce the operational burden that makes nuclear impractical for small, remotely managed facilities.

What the industry needs next from Valar Atomics is not another demonstration — it is a filed NRC license application with a defined review schedule. Until that exists, the 30 MW Utah project remains a planning ambition rather than a committed development.

---

## Key Takeaways

- Valar Atomics' Ward-250 HTGR achieved criticality on June 18, 2026 per DOE confirmation, and subsequently powered an NVIDIA DGX Spark AI workstation in Utah at approximately 100 kWth (37% of rated capacity).
- The Ward-250 uses TRISO fuel, a graphite moderator, and helium coolant — a design architecture with strong passive safety credentials relevant to unmanned or lightly staffed data center deployments.
- The NVIDIA DGX Spark device used in the demonstration draws approximately 240 W; the company is separately developing liquid cooling systems operable at inlet temperatures up to 45°C.
- Valar Atomics is targeting a 30 MW commercial facility in Utah, but NRC licensing authorization — the critical path item — has not been obtained.
- The demonstration is an engineering proof point. Commercial nuclear power for AI data centers requires licensing steps that no microreactor startup has yet completed.

---

## Frequently Asked Questions

**What is the Valar Atomics Ward-250?**
The Ward-250 is a microreactor designed by Valar Atomics that uses TRISO fuel, a graphite moderator, and helium as a coolant — placing it in the High Temperature Gas-Cooled Reactor (HTGR) category. It achieved a self-sustaining nuclear chain reaction (criticality) confirmed by the DOE on June 18, 2026, at a Utah facility.

**Has a nuclear reactor actually powered an AI data center?**
In a limited proof-of-concept sense, yes. Valar Atomics connected the Ward-250 to an NVIDIA DGX Spark AI workstation during a test in Utah, running the reactor at approximately 100 kWth. This is not a commercial data center — it is a controlled demonstration at a fraction of the reactor's rated capacity.

**When will microreactors commercially power AI data centers?**
No microreactor developer has yet received a commercial operating license from the NRC in the United States. Valar Atomics has not publicly filed for NRC design certification or a construction permit for its planned 30 MW Utah project. Commercial operation is realistically a multi-year process from the current stage.

**Why are HTGR designs favored for AI data center applications?**
TRISO fuel and helium-cooled HTGR designs offer high-temperature heat output and passive safety characteristics — the reactor cannot melt down in the traditional sense because the fuel itself retains fission products at extreme temperatures. This reduces the operational complexity of siting a nuclear plant adjacent to a data center with limited on-site nuclear expertise.

**What is the NRC's role in microreactor licensing?**
The Nuclear Regulatory Commission must authorize any commercial nuclear facility in the United States. For novel designs like the Ward-250, this involves design certification review, site-specific licensing, and environmental assessment. The agency is developing Part 53, a new licensing framework intended to streamline advanced reactor approvals, but no microreactor has completed that process to commercial authorization yet.