Valar Atomics vs Radiant: What Nuclear Microreactors Mean for Your AI Power Strategy
Mark aistrategy.guide as a preferred source Our AI strategy guides show up more often, with a Preferred badgeValar Atomics and Radiant are the two microreactor companies executives ask about most. Valar's Ward 250 went critical in Utah in June 2026; Radiant's 1 MWe Kaleidos is at Idaho National Laboratory for testing and holds a US Army order. Neither will power a production AI cluster before about 2027-28. What that means for site selection and power contracts this planning cycle.
The question arrives in some version at most infrastructure reviews this year: should our AI power plan include a small nuclear reactor? It is a fair question. Two venture-backed reactor companies hit public milestones in 2026, and both say AI is a target market. But the honest planning answer is narrower than the headlines. No microreactor will power a production AI cluster before roughly 2027-28, and the first units will be small relative to AI campus loads. What the reactor news does tell an executive is that power has become a site-selection constraint, and that is the part to act on now.
Short answer, as of 20 October 2026: Valar Atomics is further along on physics (its Ward 250 test reactor went critical on 18 June 2026 and produced electricity on 1 July) and is courting AI loads directly, including a planned 30 MWe AI facility in Utah with NVIDIA. Radiant has the stronger near-term customer (a US Army award worth up to $750 million for 15 of its 1 MWe Kaleidos units) but had not publicly announced criticality. Neither is a power purchase agreement you can sign for a 2026 or early-2027 build.
The engineering version of this topic (reactor design, heat, integration with a data center) is on ctaio.dev: nuclear microreactors for AI data centers. This page is the buyer’s view.
Why is an AI strategy page talking about reactors at all?
Because electricity is now the binding constraint on where AI capacity gets built. The US Department of Energy’s summary of the 2024 Lawrence Berkeley National Laboratory report put data centers at about 4.4% of US electricity use in 2023 and projected 6.7% to 12% by 2028. Grid interconnection queues, transformer lead times and local permitting now decide when a new site can energise, more than server delivery does.
That changes the conversation with any provider. A neocloud or colocation partner’s capacity roadmap is only as good as its power contracts. Asking a provider where its next megawatts come from, and on what date, is now a standard diligence question. Reactors are one answer some providers will give; most near-term answers are gas, grid upgrades and restarted or uprated conventional nuclear plants.
Valar Atomics vs Radiant at a glance
| Valar Atomics | Radiant | |
|---|---|---|
| Reactor | Ward 250: helium-cooled, graphite-moderated, TRISO fuel | Kaleidos: helium-cooled, TRISO fuel, factory-built |
| Size | Test unit 100 kWt; commercial unit designed for up to 5 MWe | 1 MWe, truck-transportable |
| 2026 milestone | Zero-power criticality 18 June; electricity 1 July (Utah) | Shipped to INL’s DOME test facility, August |
| Criticality status (20 Oct 2026) | Achieved | No public announcement found |
| Anchor customer signal | Planned 30 MWe AI facility in Utah with NVIDIA, feasibility study funded | US Army Janus award, up to $750M for 15 units |
| Recent capital | $1B Series B (Sequoia, Aug 2026) + $200M credit facility; about $450M earlier in 2026 at a reported $2B valuation | Over $300M Series D (Dec 2025) |
| Commercial NRC licence | Not publicly filed as of Aug 2026 | Expects NRC licences before delivering units |
Valar Atomics: physics first, AI as the stated market
Valar broke ground on the Ward 250 and an on-site TRISO fuel facility at the Utah San Rafael Energy Lab in Emery County in September 2025. Under the Department of Energy’s Reactor Pilot Program, the test reactor reached zero-power fuelled criticality on 18 June 2026 and generated electricity on 1 July, inside the 4 July 2026 deadline set by Executive Order 14301. Trade press described it as the first DOE-authorised reactor to be built and operated outside a national laboratory. It was the second program reactor to go critical, after Antares Nuclear’s Mark-0 at Idaho National Laboratory.
The AI link is explicit. Valar connected an NVIDIA desktop system directly to the Ward 250 as a proof of concept, and says it plans a 30 MWe AI facility in Utah with NVIDIA; its August 2026 funding includes an NVIDIA feasibility study. That funding was large: a $1 billion Series B led by Sequoia Capital, plus a $200 million credit facility led by Erebor Bank and JP Morgan. Earlier in 2026, Bloomberg reported a raise of about $450 million ($340 million equity, $110 million debt) at a $2 billion valuation.
Two cautions for a buyer. First, a 100 kWt test reactor is not a commercial plant; the commercial unit is designed for up to 5 MWe, and a 30 MWe facility implies several of them. Second, as of August 2026 Valar had not publicly filed a commercial licence application with the NRC, and it is part of a lawsuit challenging NRC authority over certain small reactors. The regulatory path to selling power to a third party is still open.
Radiant: the factory model and a military customer
Radiant’s Kaleidos is a 1 MWe unit designed to be built in a factory and moved by truck. It left Radiant’s El Segundo facility in mid-August 2026 for the DOME test bed at Idaho National Laboratory, where the company plans a five-phase test campaign ending in an unattended run of at least 150 hours at full power. As of 20 October 2026 we found no public announcement that Kaleidos had reached criticality.
The demand signal is strong. On 26 August 2026 Radiant announced that the US Army, working with the Defense Innovation Unit, signed an agreement worth up to $750 million for 15 Kaleidos units, the largest award in the Army’s Janus program, starting with three units at Fort Benning. Janus covers five reactor developers at five bases, with a target of sustained operation by 30 September 2028. Radiant raised more than $300 million in a Series D in December 2025, led by Draper Associates and Boost VC, partly to build a factory in Oak Ridge, Tennessee.
For an enterprise buyer, the Army contract matters as a production signal: a factory model only gets cheaper with volume, and a government order is the volume. It does not put a Kaleidos on your campus sooner. Military bases are the first customers.
What changed in regulation, and what did not
The Nuclear Regulatory Commission’s Part 53, a risk-informed, technology-inclusive licensing framework for new reactors, took effect on 29 April 2026 under a final rule published on 30 March 2026 (91 FR 15696). The DOE’s pilot program met its goal: four reactors (Antares, Valar, Deployable Energy and Aalo) reached criticality by 4 July 2026, against a target of three.
What did not change: a DOE-authorised test is not a commercial licence to sell electricity to a data center operator. Each commercial deployment still needs NRC licensing, fuel supply (TRISO fuel capacity is itself being built out) and local approvals.
What should an executive do with this?
Treat microreactors as a 2028-plus option on your power roadmap, not a 2026-27 supply. The practical moves this quarter are about power, not nuclear:
- Add power to provider diligence. For every colocation or neocloud contract, ask for the energisation date of the capacity you are buying and the source of its power. Capacity without a power date is a forecast. The infrastructure strategy framework covers the contract terms.
- Put power into site selection. If you are building or leasing your own capacity, grid interconnection and utility timelines should be screened before tax incentives and land.
- Size the claim. A single 1 MWe or 5 MWe unit supports a small inference footprint, not a training campus. Vertiv’s reference design for NVIDIA’s GB200 NVL72 assumes racks of up to 132 kW, so one megawatt goes quickly once cooling is counted.
- Watch three milestones. Kaleidos criticality and its 150-hour run at INL; Valar’s first commercial licence filing with the NRC; and first power at a Janus base. Those three tell you whether 2028 holds.
The wider supplier landscape, including the utilities and conventional nuclear deals hyperscalers have signed, is in AI infrastructure companies: the stack, layer by layer.
The verdict
Valar has proven its physics first and is pitching AI loads directly; Radiant has the more bankable customer and a factory model. Both are worth tracking, and neither belongs in a 2026-27 capacity plan. The decision for this planning cycle is to treat power as a constraint on every infrastructure contract you sign.
Sources & methodology
- Nuclear Engineering International: Valar raises $1bn for microreactors (7 Aug 2026): Series B, credit facility, Ward 250 specifications, criticality and electricity dates, NVIDIA plans, licensing status
- World Nuclear News: Valar Atomics achieves criticality in DOE Reactor Pilot Program (22 Jun 2026)
- Nuclear Engineering International: Valar’s Ward 250 achieves criticality (25 Jun 2026)
- Bloomberg Law: Luckey-backed nuclear startup Valar lands $2 billion valuation (31 Mar 2026)
- Radiant: Kaleidos has left the building (Aug 2026)
- Radiant: Janus program award announcement and Bloomberg: Radiant wins $750 million US Army nuclear reactor deal
- American Nuclear Society: Five companies, five bases, the Army’s Janus Program takes shape (28 Aug 2026)
- Radiant: Series D announcement (17 Dec 2025)
- US NRC: effective dates for sections of 10 CFR in 2026 (Part 53, 91 FR 15696)
- US Department of Energy: Department of Energy celebrates fourth criticality ahead of July 4th goal (6 Jul 2026)
- US Department of Energy: report evaluating increase in electricity demand from data centers (LBNL 2024)
- Vertiv: complete power and cooling blueprint for NVIDIA GB200 NVL72
- Methodology: facts checked against company announcements, trade press and government sources on 20 October 2026. “No public announcement found” means we searched company and trade sources and found none; it is not a claim that a milestone has not happened. Not investment advice.
