Why this is happening now
Fission did not get easier. What changed is that money, demand and the regulator lined up at the same time, which has not happened in roughly four decades.
- Demand. AI campuses need firm power around the clock, and hyperscalers have spent the last two years signing nuclear deals: Microsoft's agreement to restart the Three Mile Island Unit 1 reactor (2024), Google's agreement with Kairos Power for up to 500 MW (2024), Amazon's investment in X-energy (2024) and Meta's package of up to 6.6 GW with Vistra, Oklo and TerraPower (2026).
- Regulation. The NRC's Part 53 rule, a technology-neutral, risk-informed licensing path for new reactors, took effect on 29 April 2026, more than a year ahead of its statutory deadline. Separately, the Department of Energy's Reactor Pilot Program let companies build and operate test reactors under DOE authorisation, with a target of criticality by 4 July 2026.
- Capital. Venture rounds that would have been unthinkable for nuclear five years ago are now routine: a $1 billion Series B for Valar, $300 million for Radiant, a billion-dollar IPO for X-energy.
The pilot programme is the part most coverage underrates. It turned "when will advanced nuclear work" from a paperwork question into a construction race, and the results were public within months.
The July 4 scoreboard
Four reactors reached criticality before the federal deadline:
- Antares, Mark-0, 4 June 2026 at Idaho National Laboratory. A sodium heat-pipe design and the first advanced reactor to go critical under the pilot programme.
- Valar Atomics, Ward 250, 18 June 2026 in Emery County, Utah, and the first to do it outside a national laboratory.
- Deployable Energy, under DOE's separate Nuclear Energy Launch Pad, in early July.
- Aalo Atomics, Aalo-X, at 12:20 a.m. on 4 July 2026, the deadline itself.
One caution before reading too much into that list. Most of these were zero-power demonstrations: the chain reaction is self-sustaining but produces essentially no usable energy. It proves the core physics and the team's ability to build and fuel hardware. It does not prove a power plant.
Valar Atomics vs Radiant
The comparison investors keep making is these two, because they are pursuing almost the same machine: a high-temperature gas reactor with TRISO fuel and helium coolant, small enough to build in a factory. They differ on where they think the first dollar comes from.
Valar: speed and capital
Valar broke ground on Ward 250 in September 2025, reached zero-power criticality on 18 June 2026 and generated electricity on 1 July, according to Nuclear Engineering International. Construction to criticality in about nine months is genuinely fast for nuclear hardware. Shortly after, it used the reactor to run an NVIDIA Blackwell system on a live stream, and the two companies announced work on a 30 MWe AI facility in Utah as part of Valar's "Gigasites" plan.
In August Valar closed a $1 billion Series B led by Sequoia, with Shaun Maguire joining the board, plus a $200 million credit facility. That followed a roughly $450 million raise earlier in the year at a reported $2 billion valuation. The commercial unit is designed for up to 5 MWe, so a 30 MW site means a fleet of them.
Radiant: customers first
Radiant's Kaleidos is a 1 MWe unit designed to be moved by truck. It was the first reactor selected to test in the DOME facility at Idaho National Laboratory and left its El Segundo factory for Idaho in August 2026. The planned campaign runs from zero-power criticality to full power and a 150-hour unattended run. As of mid-October we found no public announcement that it has gone critical, having missed the July target.
Where Radiant leads is demand. On 26 August 2026 the Army's Janus programme awarded it up to $750 million for 15 Kaleidos units, starting with three at Fort Benning, Georgia, by 2028. It also has an Air Force agreement to deliver a microreactor to Buckley Space Force Base. Total private funding is a little over half a billion dollars, including a $300 million Series D in December 2025.
How I would read it
Valar has about three times the capital and a working reactor. Radiant has a signed, milestone-based military buyer for a product that fits on a truck. The military is the better first customer for a microreactor: bases pay a premium for resilience, the Army regulates its own reactors, and a 1 MW unit is the right size for a base. Data centers are the bigger prize but want tens to hundreds of megawatts, which a 5 MW unit only reaches as a fleet, and fleets need a factory, a fuel supply and an NRC licence that neither company has yet.
So the honest answer is that they are ahead in different races. If Kaleidos completes its DOME campaign, Radiant has the clearer path to revenue in 2028. If Valar turns Ward 250 into a licensed, repeatable product, its capital lets it build the factory first, and that is what decides the data center market.
The wider field, compared
| Company | Design | Status (Oct 2026) | Capital | Anchor customer |
|---|---|---|---|---|
| Valar Atomics | Ward 250 HTGR, TRISO fuel, helium cooled; commercial unit up to 5 MWe | Criticality 18 Jun 2026; first electricity 1 Jul 2026 (Utah) | $1B Series B (Aug 2026, Sequoia) plus a ~$450M round earlier in 2026 | NVIDIA collaboration on a 30 MWe AI facility in Utah |
| Radiant | Kaleidos 1 MWe HTGR, TRISO fuel, helium cooled, truck-transportable | Shipped to INL DOME test bed Aug 2026; no public criticality announcement yet | ~$500M+ total; $300M Series D (Dec 2025) | US Army Janus award up to $750M for 15 units; Air Force delivery by 2028 |
| Aalo Atomics | Aalo-X 10 MWe sodium-cooled, low-enriched uranium; 50 MWe "Aalo Pod" | Zero-power criticality 4 Jul 2026 (INL) | Not covered here | Project Ascension: 10 MWe reactor powering an on-site data center, targeted 2027 |
| Antares | R1 heat-pipe microreactor, 100 kWe to 1 MWe, TRISO/HALEU | Mark-0 zero-power criticality 4 Jun 2026 (INL), first in the pilot programme | >$140M; $96M Series B (Dec 2025) | Air Force, Space Force, NASA, DIU; Army Janus site at Fort Bragg |
| Oklo | Aurora sodium-cooled fast reactor, uprated to 75 MWe | No NRC licence yet for its Ohio campus; public company | Listed (NYSE: OKLO) | Meta prepayment for a 1.2 GW campus in Pike County, Ohio |
| Kairos Power | Fluoride-salt-cooled, TRISO; Hermes test reactor and 50 MWe Hermes 2 | NRC construction permits issued; Hermes targeted 2028 | Private | Google agreement for up to 500 MW (2024) |
| X-energy | Xe-100 HTGR, 80 MWe per module, TRISO | Seadrift (Dow) construction permit under NRC review; first fleets early 2030s | Nasdaq IPO April 2026 (XE) | Amazon investment and offtake framework |
The table mixes two kinds of company on purpose. Microreactor developers (Valar, Radiant, Antares, Aalo) build units from under 1 MW to about 10 MW and can test fast under DOE authorisation. SMR developers (Oklo, Kairos, X-energy) build 50 to 80 MW-class modules for hyperscaler campuses and are already deep in NRC licensing. The data center deals with the largest headline numbers sit with the second group; the fastest hardware progress this year came from the first.
Aalo is the one to watch for a data center specifically. Its commercial "Aalo Pod" is 50 MWe built from 10 MWe reactors, and its Project Ascension reactor at INL is meant to power an on-site data center in 2027. It also uses low-enriched uranium rather than HALEU, which sidesteps the fuel bottleneck below.
What still stands in the way
- Fuel. Most of these designs need high-assay low-enriched uranium (HALEU), enriched to between 5% and 20%. Commercial US supply is small, and every company's fleet plan depends on it growing.
- A commercial licence. DOE authorisation covers test reactors. Selling power to a data center needs an NRC licence, and no microreactor in this list has one yet. Part 53 makes that path clearer, not shorter by default.
- Manufacturing. The economic argument for small reactors is that a factory makes them cheap. That argument is untested until someone runs the factory, which is why the capital raised matters more than it would for a software company.
- Scale mismatch. A hyperscale AI campus is hundreds of megawatts. Microreactors fit inference sites, edge capacity and off-grid builds first; SMRs are the realistic path to whole campuses.
What this means if you plan AI infrastructure
Nothing in this list will power a production AI cluster you are planning for 2027. The earliest credible dates are 2027 for a demonstration data center and 2028 for military units, with grid-scale SMRs later. What is worth doing now is the same as with any early-stage supplier: track which companies convert test milestones into licences and fuel contracts, and treat power availability as a site-selection constraint you plan around, not a problem a vendor will solve on your timeline. If your board is asking about AI's energy bill, the AI ROI work is the better place to start that conversation.
Related
- Physical AI — the cluster hub, and what actually gates deployment
- World models — the research bet on the robotics side of physical AI
- AI ROI — building the business case, energy costs included
Frequently asked questions
What is a nuclear microreactor?
A microreactor is a fission reactor small enough to be factory-built and shipped, typically producing from a few hundred kilowatts up to roughly 20 MW of electricity. Small modular reactors (SMRs) are the next size up, usually tens to a few hundred megawatts per module. Most of the microreactors in this roundup use TRISO fuel particles, which retain fission products at very high temperatures and are the main reason their designers claim passive safety.
Can a microreactor power an AI data center?
Not alone at hyperscale. A single 1 to 5 MW unit covers a small inference cluster or an edge site; a modern AI campus draws hundreds of megawatts. That is why the data center pitches are about fleets: Valar describes 30 MW sites built from 5 MW units, and Aalo plans a 50 MW "Aalo Pod" built from 10 MW reactors. Larger SMRs from Oklo, Kairos and X-energy are aimed at the hundreds-of-megawatts end.
Valar Atomics vs Radiant: which is ahead?
Valar is ahead on demonstrated physics and capital: Ward 250 reached criticality in June 2026 and produced electricity in July, and the company raised a $1 billion Series B led by Sequoia in August. Radiant is ahead on committed customers: a binding US Army award worth up to $750 million for 15 Kaleidos units, plus an Air Force delivery agreement. As of mid-October 2026 Radiant had not publicly announced criticality for Kaleidos at Idaho National Laboratory, so its next milestone is the one to watch.
What is NRC Part 53?
Part 53 is the Nuclear Regulatory Commission's new, optional, technology-neutral licensing framework for commercial reactors. It was published in the Federal Register on 30 March 2026 and took effect on 29 April 2026. It replaces prescriptive light-water-reactor rules with risk-informed, performance-based requirements, which is the route most advanced reactor developers expect to use for commercial plants.
Why are hyperscalers buying nuclear power?
AI data centers need firm, around-the-clock power, and hyperscalers have carbon targets that rule out building on gas alone. Since 2024 Microsoft, Google, Amazon and Meta have each signed nuclear agreements, ranging from restarting existing plants to funding advanced reactor developers years before their first unit runs.
When will a microreactor actually power a data center?
The earliest dates on public record are 2027 to 2028. Aalo says its Project Ascension reactor at Idaho National Laboratory will power an on-site data center in 2027. Military microreactor deployments under the Army Janus programme target 2028. Commercial grid-scale SMRs from Oklo, Kairos and X-energy target 2028 to the early 2030s. Nuclear schedules have historically slipped, so treat each of these as a target, not a forecast.
Sources: World Nuclear News, American Nuclear Society Nuclear Newswire, Nuclear Engineering International, POWER magazine, Federal Register (91 FR 15696), Radiant Industries, Bloomberg. Company figures are as reported by the companies or the press and have not been independently audited.