Jul 4, 2026 · 4:22 AM
Subscribe
Home Ai

Three nuclear startups beat Trump's deadline and one already powers an Nvidia chip

Three US nuclear startups, Antares Nuclear, Valar Atomics, and Deployable Energy, pushed reactors to criticality ahead of President Trump's July 4, 2026 deadline. Valar went further on July 1, feeding live power from its Ward 250 microreactor into an Nvidia Blackwell chip, but none of the three has a signed offtake deal with a hyperscaler, and their reactors are still far from commercial-scale power.

Julian Lim
· 5 min read · 144 views
Three nuclear startups beat Trump's deadline and one already powers an Nvidia chip

Three nuclear startups pushed reactors to criticality ahead of President Trump's July 4 deadline, and one is already feeding real electricity into an Nvidia chip, though a working power plant is still years away.

On July 1, in a small data center in Emery County, Utah, a nuclear reactor that did not exist three years ago began sending live electricity into an Nvidia Blackwell chip. Valar Atomics built the reactor. Nvidia supplied the silicon. The companies say it is the first time a next-generation US reactor has delivered power directly to AI compute hardware.

That demonstration capped a run of about four weeks in which three separate startups pushed reactors to criticality, the point where a nuclear chain reaction becomes self-sustaining without outside help. Antares Nuclear's Mark-0 got there first, on June 4, at Idaho National Laboratory's Reactor and Critical Experiment facility. Valar's Ward 250 followed on June 18, at a site in Emery County. Deployable Energy's Unity reactor closed out the run late on June 30, also at Idaho National Laboratory. All three beat the July 4, 2026 deadline President Trump set in a May 2025 executive order, which directed the Department of Energy to reach at least three advanced reactor criticalities within about fourteen months.

The Mark-0 result carries weight beyond the calendar. According to the Department of Energy, it was the first time in more than four decades that a privately developed, non-light-water reactor reached criticality on US soil. Mark-0 is cooled with sodium heat pipes and fueled with TRISO compacts, the same high-assay low-enriched uranium fuel form several rival designs are betting on. Antares's CEO summed up the roadmap without much hedging: criticality in 2026, electricity production in 2027, power to the warfighter by 2028.

None of that means the lights are on yet.

Every one of these three results was what the DOE calls a zero-power fueled criticality demonstration, not a working power plant. The reactor reaches the minimum level needed to sustain a chain reaction, but it never produces meaningful heat and never has to prove it can run a turbine, cool itself under load, or hold up for a full fuel cycle. Engineers use the test to check whether their physics models, core geometry, and control rods behave the way the paperwork predicted. It is a genuine milestone. It is also, plainly, a lab result, not a power source.

Ward 250 pushes that distinction furthest. It is a helium-cooled, TRISO-fueled high-temperature gas reactor that, in commercial form, is designed to produce five megawatts of electricity, enough for a modest server room, not a hyperscale campus. It is also the first reactor authorized under the DOE program to be built and operated outside a national laboratory, on a site Valar broke ground on September 17, 2025, seven months after testing a non-nuclear prototype called Ward Zero. Getting from that groundbreaking to criticality in barely nine months is an unusually fast timeline for a nuclear project of any kind.

Money is flowing at different speeds behind each company. Antares closed a $96 million Series B in December 2025, led by Shine Capital with Alt Capital, Caffeinated, and other investors, split between $71 million in equity and $25 million in debt earmarked for its Torrance, California factory and uranium procurement. Valar has moved faster still: a $450 million round that valued the company at $2 billion, arriving barely five months after a $130 million Series A. Palmer Luckey, the Anduril founder, and Palantir chief technology officer Shyam Sankar are both investors, alongside Lockheed Martin board member John Donovan. Deployable Energy says it is self-funded, has about 20 employees as of May 2026, and is building out of Houston a shippable, container-sized one-megawatt reactor aimed at remote and defense sites, not data centers.

The program structure differs too. Antares and Valar earned their criticalities under the DOE's Reactor Pilot Program, launched in June 2025. Deployable Energy's Unity result came through a related but separate effort, the Nuclear Energy Launch Pad, which the DOE stood up in March 2026 through its National Reactor Innovation Center to widen the pipeline beyond the original pilot cohort.

The offtake question nobody has answered

Here's the part AI data center operators actually care about, and it is thinner than the criticality headlines suggest. None of the three companies has announced a signed offtake agreement with a hyperscaler. Microsoft, Amazon, and Google have all signed nuclear power agreements over the past year, including Google's deal with Kairos Power, but the reactors behind most of those contracts do not yet exist at commercial scale. Valar's Nvidia demonstration is closer to a proof of concept than a contract. It shows electrons can move from a working reactor to a live GPU. It does not show that Microsoft or Google is buying gigawatts from Valar next year, or any year soon.

Ward 250's commercial design tops out at five megawatts. A single Nvidia GB200 rack alone can draw well over 100 kilowatts, and a serious AI data center campus needs hundreds of megawatts, sometimes gigawatts. Getting from a zero-power criticality test to a reactor that reliably feeds that kind of grid connection, day after day, for years, is the harder and far more expensive part of the roadmap. That part has not happened yet, for Antares, Valar, or Deployable Energy alike.

Also read: HCLTech beats Infosys to land a $1.14 billion AI deal with Mercedes-BenzAlibaba Bans Claude Code After Hidden Anthropic Tracking Code SurfacesIsrael Turns Its Pact With Argentina Into an AI Foothold in Latin America

TOPICS
Julian Lim is an entrepreneur, technology writer, and a researcher. He started JL Data Analysis after graduating from NUS in Intelligent Systems. Julian writes about technology innovations and entrepreneurship on Business Times, Asia Pacific Magazine and occasionally contributes to Startup Fortune.
Related Articles
More posts →
Loading next article…
You're all caught up