Two Years and Four Months to a Live Reactor: Valar Atomics Wants Energy Ten Times Cheaper

A company not yet three years old plugged an NVIDIA Blackwell chip into a live nuclear reactor and hosted a website on it. Every watt that chip burned came from fission, and the page told visitors how many uranium atoms had been split to deliver it. The machine behind that stunt is Ward 250, which Valar Atomics calls the first advanced reactor ever to make power as a startup, the first advanced reactor built outside a national laboratory, the fifth new nuclear device to make power in the United States since 2000, and the first TRISO-fueled reactor turned on in the country in more than fifty years. Every earlier company that had actually produced nuclear electricity, Taylor notes, was a century-old utility, a defense contractor, or a large engineering firm.

Sarah Guo recorded the episode on site at Valar’s Utah facility, San Rafael, walking the plant with Taylor rather than interviewing him across a table. Guo, an investor who co-hosts No Priors with Elad Gil, usually talks to AI founders and lab researchers. Here the claims sit next to hardware: a control room that flew in on C-17s, a 78-inch bioshield of stacked concrete, and a core making about 100 kilowatts while they speak, splitting on the order of 10 to the 17 atoms per second.

Taylor does not have a nuclear-engineering doctorate or decades in the industry. His great-grandfather was a nuclear physicist on the Manhattan Project; his great-grandparents both lived in the secret city during World War II. He grew up assuming fission was a solved technology and that he would someday build thousands of machines, without knowing which ones. One afternoon in high school he realized the United States had actually stopped building reactors in the 1970s. "Why did nobody tell me that?" He spent ten years watching every nuclear startup that appeared, waiting for someone to finish the job. Nobody did it at the pace and scale he thought necessary, so he started Valar out of sheer frustration.

The United States stopped building reactors, Taylor says, for a simple reason: Three Mile Island. In 1979 a light-water plant lost cooling and the core melted. Nobody died, nobody was injured, and the public received no radiation dose. What failed was optics. The incident was mismanaged as public relations, fear took hold, and nuclear went out of fashion. He thinks public opinion actually flipped back fifteen or twenty years ago, once people admitted that plants can be built safely and that the accident’s lessons had been absorbed. The deeper damage was industrial. Once you stop an industry, restarting it is hard. In the decades after Three Mile Island, the United States atrophied at large-scale civil works—bridges, dams, giant power plants—and got much better at advanced manufacturing.

That history is why he thinks a reboot cannot look like the 1960s. It has to look like manufacturing: small modular reactors, built as equipment rather than poured as infrastructure. Plenty of people noticed the SMR slogan. What they missed, in his view, is that you only get there by turning machines on. You light a 100-kilowatt plant like the one behind him. You go cold-critical, as Valar did in Project Nova. Then you take the next step. Hardware iteration, not a prettier drawing, is the path. The regulatory chicken-and-egg of the last twenty years blocked that: you need operating data to go to the regulator, and you need the regulator to run a plant. The industry’s hack was modeling. Famous nuclear companies, he says, are often modeling-and-simulation shops producing "paper reactors" with precise predictions of a theoretical machine.

Congress originally wrote two pathways. The commercial one is the Nuclear Regulatory Commission, built for large-scale deployment of mature systems—an awkward cop for a team that is still learning how to bend steel. The testing pathway belongs to the Department of Energy. Most people forget that DOE was briefly called ERDA, the Energy Research and Development Agency, for about two years after Congress split the Atomic Energy Commission in two: NRC for commerce, ERDA for tests. Then Congress renamed ERDA the Department of Energy. The agency’s origin is testing reactors, and that legal fact sat unused for roughly forty years except for the odd NASA space-reactor program. The Trump administration picked it up. Executive Order 14301 called for three advanced reactors to go critical on American soil by July 4th. Ward 250 was turned on under DOE authority under that order. The chicken-and-egg, Taylor says, is finally cut: they are sitting next to a running core with real data.

Guo asks what outsiders get wrong. Empirically, Taylor says, nuclear is the safest energy source ever built, safer even than solar. Most solar is installed on roofs, and the deaths from falls, added up, exceed nuclear’s death-per-energy worldwide. Three Mile Island melted and killed no one. Fukushima, in a debate among scientists, may have killed one person from radiation. Traditional plants still have a structural weakness. After a scram there is no chain reaction, but recently split atoms keep decaying and produce about 5 to 6 percent of the prior power as heat. Light-water designs must keep cooling pumps running for roughly a day or the core melts. That is what failed at Three Mile Island and Fukushima. The old philosophy spends its effort making those pumps never fail.

Valar wants to delete active cooling entirely, because it intends to build thousands, then tens or hundreds of thousands, of reactors. Around the core sit RCCS water jackets. After shutdown, water boils, steam rises, condenses, and returns; heat leaves by geometry, with no pumps, no moving parts, and no operator input, over about two days. They already ran the experiment in Hawthorne, California, before fuel was loaded: electrical resistors dumped about fifteen Los Angeles city blocks of power into the plant, brought it to nuclear temperatures and pressures, then shut every safety system at once and watched for two days. The jackets went into natural circulation as designed. The Utah control room is the Hawthorne room, flown in on C-17s with the rest of the plant. A few days after taping, they planned to do it with neutrons: hit the red scram button, drop boron-carbide rods, kill the chain reaction, then cut site power after 72 hours at full power.

Risk, he reminds Guo, is odds times consequence. Conventional nuclear hammers the odds. Advanced designs hammer the consequence, because odds are stochastic and something will come out of left field. Valar’s safety basis for the regulator is that everything in the plant has failed, and still neither the public nor workers receive a radiation dose. Site security and trained operators still exist, but the argument does not depend on them. TRISO fuel, graphite moderation, helium coolant, and this geometry are, in his words, about as safe as humanity knows how to build a reactor. Helium’s chemistry is almost empty: it is inert, unlike wet steam chewing through turbines. The reliability lessons they did take from history are specific. Fort St. Vrain, a 1970s gas-cooled plant, suffered because graphite is hydrophilic even in desert air; moisture rusts steel and the particles clog equipment. Valar runs a helium purification system for hundreds of hours to leach water out of the blocks. Keep the working fluid simple, and the plant can stay simple.

Inside Valar the score that matters is tick rate, a term borrowed from game servers: the time between changes in game state. Translated, it is the gap between real progress marks. From the Delaware filing to the first atom split was two years and four months. From that Project Nova cold criticality, in November, to Ward 250 making power was about seven months. Taylor wants the interval driven down to minutes, a new reactor lighting every few minutes. Nuclear economics, he says, are not about uranium, which is cheap. They are about how fast and cheaply you can produce plants. Iteration and scale are the same variable, and that variable is tick rate.

He draws one line through the industry. Some companies treat nuclear as a hardware-execution problem: can you build, light, operate, and go from one plant to ten to a hundred to a thousand? Others treat it as a design problem and fill documents with rare materials and supply chains that do not exist. Taylor’s analogy is the Toyota Camry. Nuclear today is a Camry problem; the Lamborghini can wait. A simple, cheap, safe reactor you can make by the tens of thousands will beat a more efficient one-off on cost. Valar’s hiring filter matches that view. Taylor is in every hiring decision. He wants extreme agency and a record of having built hard things in the physical world. Many employees had never seen a reactor when they walked in. From inside nuclear he recruits people who loved the field, left because it was too slow, built hardware somewhere else, and can be talked into coming back—so long as they want to turn plants on rather than write papers.

The bioshield is the manufacturing exhibit. Seventy-eight inches of concrete, branded the modular citadel, is why they can stand next to the core. A conventional pour would take at least three months. Valar precasts the blocks in its own Salt Lake City factory, rated at 3,000 identical blocks a year, trucks them in, and stacks them with a crane in about 42 hours. The hard physics is the seam. At microscopic scale two blocks do not touch, so gamma rays and neutrons can slip through. Every joint follows a sine-wave tortuous path, including the vertical edges and corners; there is no straight line from inside to outside. There is also no grout, no bolts, and no screws. A seismic frame went on because Utah building code argued with the engineers, not because the stack needed mechanical locks.

They had to invent a concrete grade to make that work: dense enough to stop gammas, strong enough to self-stack, no rebar (neutrons would activate it into waste), and an atomic mix that does not become radioactive under irradiation. Two engineers, 23 and 21, spent three weeks on a "rock hunt," flying the country for samples, dissolving rocks in acid buckets in the conference room on Friday nights, running spectroscopy, and flying out again. People in the industry have been writing since the 1980s that someone should prefabricate a shield this way. When Valar unveiled it, Taylor got messages from career nuclear people saying they had never seen it done.

Beside the control room sit three unremarkable boxes that turn analog detector signals into digital ones: amplifiers and rectifiers, first-year electronics. The vendor charged $450,000 each. Valar paid $1.5 million for the set because a few months of delay was more expensive than the invoice. The next box did not get that courtesy. It is the reactor protection system, the plant’s brain: no human input, three independent channels that vote, two votes that the plant is unsafe and it shuts down. The quote was $5 million and two and a half years. After two months of haggling on schedule, Taylor pulled the team into a conference room and said they would build it. Joe, who runs instrumentation and control and dropped out of Brown, locked five people in a room. Six weeks later they had a working RPS for about $400,000.

The vendor’s response was ugly. Once it was clear Valar would not wait two years, the company told the industry that Valar was unsafe and would kill people, and volunteered for expert-network calls to say so. Taylor’s reading is calm. An industry that has barely built anything for forty years lives on 100-times markups; threaten that margin and it panics. The same pattern, he says, is everywhere you look. The secret weapon is a willingness to verticalize anything on the path to scale. He would rather buy a reactor kit and staple it together. In practice, parts are wildly overpriced and supply is constrained, so they hire the best people they can find and run toward the gunfire: sites, regulation, fuel, instrumentation, shielding. If in ten years Valar is worth $500 billion, he wants the through-line to be that they kept the hard pieces in-house.

Money follows the same contrarian cut. The standard nuclear-startup plan is a packet of designs and customer letters of intent, then a hunt for debt or project finance. Taylor watched that fail for a decade and refused it. The physics works and demand, in his view, is infinite; what remains is technology-execution risk, and American venture capital is the world’s best underwriter of that risk. Mechanical engineering, thermal hydraulics, instrumentation, and manufacturing methods are complicated, he says, but less complicated than a rocket engine. While competitors try to persuade risk-averse lenders, Valar intends to be on its fifth reactor. Only then does project finance become an option, and the equity already in the ground becomes a moat.

Customers are treated the same way. Peers crowd around hyperscalers and die in multi-party negotiations over sites and permits. Valar’s plan is to put a gigawatt on the ground, with land and fiber, on its own clock, and let load follow. "If I have a gigawatt of power with land and fiber, is someone going to put a data center there?" Yes, he says. Dedicated customer gigasites will come later, but only for buyers who want power in a year, not ones still considering it in three or four. Ask him any question long enough and you get back to speed and scale: nuclear’s missing Ford, Toyota, or SpaceX moment.

AI compute is a tailwind, and Taylor is grateful to NVIDIA, but he does not want the company’s story to be a data-center shortage. Energy is a commodity; price sets demand. Make a kilowatt-hour a cent and you induce uses that did not exist. Make it a tenth of a cent and you induce another round. Energy, he says, is the only truly scarce, irreversible resource in the universe, which is why he treats Valar’s market as infinite. The nuclear website, nuclearwebsite.com, stays up only while the reactor runs; when the plant shuts down, the page goes with it. Merch is sold only there, because he refused to sell any until the company had actually split an atom. Big compute buyers tell Guo that nuclear cannot deliver at scale until 2031 or 2032, with diesel and solar-plus-storage filling the gap. Taylor thinks they have been looking at modeling shops. Exponential curves fooled people about SpaceX’s satellite counts too. On the current industry mindset, he would not even bet 2035.

Pace, he says, is the least copyable asset a company has, and it has to be set on day one by the CEO. The first five people and the first twenty have to have it in their bones, and they have to hire for it. His own job is to walk the organization, spot the slow joint, stand up a physical war room, and turn a six-month timeline into four weeks—or eight, but not six months. There is no autopilot. He will always have to push, because he believes fission can make energy a thousand times cheaper than it is today. Somebody will do that. Whether it happens in the coming decades or the coming centuries is a function of their speed.

Cheaper energy, in his telling, is how living standards actually jump. Agrarian muscles ran on photosynthesis. Underground hydrocarbons brought winter heat, summer air conditioning, and a drive to the hospital. Aluminum was once a precious metal for royal jewelry; Charles Hall’s electrolysis, plus cheap enough electricity, turned it into the struts of the building they are standing in. Today’s comfort sits entirely downstream of "pretty cheap" hydrocarbon energy. The goal now is insanely cheap. Jet fuel has little juice left to squeeze, which is why a weekly flight to see a grandmother stays expensive. Energy ten times cheaper makes a thirty-minute Saturday hop across the country ordinary.

He has a name for the further step: hyper-techno-industrialism, everything trending toward free. A microphone comes from a factory with three inputs—people, materials and machines, and energy. AI and robots convert the human input into watts: one person coordinates a hundred robots. Ask where the machines came from and you get another factory, then another, until energy is the root input and the price of a good converges on the energy used to make it. Ten times cheaper, then ten times cheaper again, and goods approach free, which is what you need if you are going to put a lot of stuff on Mars. Guo calls that a good place to end. Taylor’s last claim is the same one he started with: the physics already work. What remains is who builds, and how fast.