Here is the contradiction the market is pricing right now: the same firms that spent a decade insisting energy was a bad venture business—too capital-heavy, too slow, too regulated—just put $470 million behind a reactor that produces, at most, enough power for 750 homes.
That is not a rounding error in the AI power story. That is the whole thesis flipped inside out.
This week the money answered from the supply side, and the answer is stranger than the problem.
On June 4, Antares Nuclear's demonstration reactor, the Mark-0, reached criticality at the Idaho National Laboratory. By late July, the company had a $470 million Series C in the bank—$370 million in equity, $100 million in debt.
The round was led by Paradigm and Caffeinated Capital.
Point72 Ventures, Industrious Ventures, and Shine Capital rounded out the syndicate. Altogether Antares has now raised $604 million, with its $96 million Series B closing only in December.
Here is what should stop you: no regulator has approved a commercial deployment. As one observer put it, the check just says someone thinks the regulator will say yes.
This is a bet on permission, not on physics.
This is a bet on permission, not on physics. What the smart money underwrites isn't cost per megawatt-hour—it's the identity of the buyer on the other side of the first purchase agreement.
Because the thing the smart money is underwriting isn't the reactor's cost per megawatt-hour. It's the off-take—the identity of the buyer standing on the other side of that first power purchase agreement.
Consider the number everyone in this round has clearly read and chosen to ignore. Lazard, which prices the cost of energy across technologies, pegs new SMRs at roughly $214 per megawatt-hour.
At that price, an SMR costs more than all but the most expensive gas turbines. On the open grid, against a merchant power curve, that is not a business. It's a science project with a burn rate.
So the VCs did the only rational thing: they stopped pretending SMRs would win on cost and went looking for a customer who doesn't buy on cost.
They found two.
The first is in Antares' own press release. The company is one of three finalists in the Pentagon's Advanced Nuclear Power for Installations program, which will test SMRs on Air Force bases in Colorado and Montana. Deployments at U.S. military installations are planned for 2028.
The military is, in the language of the reporting, a famously price-insensitive customer. A base doesn't optimize for $/MWh. It optimizes for a reliable power source that can't be cut by a downed transmission line or a strained utility.
That is the entire pitch. Antares hasn't even disclosed its pricing—and given who it's selling to, it doesn't have to. Availability, resilience, and sovereignty are the product. The kilowatt-hour is incidental.
This is why the round reads coherently despite the LCOE problem. You aren't buying into the grid economics of small nuclear. You're buying a claim on a defense procurement pipeline that measures value in mission continuity, not merchant spreads.
The second price-insensitive buyer is the one that brought all this venture money into the room in the first place: the hyperscaler.
Look across the Atlantic at Manchester's Nuclear Turbines, which raised €17.5 million led by IQ Capital, with Rhapsody Venture Partners, Zero Carbon Capital, and Empirical Ventures. Smaller round, sharper thesis.
The company's whole design premise is to strip out the steam infrastructure that makes conventional reactors big. Co-founder Jeremy Owston: "The nuclear industry has always designed reactors first, then figured out what to do with the heat. We flipped this script: we started with the cheapest way to generate power and designed a reactor to fit."
The engineering payoff is a system compact enough to deploy behind the meter—on industrial sites, manufacturing facilities, and data centers.
Behind the meter is the phrase that matters. It means the power never touches the grid, never clears a wholesale market, and never has to beat a gas turbine on Lazard's curve. It sits on the customer's own land and sells directly to the load.
For a data center operator staring at a grid interconnection queue measured in years, "we'll put the generation on your campus" is worth paying a premium for. Again, availability, not cost.
So the pattern holds on both sides of the ocean and both sides of the customer base. Antares sells resilience to the Pentagon. Nuclear Turbines sells co-location to industry and compute. Neither is competing to be the cheapest electron on the grid, because neither intends to sell onto the grid at all.
The capital is pricing captive demand. That is the thesis in one line.
And you can see the same financial logic maturing one layer up the stack. Google has reportedly committed to roughly $44 billion worth of backstops on data center leases—acting as the guarantor, per The Information. Wall Street financing techniques, applied to the compute buildout.
That is the demand side learning to make its power appetite bankable. The nuclear rounds are the supply side answering with generation designed to sell against exactly that kind of long-dated, credit-backed off-take.
When a hyperscaler will guarantee a lease and a defense program will guarantee a purchase, the reactor stops being an energy asset and becomes a contract wrapped in concrete. Venture capital knows how to price contracts. It has never known how to price merchant power. So it chose the version of nuclear that looks like a contract.
Which brings us to the fault line all this capital is quietly papering over.
Criticality at a national lab is a real milestone. It is also roughly a decade and an entire missing supply chain away from commercial-scale field deployment.
The first SMRs aren't expected in service until the early 2030s. The mass-manufacturing economics every startup invokes—the promise that building the hundredth reactor costs a fraction of the first—typically take at least a decade to materialize. No startup has reached that stage. Not one.
The U.S. supply chain for advanced nuclear is immature, and scaling production is the exact problem that has broken hard-tech companies with far more forgiving physics.
Here is the part that should give the syndicate pause. The capital migrating into this sector—reportedly software-native, growth-native—has never had to survive a physical timeline like this.
Software capital compounds on iteration speed. You ship, you measure, you fix, you ship again in a week. A reactor gives you none of that. The feedback loop between a lab milestone and a fielded, revenue-generating unit is measured in fuel qualification cycles and regulatory dockets, not sprints.
Antares says it aims to bring its first electricity-producing reactor online next year, with base deployments in 2028. That is an aggressive schedule for a technology that just reached criticality in June, and it is being underwritten by investors whose entire muscle memory is built around timelines an order of magnitude shorter.
The off-take thesis is right. There genuinely are two buyers—the Pentagon and the hyperscaler—who will pay for availability and ignore the $214 sticker. The check is not irrational.
The risk isn't demand. Demand is the surest thing in the whole equation. The risk is the gap between a reactor that works once, under lab conditions, and a fleet that ships on schedule, at cost, past a regulator that hasn't yet said yes.
That is the wager Paradigm and the rest just made. Not that the physics works—it does. Not that the customer exists—it does. But that a class of investors trained on software velocity can hold its nerve through a decade of concrete, fuel, and NRC dockets before the off-take ever pays out.
The money has correctly identified the buyer. Whether it has correctly identified its own patience is the question the next ten years will answer.