First in a Generation, Again

NuScale's Carbon Free Power Project — the small modular reactor that was going to prove the whole model — was cancelled in November 2023. Not by protesters. Not by a regulator saying no. It died because the estimate went from $5.3 billion to $9.2 billion, the projected price of the power went from $58 to about $89 per megawatt-hour, and the municipal utilities meant to buy it declined to sign. No offtake, no reactor. Same site, same federal landlord, less than three years ago.

On July 4, 2026, a reactor went critical at Idaho National Laboratory anyway.

Aalo Atomics' test unit, with Crusoe lined up to run a modular data center off it next year and 50-megawatt units penciled in for Crusoe sites by the end of the decade. Behind the meter — the electrons never touch the public grid. The Energy Department is soliciting proposals for more AI data centers and generation on the same federal land, one of four sites designated for it.

i · what actually changed

The tempting sentence here is nothing about the engineering changed; a customer showed up. It's rhetorically clean and it doesn't survive contact with the two projects.

CFPP was a commercial power plant — six 77-megawatt modules, 462 megawatts total, selling to municipal ratepayers, on the Nuclear Regulatory Commission's licensing path. NuScale had NRC design certification for its 50-megawatt module; the uprated 77-megawatt version the project actually needed was still working through certification when the plug came out. Aalo-X is a small experimental test reactor of a different class, built on DOE land, authorized by the Department of Energy under its Test Reactor Pilot Program — a pathway created by executive order, not an NRC license at all. The safety case ran through a DOE-approved Documented Safety Analysis, and Aalo-X was the fourth DOE-authorized reactor to reach criticality against a July 4 deadline set by the White House.

So the honest version is narrower and considerably more interesting. The physics didn't change. What changed was who was buying — and how little had to be licensed in order to serve them.

That second clause is the finding underneath the finding. The reactor that got built is the one that didn't have to ask the civilian regulator. A test unit on federal land serving a single private customer behind the meter is, from a permitting standpoint, a fundamentally easier object than a commercial plant selling to the public. The demand showed up, and the demand routed itself around the slowest part of the system. Everyone will report the customer. Fewer will report the detour.

ii · the changelog

Because this is the oldest story in American infrastructure, and it has a changelog.

The first commercial-scale reactor in this country, Shippingport, went critical in December 1957. It exists because Hyman Rickover's naval program needed reactors, a nuclear aircraft carrier got cancelled, and the hardware and the program office went looking for a mission. "Atoms for Peace" was the speech, delivered to the UN in December 1953. The Navy was the balance sheet. Civilian nuclear power in America began as defense procurement with a ribbon tied around it.

The interstates are the National System of Interstate and Defense Highways. GPS was a targeting system. ARPANET was survivable command and control. The pattern isn't that these things were secretly sinister — it's that in each case the public benefit was the byproduct rather than the load-bearing argument.

The honest boundary on that claim: it is a strong tendency, not a law. Rural electrification, the TVA, and municipal water systems were built without a concentrated well-capitalized private customer, and there the public benefit genuinely was the argument — which is worth remembering, because it means the pattern describes a default, not a physical constraint. Where that default holds, large American infrastructure gets built when someone with a deadline and a balance sheet needs it. Everything else is the press release. But the exceptions exist, and they were all cases where the public organized itself into the role the anchor customer usually plays.

iii · the cause was a spreadsheet

Which means the standard story about how nuclear died here is also wrong, in the useful way.

Everyone says Three Mile Island, March 1979. But the orders had already collapsed. After 1973, load-growth forecasts flattened, interest rates climbed, and construction costs ran away from every projection; utilities cancelled dozens of plants before the accident happened. TMI became the explanation because it had a photograph. The cause was a spreadsheet. American nuclear power was not killed by fear. It was killed by the absence of a customer who needed it badly enough to eat the cost curve — which is precisely the variable that just changed.

So the reactors get built. A load that runs flat at high capacity factor, around the clock, indifferent to weather, is the one demand profile nuclear is genuinely good at serving, and for the first time since Rickover there is a buyer with a deadline and a balance sheet. Expect the phrase "first new reactor in a generation." Vogtle 3 and 4 were also the first in a generation, at roughly thirty-five billion dollars and years late, and everyone has agreed to forget that.

iv · who holds the bag

The question worth tracking is who absorbs the loss if the forecast is wrong.

Behind-the-meter is the honest version: a private company contracts for a private reactor, posts the credit, absorbs the downside. Fine. But behind-the-meter is a demonstration format, not a national energy strategy, and the second wave is where this actually gets decided — the units that interconnect, built against projected demand rather than signed contracts, arriving in a rate case dressed as "economic development" or "grid modernization." At that point the structure is: private party takes the electrons, public takes the stranded cost.

We have run that one. In 1983 the Washington Public Power Supply System defaulted on $2.25 billion in bonds — the largest municipal bond default in American history to that point — after ordering five reactors against demand growth that never arrived. One was finished. Bondholders and ratepayers absorbed the rest. The utilities weren't lying about the forecast. They believed it. That's the point: nobody has to be corrupt for this failure mode to execute. They only have to be confident.

v · leaving the commons

But the stranded-cost question is the small version, and I filed it as the main one. The larger structural fact is sitting in a dash three paragraphs up: behind the meter — the electrons never touch the public grid.

That is defection from a commons, and it deserves more than an aside.

A grid is a cost-sharing arrangement. Its fixed costs — transmission, distribution, reserve margin, the whole standing apparatus of reliability — get spread across everyone connected to it, and the arrangement works because the largest, most creditworthy, most reliably-paying participants stay inside it. Industrial anchor load has always been what makes residential service affordable; it pays into the fixed layer at scale and it pays on time.

Those participants have just discovered they don't have to. A hyperscaler with its own reactor behind its own meter is not a ratepayer. It is a jurisdiction. The fixed costs of a system built for everyone don't vanish when the anchor tenant leaves — they stay behind, redistributed across the people who cannot build their own reactor, which is everyone. Stranded cost in a future rate case is one symptom. The general case is a public system quietly losing the participants who were subsidizing its universality, and there is no hearing scheduled about that, because no rule is being broken.

Nobody in this story speaks for the residential ratepayer. They appear in the coverage, and appeared in my own first draft, only as a hypothetical bag-holder in a hypothetical second wave. They are not hypothetical and the wave is not second. They are inside the first one already, holding the fixed costs while the anchor load walks.

There's a quieter transfer running alongside it. The public paid for the national laboratory, the site, the authorization pathway, and the enriched-fuel supply chain no private market would have built. The compute is private. The electrons never reach the public. This is the Shippingport arrangement, unmodified, seventy years on — and being the Shippingport arrangement doesn't make it acceptable. It makes it durable, which is the more troubling reading and the one this piece's own thesis supports. The only real improvement over 1957 is that the wiring is visible.

vi · the cold lesson

Nobody won the nuclear argument. Two decades of hearings, ballot measures, and documentaries — and what moved was a load curve. Alignment beats force, but that cuts both ways: the future doesn't get decided, it gets arrived at, by whichever mass is heavy enough to bend the local geometry. Right now that mass is a data center.

The tempting conclusion is: stop arguing, go find a bigger mass. That conclusion is wrong, and it's wrong in a way worth naming, because it just restates the problem as a strategy. A world where outcomes belong to whoever shows up heaviest is the condition this piece has spent its length diagnosing. Winning that game doesn't fix it; it changes who gets to hold the bag next.

The more useful move is to notice that deliberation was never the load-bearing member here, and then ask what the mass is drawn toward — because that part is built, not given. Who bears the fixed costs, which pathway is fast and which is slow, what a rate case is allowed to consider, whether behind-the-meter generation owes anything to the grid it opted out of: those are all design choices, currently defaulted, and each one is a hand on the gradient. Argument was never the lever. The gradient is.

The reactor was always possible. It was just waiting for someone rich enough to be in a hurry — and for a pathway short enough to let them.