The Rung You Can't Streamline

Two curves, two directions

Hyperscalers now have about 9.8 GW of nuclear capacity committed to AI data centers across 13 announced deals, and every major U.S. hyperscaler has signed at least one. Meta's package alone runs up to 6.6 GW across TerraPower, Oklo and Vistra. Amazon has a 1.92 GW agreement for power from Susquehanna. Google has a Kairos fleet deal. Switch has a framework agreement with Oklo for up to 12 GW.

The rules are moving to match. The NRC issued Part 53 in March, its first new reactor licensing framework since 1989 and the first major rewrite of licensing standards since 1956. In July it proposed a broader package that it described as stripping out rigid frameworks and unnecessary conservatism. Separately, DOE's Reactor Pilot Program authorized test reactors outside the NRC process, and at least five of them have now reached criticality.

That's one curve. It points up and to the right.

Here's the other one. GAO reported last week that from July 2024 through June 2026, the NRC lost about 500 staff on net, mostly through voluntary retirements. At a May Senate hearing, Commissioner Bradley Crowell put it at 510 departures in 16 months against 59 new hires, with attrition expected to reach 14 to 15 percent this fiscal year compared with a historical average below 10. The FY2027 budget request asks for 2,606 full-time equivalents, about 7 percent fewer than last year. Chairman Ho Nieh's answer was that the agency is working smarter, and he attributed the departures to private-sector competition and work-life factors. Both things can be true. Smarter process helps. It doesn't replace the judgment that walked out the door.

Both curves are real. They just don't point the same way.

Give the reform its due

It would be easy to read that as a scandal. I don't think it is, and the case for reform deserves a fair hearing before anyone pokes at it.

A lot of the existing rulebook was written around large light-water reactors. Applying it line by line to a 50 MW molten-salt or gas-cooled design produces paperwork that doesn't buy any safety. Part 53's core move is to replace fixed design requirements with a demonstration that the plant meets safety performance targets. That's a reasonable idea. It's how mature engineering disciplines usually regulate.

Many advanced designs do have real physics advantages. Strong negative temperature feedback, natural-circulation cooling and low-pressure coolants aren't marketing. They change what can go wrong.

There's also a working model for safe nuclear at scale. Naval Reactors has run a large fleet for decades with an exceptional record. I started my career in Navy nuclear operations, and the lesson I'd carry from it is simple: the record came from culture, competence and conservative decisions, not from the thickness of the rulebook.

So the question isn't whether to streamline. It's what the streamlining relocates.

You don't remove a bottleneck, you relocate it

Licensing was a bottleneck. Everyone agrees on that, including people who worry about safety. The reforms are designed to remove it.

They won't remove it. They'll move it.

Under a prescriptive rule, safety is mostly embedded in the requirements. A reviewer checks a design against a list. Under a performance-based rule, safety moves into the quality of the applicant's safety case and the judgment of the person reviewing it. The burden shifts from the rulebook to the reviewer. A flexible rule is only as good as the bench reading it.

That bench is the part that's shrinking. Fewer, less experienced reviewers will face more applications, more novel designs, and a framework that leans harder on their judgment than the old one ever did.

The same move is happening one layer up. Executive Order 14300 directed that the Advisory Committee on Reactor Safeguards be cut back to its statutory minimum and focus only on issues that are truly novel. ACRS is the independent second look, the group that reports to the Commission rather than to the staff. For first-of-a-kind non-light-water reactors, almost everything is novel. That's where the second look earns its keep.

None of this shows up in the deal announcements. It shows up years later on an operating floor, if it shows up at all.

Chernobyl was a disclosure failure

People reach for Chernobyl as the case against nuclear. I think it's the case for review done right.

The RBMK reactor had a positive void coefficient at low power and control rods with graphite tips that briefly added reactivity when first inserted. That second flaw was observed at Ignalina in 1983, three years before Chernobyl. The knowledge didn't reach the operators who needed it.

So the lesson isn't only "review the design." A design review is worth nothing if what it finds gets buried, diluted or never asked. Chernobyl was a design flaw that survived because the review-and-disclosure system around it was weak.

That's the system we're now changing, at speed.

The copy-exact problem

The new wave adds a risk that one-off plants never had.

Factory-built, standardized reactors are a real quality advantage. In semiconductors we called the discipline Copy Exact: once a process is proven, every fab runs it identically. It works beautifully. It also means a latent defect ships everywhere at once.

The 737 MAX is the cautionary version: one design assumption, propagated across a fleet, found in service. When the plan is hundreds or thousands of identical units, the cost of a missed flaw scales with the fleet. That argues for putting the heaviest scrutiny on the first design.

Then there's export. The IAEA sets standards and runs peer reviews, but it can't enforce them. If U.S. vendors sell standardized reactors to countries with thin regulatory capacity, a U.S. approval becomes the de facto safety review for buyers who can't check the work themselves. The quality of that approval travels with the product.

The supply chain is a safety system

There's a quieter version of the same gap in the supply chain.

The pool of suppliers holding the ASME N-stamp or keeping NQA-1 quality programs has been shrinking. The OECD Nuclear Energy Agency has flagged counterfeit and falsified-certificate items already delivered into the nuclear industry in several countries.

Anyone who's run supplier quality through a demand spike knows how this goes. Qualification takes years, demand arrives in quarters, and the pressure to accept "close enough" peaks right when inspection capacity is thinnest. A counterfeit valve is far more likely than a meltdown, and far less likely to make headlines.

The horizon mismatch

The flattering number is the gigawatts. The operating number is reviewer-years.

Even the gigawatts need a second look. A large share of that 9.8 GW comes from power agreements with plants that already run, like Susquehanna and Clinton. Those reactors have decades of operating history and an established oversight regime. The part that needs new review is the new-build share: the TerraPower, Oklo, X-energy and Kairos units. That's a smaller number than the headline, and it's almost entirely first-of-a-kind.

Most of the SMR-based data center deals point to 2030 or later. That makes them feel like a future problem. They aren't. The safety review for a reactor that runs in 2031 happens in 2026 and 2027, done by whoever is on the bench right now. Reviewer competence takes years to build. You can't hire it in the quarter you need it.

The deal-flow curve is set by capital, which moves fast. The oversight curve is set by experience, which doesn't. That's what makes it a constraint rung in the same family as consent. Money can fund the hiring. It can't compress the learning.

What I'd watch

If you want to know whether the follow-up is keeping pace, these are the dials:

• Reviewer capacity against the docket. Headcount and experience of NRC technical reviewers compared with the number of active advanced-reactor applications. GAO noted the NRC had a plan to use its new hiring authorities but no metrics to show whether they work. That's the first metric worth publishing.
• Lead-unit hours before fleet orders. How much real operating time a first-of-a-kind design logs before multi-unit deployments commit. Measured performance should come before mass production.
• The DOE-to-NRC handoff. Whether operating and transient data from the DOE test reactors is published and used to check vendor safety claims before commercial licensing.
• Peer accountability for new entrants. Whether startup operators join INPO-style peer review before fuel load, not after their first event.
• Qualified suppliers against the order book. N-stamp and NQA-1 supplier counts compared with announced unit orders.
• Export review standards. Whether a design has to be proven at home before it's sold abroad.

None of these require slowing the buildout. They require building the oversight on the same schedule as the deals.

The more important conversation

We spend a lot of public attention on whether AI might someday turn against us. That's a fair question. It's also speculative, and it's mostly out of any one person's hands.

This one isn't. The power plants being built to run AI are concrete, near-term and governable with tools we already know work: independent review, peer accountability, conservative operations and honest disclosure. We've done it before, at scale, under pressure.

The deals are moving at deal speed. The question is whether the review keeps up, or whether we learn what it missed the way Chernobyl did: three years after someone had already seen it.