# “Thirty to Fifty Cents of Every Overrun Dollar”: Where US Nuclear Financing Actually Gets Stuck

Source: Realpha Blog (blog.getrealpha.com)
Original article and charts: https://blog.getrealpha.com/en/blog/columbiaenergy-2026-10-20-doe-s-julie-kozeracki-on-the-federal-role-in-us-nu/

> Notes from the October 20, 2026 Columbia Energy Exchange with Julie Kozeracki of the DOE's Office of Energy Dominance Financing: how a 30–50% investment tax credit caps overrun risk, why a $17.5 billion long-lead-item order buys back three years, and why the real obstacle to new build sits with equity investors rather than permitting. Educational notes, not investment advice; no price targets or stock recommendations.

Published: 2026-10-21
Locale: en
Tags: nuclear, energy policy, infrastructure finance, us power, policy risk


![A nuclear plant yard at dusk, a line of identical large equipment modules on rail cars shrinking from the foreground toward a distant containment dome, an engineer's back visible beside the nearest module](/covers/columbiaenergy-2026-10-20-doe-s-julie-kozeracki-on-the-federal-role-in-us-nu-cover.png)

> In the economic sphere, an act, a habit, an institution, a law produces a series of effects. Of these effects, only the first is immediate; it appears at the same moment as its cause, and so it is seen. The others emerge only in succession, and they are not seen.
>
> —— Frédéric Bastiat, "That Which Is Seen, and That Which Is Not Seen" (1850, my own translation)

In the October 20, 2026 episode of Columbia Energy Exchange, host Jason Bordoff interviewed Julie Kozeracki, acting co-chief investment officer and nuclear sector lead at the US Department of Energy's Office of Energy Dominance Financing. She put a number on something I had never done the arithmetic for: with the investment tax credit running between 30% and 50%, the federal government picks up thirty to fifty cents of every dollar a project overruns. Her office also issued a $17.5 billion conditional commitment in June to buy the longest-lead equipment for ten AP1000 reactors up front. Her read is that new build is now held up by utilities worrying about their shareholders, with licensing and technology risk behind that — a read that holds for the AP1000 specifically, since by her own account reactors nobody has built yet belong with grants rather than senior debt.

## What the episode covers

The show is the weekly podcast of Columbia University's Center on Global Energy Policy. Many listeners will know the office by its old name, the Loan Programs Office; it was created by the Energy Policy Act of 2005 with nuclear in mind, and after its 2022 expansion it holds more than $200 billion of lending authority across energy infrastructure. Kozeracki flagged the clock on that: a large share of the authority expires in under 24 months, which is why they are moving fast.

The backdrop matters. The US has 94 operating reactors at 54 sites supplying roughly a fifth of national electricity, and only three new units have come online in the past thirty years. Four executive orders signed in May 2025 set the target: ten large reactors under construction by 2030, plus five gigawatts of uprates at existing plants. Bordoff kept circling one question — with power demand surging and hyperscalers carrying enormous balance sheets, why does any of this need federal money?

## Six things worth keeping

### 1. "Complete design" means something stricter than outsiders assume

Kozeracki's line stuck with me: a complete, constructible nuclear design is not "it's a pressurized water reactor and it's 1,100 megawatts" — it is that the pipe goes *here*, and not three quarters of an inch over there. When Westinghouse received its design certification from the Nuclear Regulatory Commission in 2006, the team believed they were about 50% design complete. In hindsight they put it closer to 5%, because the regulator concentrates on safety-related components, and an AP1000 contains something like 17,000 cables. Those details only close out once you are building the thing.

![Two horizontal bars of starkly different lengths: the upper one runs about half the full width, the lower only a short stub, showing the gap between the design completion claimed at the time and what it looked like in hindsight.](/figures/design-completion-gap-en.svg)

### 2. America removed its own learning curve

Those 94 reactors were built from more than fifty different designs. Kozeracki describes that as having taken standardization away from ourselves. Which is why the $17.5 billion matters less for its size than for what it buys: containment vessels, steam generators, structural modules — standardized to the point of being interchangeable between projects, organized as five two-packs of AP1000s across five loans. Placing those orders now, she estimates, pulls roughly three years off the commercial operation dates. The side effect she described is the interesting part: scoping the conversation down to procurement, rather than full greenfield construction, is what moved utilities off the sidelines into serious planning.

![On the left, fifty small blocks of varying sizes scattered in disarray; on the right, five neatly aligned pairs of identical large blocks, showing the shift from every plant being different to standardized and interchangeable.](/figures/fifty-designs-to-five-packages-en.svg)

### 3. Interest expense, not concrete, is the dominant cost

Capital goes out early and construction runs long, so interest accruing during construction becomes disproportionate. Her own books illustrate it: about $11.5 billion of loans for Vogtle units 3 and 4 produced an estimated $3 billion of interest savings passed through to Georgia ratepayers, measured against what an investment-grade utility could have raised in the bond market. The spread itself is modest; multiplied by that much capital over that many years, it moves total project cost. Cheap, patient debt changes the final price of the electricity.

![Two bars of equal width whose lower engineering-cost segments are the same height, while the upper interest segment is clearly tall on one and short on the other, with the difference arrowed toward customer bills.](/figures/interest-not-concrete-en.svg)

### 4. The overrun risk already has a ceiling on it

This was the most concrete stretch of the conversation. Kozeracki repeated the utility refrain — "this costs more than we expect, the government needs to step in" — and answered it: the federal government has already committed to pay thirty to fifty cents of every overrun dollar. On top of that, loans can be structured with meaningful contingency, and recent ratepayer-optimized loans carry bullet maturities, 25, 27, and 30 years of interest only with no principal repayment. Her arithmetic: subtract the federal share, then discount whatever remains across three decades. For the ceiling itself, underwriters use Vogtle — a project that began from a wholly incomplete design, went back to the regulator every time the field differed from the drawings, survived a Westinghouse bankruptcy, multiple changes of EPC contractor and COVID, and landed near $15,000 per kilowatt of overnight cost in today's dollars. Assembling that pile-up again, she argues, is hard.

![One dollar cut into ten cells, with three to five cells on the left marked as taken up by the federal government and the rest borne by the project, two dashed lines marking the upper and lower edges of that share.](/figures/overrun-split-federal-share-en.svg)

### 5. The bottleneck sits in the boardroom, not the licensing queue

Kozeracki is complimentary about the NRC and used Palisades as the example: in 2023 most people thought restarting a reactor that had entered decommissioning was impossible, and within a few months the staff produced an inspection guide for restarts. Her evidence that licensing is not the constraint is blunt — there are six active combined operating licenses for large light water reactors right now, in South Carolina, Florida, Virginia and Michigan. What stops things is the equity market's reflex of "whatever the cost estimate is, double it." She also named a very practical fear: utility investors buy low-risk returns on capital expenditure, and if one or two companies step forward saying the numbers work for their board and their commission, investors may rotate into the three peers that haven't made that decision yet. So the utilities genuinely advancing are advancing quietly.

### 6. Three restarts left, and a partner who wants to stay thirty years

Between 2010 and 2022 the US shut down 13 perfectly good reactors. Kozeracki says the internals of most of them are cut into or otherwise too far gone; the three recoverable ones — Palisades, Three Mile Island (now Crane), and Duane Arnold in Iowa — all have closed loans, the last at nearly $2 billion. Asked why Constellation needed a government loan when it could finance a restart itself, her answer was about duration: these assets should run 80 years, into the 2050s and beyond, thirty years is a long time, and the last round of closures took out good machines. Having the federal government formally on the paper as a long-term partner makes her more comfortable.

![Thirteen circles in a row, three of them solid and ten struck through, showing that among the units already shut down only three can still be brought back.](/figures/thirteen-closed-three-left-en.svg)

She was also specific about what industry wants from Congress, which is useful for reading the next phase: direct pay for the investment tax credit, since a 50% credit on a $10 billion AP1000 means $5 billion of credits that the market cannot absorb without heavy losses in monetization; extending energy-community status from coal sites to nuclear sites for the investment credit as well as the production credit; and an opt-out from accounting normalization so the benefits can flow to ratepayers up front.

## The incentives are out. Why is nobody moving?

This is the confusion I get asked about most often: the policy was announced, so why do the companies behave as though they never heard it? The episode supplies a clean dissection. Subsidies address how much you lose when you lose; they do nothing about who gets punished first for moving. A utility executive faces next quarter's share price, and investors hold three near-identical substitutes. The counterintuitive result: the better the policy, the lonelier the first mover.

![Four company blocks side by side where only the first raises a flag to announce, and the arrows of capital above all bend around it toward the other three.](/figures/first-mover-penalty-en.svg)

That frame travels to any policy-driven theme. I now separate two questions that tend to get merged — how far the policy caps the downside, and whether anybody has already spent money. The first lives in press releases. The second lives in purchase orders, long-term contracts and construction permits, the things you cannot quietly walk back. When Kozeracki says utilities have already cleared their boards and their commissions and are working quietly, the testable version of that claim is whether orders get placed in the coming months.

## How do I know a cost estimate this big is believable?

I laughed at "double whatever the estimate is," because that is my own reflex with forward numbers. But the episode offers something firmer than instinct: ask which copy this is. Kozeracki breaks the Vogtle overrun into parts — incomplete design, field-versus-drawing resubmittals, a bankrupt vendor, contractors changing hands, a pandemic — and every one of them is the kind of thing that stacks up on a first-of-a-kind. Whether that risk has been retired is a checkable fact: the NRC completed its review of the twentieth revision of the design document last week, which means future AP1000s can match Vogtle unit 4 as built.

So I replace "is this estimate credible" with two smaller questions: has this design been built and operated, and is this batch standardized enough to be interchangeable? Only when both answers are yes does a cost ceiling have an anchor. She drew the boundary honestly too: for reactors not yet built or operating, you cannot put those cost and schedule parameters around the project, which is why grant funding fits them better than senior debt. That sentence carries more weight than anything optimistic she said.

## Paying for the things nobody sees

Asked at the end why it matters that nuclear works in the US, Kozeracki did not reach for decarbonization. She reached for the counterfactual. Three years ago, she notes, serious people argued seriously that solar and batteries removed the need for nuclear in a decarbonized grid — yet what turns on at night in those markets is mostly gas, with California also importing hydro and nuclear. She likes looking at the completely steady layer at the bottom of a grid chart, because that layer is what gives every other resource room to flex. The most persuasive case, she says, is Diablo Canyon: when people tried to work out how to replace a round-the-clock resource, they discovered how hard it is.

![A stacked power chart where the upper layers swing wildly while the bottom layer is perfectly flat, showing that the flexibility of everything else rests on that steady base.](/figures/the-flat-bottom-layer-en.svg)

What stays with me is the pricing failure underneath that. The steady layer makes the whole system work better, and nothing compensates the asset for providing it. That observation reaches past electricity into how we value anything: the more quietly something holds other things up, the less likely it is to be priced.

## Further reading

- Columbia Energy Exchange, October 20, 2026, Julie Kozeracki on the federal role in US nuclear financing (the source for these notes)
- Public announcements from the DOE Office of Energy Dominance Financing: the $17.5 billion long-lead-item conditional commitment in June, the conditional commitment of over $4 billion to Vistra for uprates, and the nearly $2 billion Duane Arnold restart loan
- NRC records for AP1000 design certification revisions
- Public cost and schedule records for Vogtle units 3 and 4, plus Georgia Public Service Commission filings

## The one thing to take away

The idea I kept is this: value something by pricing the version of the world without it. Kozeracki did not win the argument by listing nuclear's virtues; she removed Diablo Canyon and made you read the replacement invoice. Whatever quietly holds a system together is the hardest thing to price, because it has never been absent and you have no control group.

Here is something I have tried that needs no stock involved: pick one thing you use daily and treat as given — a step a colleague absorbs for you, a chore someone at home does without mentioning it, a routine you have kept for years — and write down what you would replace it with tomorrow if it vanished, how many hours, how much money, and who would do it. Some lines will come back blank, and those blanks are the steady layer you have never been paying for.