Who Owns the Kitchen

On July 10, Marketplace ran a segment explaining how a data center can lower your electricity bill. It used a buffet. You are throwing a party, the food costs a thousand dollars, and the more guests who chip in, the less each one pays. The data center is the cousin who shows up with four friends. The fixed cost of the grid is the spread.

The analogy is sound, and the segment did not stop there. It ran the party past the point where the food runs out. More friends arrive, they are hungry, you have to go buy more, and now everyone pays more. Ryan Hledik of the Brattle Group named the condition precisely: the states that saw prices fall already had room on the grid to take the new customers without building anything. He named North Dakota. He also named where the condition has already failed, which is PJM, the mid-Atlantic and parts of the Midwest, where demand has outrun the available generation.

That is honest work under a six-minute constraint. It is also, in one respect, incomplete, and the gap is where the money lives.

In the buffet, buying more food is a burden. The host absorbs it and hates it. In the American electricity system, buying more food is the business model. A regulated utility earns an authorized rate of return on the capital it deploys. New generation, new transmission, new substations: these are not costs the utility grudgingly bears. They are the asset base on which its profit is calculated. The host does not resent the grocery run. The host is paid a guaranteed margin on every tray.

The analogy works until you ask who owns the kitchen.

01

The Condition, and Who Deleted It

Everything turns on the word headroom. The Electric Power Research Institute study underlying the Marketplace segment found that average residential rates in the average state would have been about 6% higher without the data centers built from 2019 to 2024. That finding is real, and EPRI is the electricity industry's own research arm, not an advocacy shop.

Read the sentence again, though, because every load-bearing word is a limit. Average. 2019 to 2024. Past tense. It is a measurement of the era when the grid had empty seats, and it says nothing about what happens once they are filled. Asa Watten, the EPRI co-author, said exactly that. Hledik said exactly that. Marketplace put it in the headline: in some places, for now.

Now set that beside the Secretary of Energy. Speaking at the Cato Institute, Chris Wright said data centers are “the answer to drive down electricity prices. They're not the problem. They're the answer.” At an Amazon Web Services summit, he said electricity prices are “not one of the downsides.” At the White House signing of the Ratepayer Protection Pledge, the President went further, saying prices for American communities “will not go up, but in many cases, will actually come down.”

No condition. No geography. No window. The researchers who produced the finding scoped it carefully. A public radio business program scoped it carefully. The administration citing the finding did not scope it at all.

This is not a dispute about the mechanism. The mechanism is real. It is a dispute about its reach, and the reach is the entire question, because the conditions under which the claim holds are anticorrelated with the places the load is actually going.

02

Where the Headroom Is, and Where the Load Is

Headroom exists where an industrial base died and the wires stayed. North Dakota, with surplus coal, gas, and wind. Stranded capacity in places built for a demand that never arrived or has since left.

Data centers want fiber trunks, low-latency proximity to enterprise customers, water, and workforce. That points at Loudoun County, at Dallas, at Columbus, at Atlanta. It does not point at Bismarck, except at the margin, and the margin is small and self-terminating. Whatever gets built in a low-density state fills the empty seats, and then the seats are full, and the next data center is back to greenfield construction like everywhere else.

So the exception does not generalize. It expires. And in the meantime the population-weighted reality is already visible: in Virginia, the state with the highest data center concentration in the country, electricity prices have risen by up to 267% over the last five years. Utilities received data center interconnection requests totaling at least 700 gigawatts in 2025 alone, against total United States consumption of 477 GW across all of 2023.

The evidence that data centers lower prices comes from the era of spare capacity. The buildout being justified with that evidence is the thing that ends the era of spare capacity. They are spending the credibility of a condition they are in the process of destroying.

03

The Escape Hatch

There is a real answer to the headroom problem, and its defenders will reach for it, so it is worth taking seriously rather than dismissing. Retired generation is not demolished generation. Plants that closed for economics rather than engineering still have the site, the interconnection, and the wire. Restart one and you have added supply without a greenfield build.

That argument is correct. The question it invites is the only one that matters: who pays for the restart?

There are three places the cost can go. Into the rate base, where the ratepayer funds it and the utility earns a return on it. Into a federal instrument, where the public funds it and it never appears on a bill. Or onto the tenant, through a power purchase agreement that covers the refurbishment capital, which is what the Ratepayer Protection Pledge implies and what the pledge has no power to compel, being nonbinding, with no enforcement mechanism ever specified.

Only the third supports the claim being made. Here is where it actually went.

The Crane number is the one to sit with. In its 2024 annual report, Constellation told its shareholders the project would require approximately $1.6 billion of cash from operations. Not from federal credit. From operations, backed by a twenty-year power purchase agreement with Microsoft that supplied the revenue certainty making the economics work.

The company said it could pay. Fourteen months later, the Department of Energy committed a $1 billion credit facility anyway, guaranteed through Title XVII and funded by the Federal Financing Bank, at 0.375% over Treasury on a term running to 2055. That is an extraordinarily cheap thirty-year loan for a merchant power project, below anything Constellation could obtain in the private market. DOE's own project description states the purpose: supporting 835 MW of baseload power for data centers. The plant's entire output is contracted to one of the largest companies on earth. Crane is also expected to qualify for the clean electricity production tax credit under Section 45Y for its first ten years.

One detail sharpens rather than softens the point. As of the November 17 closing, Constellation had not drawn a dollar, and is not required to make its first borrowing until May 2027. This is a commitment, not a disbursement. But the commitment is the decision. The underwriting is done, the credit-risk judgment is made, and the full faith and credit of the United States is behind the project now, whether or not a single advance is ever requested. The government did not respond to a funding gap. It extended the backstop before the borrower had asked for a cent of it, to a company that had already told its shareholders it did not need one.

A guaranteed loan is not a grant, and that distinction should be made honestly. If Constellation repays, the public's cost is the interest-rate subsidy and the contingent default risk, not the face value. But the direction is not ambiguous. The public balance sheet stands behind up to 80% of the eligible cost of a project the private parties had already stated they could fund themselves, at below-market cost, on a thirty-year term, with a production tax credit layered on top.

The Ratepayer Protection Pledge says the tech companies will cover the cost of new generation. At Crane, the federal government stood behind up to 80% of it.

04

What Nobody Notices

Set the ledger against the scale of federal revenue and the reason none of this drew a headline becomes obvious.

Take the most conservative tariff figure available, one that no litigation can claw back: customs duties raised roughly $79 billion in calendar year 2024, before any of the contested authorities existed. That is about $216 million a day. Against that floor, the entire Palisades guarantee is about seven days of collections. The Crane commitment is under five. The disbursed Palisades total is a little over two.

Nobody notices $1.3 billion.

That is the whole trick, and it is not a conspiracy. It is arithmetic. A number can be simultaneously enormous and invisible, and it is invisible precisely because the numbers around it are so much larger. Enough to restart a nuclear plant. Small enough to move through six routine agency press releases and read as administrative housekeeping.

05

The Instrument Already Exists

Here is what makes this a proposal rather than a complaint. The mechanism that would fix it is not hypothetical. It is the one they are already using.

Federal credit programs work through the Federal Credit Reform Act: an appropriation covers the expected loss on a portfolio, not its face value, so a dollar of subsidy budget authority supports many dollars of lending. That is the real leverage, and it is how a $1.52 billion guarantee costs the government a fraction of $1.52 billion. The Department of Energy has now deployed this instrument at Palisades and committed it at Crane. It works. It is the tool that can move restart capital off the ratepayer.

So the policy ask is narrow and it is aimed at people who have already conceded the tool: if federal credit is going to finance generation restarts that anchor data center load, then access to that credit should carry a condition foreclosing recovery of the same capital expenditure through a state-regulated rate base. Not a pledge. A covenant. Pledges are nonbinding and this one was never given teeth. Loan agreements have conditions precedent, and conditions precedent are enforceable because they are contractual.

The distributional argument has to be made out loud rather than assumed: moving a cost from the ratepayer to the federal balance sheet does not make it free. The public still carries it. But the ratepayer base is captive and regressive, with no ability to opt out of the electric bill, while the federal tax base is broader and progressive. That is a real difference and it is the reason to prefer one door over the other.

The Question

If the public is subsidizing nothing, then there is nothing to hide.

Every federal loan carries a credit subsidy cost, computed by the lending agency under the Federal Credit Reform Act. It is a government calculation, not a corporate secret. It states, in dollars, what the taxpayer is carrying. If that number is zero at Crane, the Department can say so tomorrow and this piece is wrong. If it is not zero, then the public is financing generation for a private tenant while the Secretary of Energy tells the public that the tenant will drive their bills down.

We have asked.

The three documents they produced are the ones already on EDGAR. The one number that would settle the question, the credit subsidy cost the government computed before it made the commitment, is still, as they put it, under review. This is one promised document. It is not the answer. The answer, or the continued absence of one, will be published here.