Kilowatt Kings

---

There is a particular kind of danger that arrives not when a sector is hated but when it is reinvented. For a hundred years, owning an electric utility meant owning the most predictable cash flow in the economy. People keep the lights on in recessions and booms alike; regulators grant the utility a fixed return on its assets; the dividend arrives every quarter; the stock barely moves. Utilities were where you put money you could not afford to lose and did not expect to grow. The whole sector was a synonym for safe and slow — a bond proxy, traded on its yield, valued at low-teens multiples of earnings because nobody expected those earnings to do anything but plod gently upward with the population.

That sector no longer exists, at least not in the market's mind. In the space of about two years, a handful of independent power producers — Constellation Energy, Vistra, NRG, Talen — have been mentally re-filed by investors from "utilities" into "AI infrastructure." Their stocks have multiplied. Their valuations have detached from a century of precedent. Constellation Energy now carries a market capitalization near $98 billion and trades at a forward price-to-earnings ratio around 26 times; Vistra is valued near $52 billion at roughly 18 times forward earnings, while carrying about $19.6 billion of net debt. These are not utility multiples. These are growth multiples, the kind the market pays for software and semiconductors — applied, for the first time in the history of the business, to companies that boil water to spin turbines.

The justification is a single word, and the word is data centers. The story is genuinely powerful, and it is worth stating in its strongest form before taking it apart.

The strongest version of the bull case

Artificial intelligence is voracious for electricity. Training and running large models requires vast data centers packed with power-hungry chips, and those data centers need electricity on a scale that dwarfs anything the grid has had to absorb in decades. Goldman Sachs has projected that U.S. data-center power demand could roughly double by 2027 and rise on the order of 165% by 2030. After twenty years of essentially flat American electricity demand — efficiency gains canceling out growth — the country is suddenly staring at a step-change in load, and the companies that own generation, especially carbon-free, always-on generation, hold something newly precious: the one input the entire AI build-out cannot proceed without.

Nuclear sits at the heart of this. AI data centers want power that runs 24 hours a day, every day, and that does not emit carbon — because the hyperscalers buying it have made loud climate commitments. Nuclear is almost the only source that delivers both at scale. And so the owners of America's nuclear fleet have found themselves courted by the richest companies on earth. Constellation, which operates the largest nuclear fleet in the United States and generates around 90% carbon-free power, has signed long-term contracts covering more than 5,650 megawatts of clean energy for hyperscalers — including a deal to restart a reactor at Three Mile Island (rebranded the Crane Clean Energy Center) to supply Microsoft, and a 20-year, roughly 1.1-gigawatt agreement with Meta. In January 2026, Vistra signed its own headline deal: 20-year power-purchase agreements with Meta for about 2,600 megawatts of nuclear output across three plants in the PJM region.

These power-purchase agreements — PPAs — are the bull case made concrete. They are long (twenty years), large (gigawatts), and signed by the most creditworthy buyers imaginable. They convert a power producer's once-volatile merchant earnings into something that looks like contracted, visible, growing cash flow. Constellation has gone so far as to put a multi-year earnings bridge in front of investors: adjusted operating EPS of $11.00 to $12.00 in 2026, and a path to base EPS growth above 20% a year through 2029. Twenty percent compound earnings growth from a utility — that is the number that justifies the re-rating, and if you believe it, the stocks are not even expensive.

That is the case. It is coherent, it is partly true, and it is exactly the kind of story that sweeps a sector to prices from which it cannot easily return.

The forecast quietly being cut

Now the part the bull case does not put on the slide. Every dollar of the re-rating depends on the demand forecast being real — on those data centers actually getting built, actually drawing the projected load, actually on the projected schedule. And the single most authoritative body on the question, the organization whose entire job is to know how much electricity the eastern United States will need, has started revising the number down.

In January 2026, PJM Interconnection — the grid operator that runs the thirteen-state eastern system from the mid-Atlantic to the Midwest, the largest grid operator in America — cut its peak-demand forecast for the summer of 2027 to about 160 gigawatts, down from a prior outlook near 164 gigawatts. Four gigawatts may sound modest, but the reason is the story. PJM lowered the forecast because many of the projects baked into the previous outlook — data centers prominent among them — do not yet have firm electric-service or construction commitments. In plain English: a large share of the projected AI demand was speculative. Developers had announced projects, utilities had pencilled them into the forecast, and on inspection the projects lacked the firm commitments that distinguish a real load from a press release. PJM is now applying stricter vetting to data-center requests precisely because it can no longer tell, from the announcements alone, what is real and what is hype.

PJM was careful to say it still expects enormous long-term load growth through the 2040s — it cut the near term (the forecasts through roughly 2032) while reaffirming the distant future. But that distinction cuts the wrong way for the bulls, not the right way. The near-term forecast is the one that underwrites the cash flows a 26-times multiple is paid for today; the far-future forecast is the unfalsifiable part, two decades out, that no one can disprove and everyone can dream on. When the grid operator trims exactly the dates that matter for present valuation and leaves intact only the dates too distant to test, it is telling you which part of the demand it can actually see — and it is not the part doing the heavy lifting in the stock price.

Read that again with the valuations in mind. The companies have been re-rated from bond proxies to growth stocks on the strength of a demand forecast. The body that produces the demand forecast is trimming the near-term portion of it, and is openly saying the reason is improved vetting — that some of the demand it had penciled in may not be firm. When the grid operator itself — an entity with no stock to talk up, whose only mandate is to keep the lights on and not over-build — starts stripping out unconfirmed data-center load, the market should listen more carefully than it did to the developers' announcements. The market, for now, mostly has not.

The price action when the news broke was a tell. On the forecast revision, the largest independent power producers fell together: Vistra dropped as much as 3.1%, Constellation as much as 2.3%, NRG as much as 2% — a correlated lurch that revealed how tightly these stocks are now bound to the single variable of projected AI load. They no longer trade like utilities, on their dividends and rate bases. They trade like leveraged bets on a forecast, and they move as one when the forecast moves. That is the signature of a sector that has stopped being defensive and become, instead, a thematic trade.

What the bears at the rating desks already see

This is not a lone skeptic's reading. Morningstar — hardly a sensationalist outfit — assigns both Vistra and NRG a High Uncertainty Rating and has had them trading in 1-star territory, its label for stocks it considers significantly overvalued. Its analysts' description of how the sector got here is worth quoting in spirit: the shares shot up as electricity-demand estimates soared, and have wobbled as regulators introduced price caps and those same demand estimates began to moderate. The thesis that lifted the stocks — soaring demand forecasts — is the thesis now being revised, and the stocks have begun, fitfully, to feel it.

The deeper problem is structural, and it is the mismatch the bull case papers over: the timing. The PPAs are twenty years long. The data centers take years to build. The nuclear restarts and new builds that are supposed to feed them are among the most schedule-cursed projects in all of industry. Reactor projects run over budget and behind schedule with grim regularity; small modular reactors — the technology the optimists invoke for the back half of the decade — may not reach meaningful commercial deployment until past 2030, if then. So the cash flows that justify a 26-times multiple sit years in the future, while the demand that underwrites them is being marked down in the present. A long-duration promise valued at a growth multiple is the most fragile thing in markets, because its entire worth lives in a forecast, and forecasts that far out are exactly the ones most likely to be wrong — and, as PJM has just demonstrated, most likely to be revised toward sobriety as the speculative projects fail to firm up.

The circularity nobody wants to name

There is a circularity at the center of the AI-power trade that mirrors the circularity in the AI-chip trade, and it deserves to be named plainly. The hyperscalers signing these twenty-year PPAs — Microsoft, Meta, Amazon, Google — are themselves making enormous, partly speculative bets on the future demand for artificial intelligence. They are contracting for gigawatts of power on the assumption that AI workloads will grow to fill them. The power producers, in turn, are being valued on the assumption that those contracts are rock-solid signals of real, durable demand. But the contracts are only as solid as the AI demand that justifies them in the buyers' own minds. If the AI build-out disappoints — if the workloads do not materialize on the projected curve, if efficiency gains in chips and models reduce the power needed per unit of intelligence, if the hyperscalers' own capital discipline reasserts itself — then the demand the whole edifice rests on softens at the source.

A twenty-year PPA signed by Meta is a strong contract; Meta will pay. But the marginal demand, the next tranche of projected load that has not yet been contracted and that fills out the bullish 2030 forecasts, is precisely the speculative portion PJM is now discounting. The signed deals are real. The forecast above and beyond the signed deals — which is where the growth multiple lives — is the part being cut. Investors are paying a growth multiple for the contracted base plus an enormous assumed wedge of future demand, and that wedge is the soft part. You are not buying twenty-year cash flows at 26 times earnings. You are buying twenty-year cash flows plus a forecast, and paying the growth multiple on the forecast.

We have re-rated boring before

The market has a recurring habit of taking a stable, slow, defensible business and, in the grip of a story, repricing it as a growth stock — and the repricing, not the business, is what later breaks. It happened to the railroads when they became "the new infrastructure." It happened to pipelines in the shale boom, sold as growth and then re-rated brutally back to income when the volumes and the distributions disappointed. It happened to consumer-staples "bond proxies" in the low-rate era, bid to growth multiples for their yield and then de-rated when rates rose. In each case the underlying business survived and even prospered; what cratered was the multiple, the portion of the price that the story had inflated beyond what a sober view of the cash flows could support.

Utilities are now living through their version. The business is real, the AI demand is partly real, the PPAs are real. But a utility is, in the end, a capital-intensive, regulated or merchant generator of electricity, with a cost of building that runs over budget, a schedule that runs late, and a demand curve set by forces — economic growth, weather, efficiency, and now the uncertain trajectory of AI — that no power producer controls. To value that business at the multiple of a software franchise is to bet that this time the boring business has permanently become a growth business. The grid operator's January forecast cut is the first data point suggesting that the boring business is, underneath the story, still the boring business — one whose demand can be over-estimated, and is now being trimmed.

What breaks, and how

It is worth being precise about the mechanism of risk, because it is not "the lights go out" — the lights will not go out, and these companies are not going bankrupt. The risk is narrower and more financial: it is de-rating. A stock trading at 26 times earnings on the promise of 20%-plus growth does not need the growth to vanish to fall hard; it only needs the growth to slow, or the confidence in the growth to slip. If AI power demand grows at, say, half the breathless forecast — still an enormous outcome by any historical standard — the earnings might grow nicely, but the multiple would collapse from a growth multiple toward a utility multiple, and the stock would fall by a third or a half even as the business kept expanding. That is exactly what happened to the pipelines and the staples. The earnings were fine. The multiple was the trade, and the multiple was wrong.

For these stocks, the de-rating trigger is sitting in plain sight: a demand forecast that the most credible forecaster is already cutting, regulators introducing price caps that limit how much of the AI boom flows to producer profit rather than consumer bills, nuclear restart and new-build schedules that have every historical reason to slip, and a circular dependence on a hyperscaler capex cycle that is itself the most-watched bubble question in the market. Any one of these, sufficiently disappointing, takes the multiple down. Several of them are correlated, which means they are likely to disappoint together, in the same risk-off moment, just as they rallied together in the same risk-on years.

The leverage and the gas underneath the nuclear story

There is a further wrinkle the clean nuclear narrative tends to obscure, and it sharpens the risk rather than softening it. The two flagship names are not pure carbon-free plays, and they are not unlevered. Vistra acquired its nuclear fleet only recently — through the 2024 purchase of Energy Harbor — and the bulk of its generation remains a large fleet of natural-gas plants whose earnings rise and fall with commodity prices it does not control. So the "nuclear AI utility" thesis is, for Vistra, bolted onto a merchant gas business with all the cyclicality that implies, and it is carried on roughly $19.6 billion of net debt. Debt is the amplifier that turns a multiple de-rating into something worse: a company valued for secular growth, leaning on borrowed money, whose merchant earnings swing with gas prices, is precisely the profile that performs beautifully while the story holds and painfully when it cracks. Leverage does not care about the narrative. It compounds whatever direction the cash flows take.

Constellation is cleaner — the largest U.S. nuclear fleet, around 90% carbon-free — and that cleanliness is exactly why it commands the richer multiple, near 26 times forward earnings against Vistra's 18. But "cleaner and more expensive" is not the same as "safe." It means investors have paid up the most for the name most completely dependent on the AI-demand forecast holding, and have the least valuation cushion if it does not. The premium that looks like quality in a bull market is the same premium that has the furthest to fall when the growth assumption is questioned. Pay 26 times for 20% growth and you are insulated by nothing; the price is the forecast, expressed as a multiple, with no margin of safety left over. The market has arranged things so that the highest-quality asset carries the highest forecast risk — which is usually how the most respectable names end up doing the most damage when a theme unwinds.

Even the smaller names in the basket — NRG, Talen, GE Vernova, Cameco, Dominion — have been swept up by the same current, bid as a group whenever the AI-power theme runs and sold as a group when it stumbles. Talen has paired with reactor developers to assess next-generation small modular deployments; NRG and the rest have ridden the same demand estimates. The correlation is the warning. A genuinely diversified set of businesses moving in lockstep is not trading on its individual fundamentals. It is trading on a single shared story, and a single shared story has a single point of failure.

The tell in the dividend

Here is the cleanest way to see what has happened. Ask the oldest question you can ask of a utility: what is its dividend yield? For a century, that number — usually a healthy 3, 4, 5% — was the investment case; you bought the utility for the yield and held it through the boredom. Now look at the AI-power darlings, and the yield has been compressed toward irrelevance by the surge in the share price. The thing that made a utility a utility — the fat, safe, defining dividend — has been bid away, because the buyers are no longer there for income. They are there for growth, for the AI theme, for the next leg of a momentum trade. The yield's disappearance is the re-rating made visible: these are not being held as utilities anymore, by anyone, for anything.

And that is the quiet danger. The widows and pension funds who owned these names for safety have, in many cases, been replaced by momentum and thematic capital that owns them for the story — capital that will leave as fast as it came the moment the story wobbles. A sector whose entire historical identity was stability has been hollowed out and refilled with the least stable money in the market, at the least stable point in its valuation history, on the strength of a forecast the grid operator is busy revising downward. The kilowatt kings are real companies with real reactors and real contracts. But they are being priced for a coronation that depends on a number already being cut — and the market that crowned them has shown, again and again across a century of manias, that it takes back the crown far faster than it bestows it. The lights will stay on. The multiples are another matter entirely.