Two men said the quiet part out loud last week, and they contradicted each other without meaning to.
Rolls-Royce CEO Tufan Erginbilgic told analysts a "big hyperscaler" nuclear deal would be signed "early next week" — small modular reactors, the future of always-on AI power, imminent.
NextEra CEO John Ketchum, announcing the largest AI-power deal of the year, chose natural gas and batteries. No reactors. Not one.
The Department of Energy had explicitly invited nuclear. Its November 2025 Request for Offers, which reportedly applied to the Paducah site among others, asked bidders to integrate small modular reactors and nuclear capacity, per TFTC. The winning plan ignored it.
That gap — between what the industry keeps promising and what it actually builds when real money and real deadlines are on the line — is the whole story of the AI power crisis. Everything else is press-release theater.
That gap — between what the industry keeps promising and what it actually builds when real money and real deadlines are on the line — is the whole story of the AI power crisis.
Last week we traced AI capex migrating off hyperscaler balance sheets into structures that don't show up as corporate debt. This week that thread hardened into concrete: a $100 billion campus in Kentucky and a $15 billion loan package in Texas. The abstraction now has an address.
Start with the number nobody put in a headline.
To power an AI campus in Paducah, Kentucky, that will support up to 1.8 GW of utility capacity, Brookfield and NextEra are building up to 4.6 GW of dedicated generation from scratch — 2 GW of natural gas plus 2.6 GW of batteries.
That's roughly 2.6 gigawatts of nameplate generation for every gigawatt of campus load (a 2.6x ratio), per TFTC. A 2.6x overbuild, funded entirely privately, with zero cost passed to existing ratepayers.
You do not build 4.6 GW to serve 1.8 GW because you're flush. You build it because you don't trust the grid to deliver a single reliable electron — and you're right not to.
The tell is in Ketchum's own framing. "The data center will bring its own power, pay for its own power infrastructure," he said, calling it "a proof point for how AI infrastructure should be built in America."
Read that again. The largest utility in the country is celebrating the fact that the grid it operates cannot serve the load, so the customer has to build a private power plant to get connected at all.
Data Center Knowledge called Paducah a testbed for "bring-your-own-power" campuses. That phrase should chill anyone who owns these stocks. It means the grid is no longer infrastructure you plug into. It's a thing you replicate yourself.
Now the second discipline: don't take that 4.6 GW at face value either.
Neil Osnato of Persistence Analytics Group put it precisely. "The first thing I would not do is add those resources together and call it 4.6 GW of equivalent firm generation. A gigawatt of battery capacity is not the same thing as a gigawatt of continuously available generation."
Batteries discharge and then they're empty. The 2.6 GW of storage smooths peaks and covers gaps; it does not run a training cluster around the clock for months. The firm, dispatchable spine of this project is the 2 GW of gas.
"Bringing your own power is a major improvement," Osnato said. "It is not the same as bringing your own grid." And his warning generalizes to every deal in this space: "Do not count dedicated megawatts by nameplate. Verify what they can actually deliver when the system is stressed."
Apply that same test to the nuclear story, and it collapses.
The celebrated number is 9.8 GW — roughly 13 announced deals committing that much nuclear capacity to AI data centers. Microsoft restarting Three Mile Island. Google's Kairos SMRs. Meta's up-to-6.6 GW portfolio across TerraPower, Oklo, Vistra, and Constellation, per Nexi.fund.
It sounds like a power system. Check the delivery dates and it's mostly a press-release system.
Less than 10% of the nuclear capacity the industry needs will be available by 2030. TerraPower reportedly won the first construction permit for a commercial non-light-water reactor in over 40 years — a genuine landmark, in March 2026 — for a plant that will not produce power before 2030, per Nexi.fund.
Google's first Kairos reactor targets "around 2030." Amazon's X-energy units, early 2030s, per ROC Telecom.
Microsoft's Three Mile Island restart — 835 MW, targeted for 2027 — reportedly delivers this decade, and that's not new nuclear at all. It's an existing reactor being switched back on.
Here is the timeline reality, in the words of one industry writer: "the reactors arrive years after the hardware they will power, which will refresh several times over before the first nuclear electron flows."
Sit with that. The GPUs going into Paducah will likely be depreciated, retired, and replaced before any SMR sends its first electron to an AI campus. You are being sold reactors to power hardware that won't exist by the time the reactors turn on.
Which is why, when the DOE asked for nuclear at Paducah, the market answered gas. And batteries. Because those are the only firm, dispatchable, reportedly buildable-by-2031 electrons on the menu.
The nuclear rush is real as a decade-out bet. As a source of power for this AI buildout, it is fiction — marketing that papers over a gas-and-debt present.
Rolls-Royce's own numbers admit it. The company now expects 25% annual power-generation growth through 2030, and CFO Helen McCabe said data-center power orders grew more than 50% in the first half, per CNBC — but the demand is for gas engines and backup engines that run around the clock "because they can't get connected to the grid." Even the company teasing SMR deals is booking its actual revenue on combustion.
Now connect the physical squeeze to the financial one.
If the grid can't deliver the power, the developer has to build it — and a 4.6 GW private generation stack plus $100 billion of campus isn't going on a balance sheet as equity. It goes into project debt.
Enter the second deal of the week: Morgan Stanley reportedly leading a $15 billion debt package for a Google-backed Anthropic data center in Texas, per briefs.co.
This is exactly the mechanism we flagged last week, now made concrete. The capex is being routed off the hyperscaler's books into structured, project-level borrowing, backstopped by a strategic partner.
Physical scarcity and financial fragility are compounding. You have projects that must overbuild generation 2.6-to-1 because the grid failed, financed increasingly with debt that sits at arm's length from the companies whose names are on the AI.
Osnato's discipline applies to the financing too: verify what it delivers when the system is stressed. Nameplate capex commitments are not firm balance-sheet strength any more than 2.6 GW of batteries is firm power.
So here's what to actually watch, and it's narrow enough to be useful.
The Trump administration reportedly named four federal sites — Paducah, Idaho National Lab, Oak Ridge, Savannah River — as templates to replicate, per the DOE. Paducah and Savannah River both reportedly paired the site with dedicated bring-your-own generation, per TFTC.
The tell is simple. If the next federal-site deals still require the developer to build their own power plant rather than interconnecting to existing grid capacity, the constraint is structural, not cyclical. The grid isn't catching up.
And if any of them are backed by SMRs delivering firm power before 2030, come find me. I'll retract every skeptical word.
Until then, the trade is unglamorous and clear. The AI boom does not run on chips or capital. It runs on firm, dispatchable electrons — and right now those come from gas, get smoothed by batteries, and get paid for with debt.
The reactors are a story for the 2030s. The power bill is due now.