Newsletter · · Ashutosh Agarwal

The Grid Says No as Data Centers Bring Their Own Power - Powering AI: Grid, Gas, Generation & Nuclear - Week of August 4, 2026

A synthesis of what power, grid, gas and nuclear podcasts said for the week of August 4, 2026, as the largest US grid operator warned data centers without their own contracted generation could face involuntary power cuts, a $100 billion 'bring your own power' campus took shape in Kentucky, and Meta's latest data-center bond cleared at a 7.53% yield.

Powering AI: Grid, Gas, Generation & Nuclear

Week of August 4, 2026: The Grid Says No as Data Centers Bring Their Own Power

The grid just started saying "no." Not politely, not eventually, out loud, this week, from the biggest power market in the country. The nation's largest grid operator warned that data centers that show up without their own contracted generation could get their power involuntarily cut as soon as mid-2027. That single sentence reframes everything the podcasts have been circling for a year: this is no longer a story about who wants to build data centers. It's a story about who can actually plug one in, and what the rest of us pay while they figure it out.

TL;DR

  • The buyer now brings the power plant. A $100B campus in Kentucky pairs a 1.2 GW data center with 2 GW of new gas generation and its own uranium enrichment, the "bring your own power" model the White House is openly pushing, and the shape every mega-deal is starting to take.
  • The bottleneck moved from chips to electrons to labor and steel. One big utility says it will triple in five years the grid capacity it built over the prior 125, ~90% of it data-center driven, and it can't hire electricians fast enough. Turbines are sold out for four years. Copper and electrical steel are the quiet chokepoints.
  • The financing is the tell. Meta's latest data-center bond cleared at 7.53%, the highest yield for a blue-chip data-center deal since early 2025. The build isn't slowing, but the money is getting pickier, and that filters straight back to who gets the power.

What's New

The grid operator blinked, and it matters more than any earnings beat. This week on The TreppWire Podcast, Stephen Buschbom and Steve Haley walked through the most important number of the week: PJM, the operator of the largest US power grid (and the one serving Northern Virginia's "data center alley"), warned that data centers lacking sufficient contracted generation could face involuntary power cuts as early as mid-2027, after a capacity auction failed to secure enough supply to cover a projected shortfall of nearly seven gigawatts starting in June 2028. In plain terms: the market cleared, and there still wasn't enough electricity to go around. When the grid itself can't promise the power, the entire "just build more data centers" thesis has to route around a hard physical wall.

The answer to that wall has a name now: bring your own power. Also on TreppWire, and confirmed from the site itself on Bloomberg Talks by US Energy Secretary Chris Wright, is the NextEra-Brookfield project in Paducah, Kentucky, more than $100 billion, on leased federal land where the government once enriched uranium for the Cold War. Wright laid out the anatomy: 2 gigawatts of new natural-gas generation, a 1.2 GW data center, and 800 megawatts pushed onto the regional grid for everyone else, plus a new company, General Matter, building modern uranium enrichment on 100 acres already under construction. TreppWire added the fuller spec: a 1.8 GW campus with up to 2.6 GWh of battery storage, ~8,000 construction jobs. Wright framed it as the core of the Trump "ratepayer protection" pledge: make the hyperscaler pay for its own power so ordinary customers don't eat the bill. His most quotable reassurance to anyone worried about an AI bust: even if there's overinvestment, "it definitely won't leave the grid... we would love to have the extra power" for reshoring. (He also dispatched the water scare with a shrug: a modern data center "uses as much water as a McDonald's.")

The scale of the grid build is genuinely hard to picture, so here's a yardstick. On The MAD Podcast with Matt Turck, Samsara CEO Sanjit Biswas relayed a stat from a large US utility he'd just visited: over the last 125 years they built a certain amount of grid capacity, and over the next five years they plan to triple it. Ninety percent of that new demand, they told him, is data-center related. The binding constraint isn't ambition or even money; it's people. "There are just not enough electricians out there right now," Biswas said, which is why Meta is running programs to reskill workers into the trade. Keep that in mind every time a company guides to a big backlog: the backlog is real, but so is the labor wall standing between backlog and revenue.

And the turbines? Already spoken for. On Renewable Rides, Voltus president Matthew Plante gave the single most vivid operator detail of the week. A data-center developer in Missouri came to Voltus because it needed power fast and couldn't wait: "GE Vernova is sold out of turbines for the next four years," he said, and interconnection queues are jammed. So Voltus did something clever: in four months it signed up enough factories and businesses to free 100 megawatts of flexible demand (the developer pays each participating business about $60,000 a year to dial back when asked), letting the data center connect years sooner. His other bombshell number: in MISO, the grid spanning twelve mid-continent states, capacity prices have gone from $500 to $90,000 per megawatt-year, a 180x jump, purely on load growth. That's the price signal screaming that power, not compute, is the scarce thing.

Follow the money, and you can see the build getting harder to finance. On Wall Street Unplugged, Frank Curzio zeroed in on a detail most people missed: BlackRock's $12.5 billion bond to fund Meta's Texas data center cleared at a 7.53% yield, the highest for a blue-chip data-center deal since early 2025, about 2.8 percentage points over Treasuries, the widest spread for AA-or-better paper in three years. Roughly $600 billion of AI debt has been raised in 18–19 months; the rising cost of the newest slug says the market is starting to demand more to keep funding it. Curzio also cited a Morgan Stanley report projecting a 38 GW power shortfall against the 68 GW needed through 2028, coming up 55% short, and hyperscaler spend running near $700 billion a year. His trade read: the companies that already own power (he pointed to Bloom Energy) win as power gets scarce, and equipment names that sold off on strong prints (GE Vernova, Corning) are getting caught in the broader de-rating, not judged on fundamentals.

The Debate

This week the bull case largely made itself, but it's worth steel-manning both sides, because the one genuinely two-sided fight showed up in gas.

Bull: this is one supercycle, not four separate trades. String the week together and the same tide lifts everything. PG&E's Sumeet Singh, on Factor This, made the utility bull case better than any sell-side note could: demand that grew "1% year over year" for two decades is now a "generational" surge, and, crucially, the grid today is only 40–45% utilized on average because it's built for a handful of peak days a year. Slot new data-center load into that slack and you spread the system's fixed costs across more electrons, which is why PG&E models that every 1 GW of data-center load can cut customer bills by ~1% or more. That's the rate-base flywheel: more load, more investment, more earnings base, and, done right, lower bills, which keeps regulators friendly. Layer on turbines sold out four years deep, MISO capacity up 180x, copper as ~6% of a data center's cost with no substitute, and uranium riding a nuclear renaissance, and you get a durable, multi-year re-rating of grid, utilities, gas and fuel together.

Bear: every gas-shortage prophecy has been wrong before. The cleanest counterpunch came on Big Digital Energy, where three oil-and-gas veterans took apart a widely-shared 18-month model (by ex-Citadel analyst Matt Smith) calling for a gas shortage by 2028. The math they were rebutting: US production ~110 Bcf/d, LNG export capacity going from 15 to 35 Bcf/d (+20 Bcf/d), plus ~5 Bcf/d of new AI data-center demand, roughly 25 Bcf/d of new pull against ~20 Bcf/d of supply growth, draining the ~4 TCF of storage toward zero by 2030, which makes the ~$3.50 forward curve, in Smith's words, "not a price, that is a prayer." The veterans' retort, earned over decades: "not once has there ever been a gas shortage... the producer always comes through." Their specifics, the Permian alone throws off ~25 Bcf/d of associated gas, ~393 TCF of proved shale reserves sit in the ground, and demand simply curtails when prices spike (as it did in 2001), are the real bear case for the whole gas-for-power leg. And they made a point every long should internalize: those "ironclad" LNG export contracts are only ironclad until there's a domestic shortage and a president decides "that's our gas." The bull's tidy story assumes the plumbing behaves; the bears have watched the plumbing surprise people for thirty years.

The honest read: the load growth (PJM, PG&E, Samsara) is about as one-sided as evidence gets, nobody this week argued the demand isn't real. Where the tape genuinely splits is on gas price and the fuel mix, whether scarcity sticks or supply floods in as it always has.

The Names in Play

GE Vernova (GEV) is the name the podcasts kept circling from opposite ends. On The Uptime Wind Energy Podcast, the hosts noted that on GE Vernova's Q2 call, analysts barely asked about wind, they wanted gas turbines. GE guided that gas-turbine orders should peak in 2026 and then trend down, because the order book is already ~5 years out and buyers are "already in line." That's the bull-bear knot in one data point: turbines sold out four years deep (Voltus confirms it from the buyer side) is pricing power today, but a 2026 order peak is exactly the "front-loaded bookings ahead of an air-pocket" the bears fear. The Uptime crew also flagged a regulatory wrinkle worth watching: data centers (including SpaceX) chaining lots of small gensets to slip under an EPA size threshold, a loophole regulators are moving to close, which matters for how much behind-the-meter gas actually gets permitted.

Bloom Energy (BE) got a direct nod from Curzio as a "owns the power" winner selling off with the group. Oklo (OKLO) surfaced only in passing, as a future nuclear partner inside Schlumberger's behind-the-meter power push, a reminder that SMR names are still being discussed as someday options, not near-term supply. And Schlumberger (SLB) was the sleeper: the Big Digital Energy crew flagged its alliance with Liberty (LBRT) to sell "complete your well to power your data center" behind-the-meter power, a line SLB is guiding to double from ~$1B to ~$2B next year. Oil-services engineering, redirected at GPUs.

On the buyer side, NextEra (NEE) is the clearest beneficiary of the bring-your-own-power model via Paducah, and Meta remains the marginal price-setter for data-center debt.

Read-Throughs

The most useful way to hold this week is as a chain, where a move in one link forces a move in the next:

  • Turbines to gensets. GE Vernova sold out four years deep pushes developers toward on-site generation now, which is bullish demand for Cummins (CMI) and Caterpillar (CAT) gensets, but the EPA-loophole crackdown the Uptime hosts described is the regulatory cap on how far that substitution runs.
  • Everything to copper and electrical steel. This is the quietest, highest-conviction read-through of the week. On Mining Stock Daily, Scott North cited a study modeling copper at ~83% of the total mineral mass demanded by AI data centers, with grid transmission and distribution taking ~two-thirds of that copper, and argued the real bottleneck is processing and refining capacity, not ore in the ground. He also flagged grain-oriented electrical steel and rare-earth permanent magnets as overlooked constraints, with copper mine timelines of 15–17 years. On The David Lin Report, Brien Lundin put ~6% of a data center's cost as pure copper, "you can't innovate around it," and sees copper "over $8" a pound (from ~$6) within a few years. And on The Canadian Investor, the hosts called copper "one of the cleanest plays" on AI plus electrification, citing a view that copper "needs to triple," with China already generating ~2.5x US electricity. The read-through for Freeport (FCX) and the transformer/conductor supply chain: the load-growth thesis is a copper-and-steel thesis one layer down.
  • Gas E&Ps to LNG. The tension Big Digital Energy raised, that a domestic shortage could void export primacy, got sharpened on TFTC, where energy economist Anas Alhajji warned that within two to three years "US LNG is going to fight, literally fight, with the data centers over gas," pointing to a $15B AI center south of Dallas building its own gas plant. For gas-weighted E&Ps that's demand tension = pricing power; for LNG exporters it's the political risk the bears are pricing.
  • Uranium to the fuel cycle. Lundin's "set it and forget it" framing on uranium, plus the General Matter enrichment build at Paducah Wright described, points at both physical uranium (Cameco, CCJ) and the enrichment/conversion chokepoint, the part of the nuclear renaissance with the least new capacity and the longest lead time.
  • Hyperscalers fund the whole thing. The through-line under every deal: Meta's 7.53% bond, the $700B/yr spend, BlackRock's $12B Meta financing and $40B Aligned acquisition (per TreppWire). The buyers' balance sheets are the fuel; when their financing gets pricier, every link upstream feels it.