Newsletter · · Ashutosh Agarwal

Powering AI - Grid, Gas, Generation & Nuclear - Week of June 23, 2026: Hyperscalers Go Off-Grid, Turbines Sold Out to 2031

Powering-AI newsletter for the week of June 23, 2026. The bottleneck has moved from whether AI needs power to who can physically deliver a turbine or a megawatt-hour, with hyperscalers going off-grid (a 2.67 GW Microsoft-Chevron West Texas gas plant the new template) and gas turbines now sold out to 2031.

Powering AI: Grid, Gas, Generation & Nuclear

Week of June 23, 2026: Hyperscalers Go Off-Grid, Turbines Sold Out to 2031

Powering AI: Grid, Gas, Generation & Nuclear, Tuesday, June 23, 2026

There's a number that reframes this whole trade, and a fund manager just back from the industry's annual gathering said it out loud: order a gas turbine today and the manufacturer will hand it to you "sometime in 2031." That's not a supply chain. That's a queue with a velvet rope. And this week the tape was wall-to-wall with people explaining what happens when the most powerful companies on earth show up at that rope all at once.

The short version: the demand is no longer the debate. The plumbing is.

TL;DR

  • The bottleneck has moved from "will AI need power" to "who can physically deliver a turbine, a transformer, or a megawatt-hour", and the people closest to the equipment are the most bulled-up.
  • Hyperscalers have stopped waiting on the grid. A 2.67 GW Microsoft-Chevron off-grid gas plant in West Texas is the new template, and roughly a quarter of data-center capacity in development now plans to bring its own power.
  • The bear case isn't "no demand." It's an interconnection bubble, peak-priced equipment, and a five-year visibility cliff, and several smart guests said the back half of the decade is a coin flip.

What's new

The turbine line is the whole story. On Going Nuclear (Justin Huhn and Trevor Hall, back from the World Nuclear Fuel Markets Conference), a GE Vernova representative said the quiet part plainly: turbines are "sold out for four years out into the future," and once supply chains were finally in place, "the line was around the block." That's an operator, not a bull. It dovetails with Tema's Chris Semenuk on Other People's Money (also simulcast on Monetary Matters): GE Vernova sitting on a ~$90B equipment backlog plus an ~$80B service tail, Eaton's data-center equipment sales up 240% in a quarter, and Caterpillar's engines now running data centers rather than just backing them up.

Hyperscalers are building their own power plants. The Motley Fool Hidden Gems crew broke down the morning's Microsoft-Chevron "Project Kilby": a 2.67 GW behind-the-meter gas facility in Pecos, Texas, roughly thirty times a standard data center, on a 20-year supply deal, with GE Vernova turbines and Caterpillar's Solar Turbines doing the work. Matt Frankel's line: off-grid gas is "literally the only viable way" to get capacity on a workable timeline against a four-to-seven-year interconnection wait. Meta's doing it in Louisiana, Alphabet and Amazon are permitting their own campuses, and roughly a quarter of capacity-in-development is now off-grid. Why it matters: incremental gas-fired demand that never touches a utility queue, and a direct order flow into turbine and genset OEMs.

The equipment crunch is now showing up in price. S&P Global's analysts on The Decisive Podcast said transformer and switchgear lead times have doubled, high-voltage gear is north of a year, and power-equipment prices are up ~5% year-over-year even as the rest of non-residential construction sags. Their framing is the one to steal: hyperscalers are buying "the way that we have historically thought about... defense spending", not worried about cost, buying everything they can. That's pricing power for whoever makes the box.

Uranium's term market keeps grinding higher while nobody's watching. On Going Nuclear, spot is parked in the mid-$80s, but the reported long-term price is $94-95/lb, three- and five-year forwards both sit above $100, and some base-escalated contract portions are already breaching triple digits, with US production incentive pricing pegged north of $130/lb. Urenco's CCO told World Nuclear News the enricher is adding 4.6 million SWU and has "completely reversed" a post-Fukushima plan to shrink, just as the US-Russia suspension agreement pulls the largest enrichment supplier offline on January 1, 2028.

The debate

This is the rare week where both sides actually showed up.

The bull is a synchronized, multi-year re-rating: grid, utilities, IPPs, gas and uranium pulled up together by a demand shock with no modern precedent. Wood Mackenzie's team on Interchange Recharged put the scale in historical terms, ~435 MW of new data-center capacity announced every month, the fastest surge since the WWII mobilization of 1939. Equipment makers have pricing power, utilities are signing 15-year large-load contracts, and the fuel chain behind the reactors is structurally short.

The bear isn't a demand denial, it's a timing-and-froth argument. Wood Mac's own Tom Falcone (LPPC) noted that when utilities demand real commitments, about two-thirds of interconnection requests simply vanish; PJM shows ~78 GW of committed load against just ~26 GW of accredited generation. UT Austin's Joshua Rhodes was blunter on Renewable Rides: the ERCOT queue has ballooned to 435 GW (90% data centers) against an all-time grid peak of 85.5 GW, "I called it a bubble... it's just gotten more bubbly", and warned the rate-base names will "still be paying off the mortgage" on peak-priced gear if demand cools. On Catalyst, Energy Impact Partners' Andy Lubershane was explicit, high conviction on the next 3-5 years, notably low beyond that, which is precisely the window where today's 2031-delivery orders land.

The pull-quote of the week, from a fund manager fresh off a week with utility CEOs: behind-the-meter was "the cool kid on the block," but after talking to the hyperscalers, "I think behind the meter's nonsense", grid plus backup is still cheapest. File that against the Project Kilby template and you have the year's central argument in two sentences.

The names in play

GE Vernova is the cleanest expression of the thesis and the one guests kept circling, backlog visibility into 2029-2031, FCF inflecting hard, though the execution risk (does the backlog convert on schedule?) is real. Semenuk's relative-value call on Monetary Matters is worth noting: he'd rather own NextEra, 8-9% EPS growth plus a ~4% yield to 2035, buying PJM's "data center alley" via the pending Dominion deal, at ~20x than Constellation at ~35x, and he'd skip the pre-revenue SMR "framework" names. He's also fishing the SMID tier for picks-and-shovels, Powell and Bel Fuse, legacy suppliers suddenly selling into utilities and data centers. Eaton and Caterpillar are the other two the operators kept naming.

Read-throughs

Pull one link and the rest move. Gensets (Caterpillar's Solar Turbines, and the on-site engine demand Cummins should ride) are now primary power, not backup. Copper is the sleeper: on The David Lin Report, Copper Giant's Ian Harris flagged AI's share of US power going from 4.4% to ~12% by 2028, a ~400kt deficit against ~15 days of global stockpile, and a 2.5-year wait for large mining transformers, copper now trading more in sympathy with NVIDIA than gold (read it through to Freeport). Enrichers and physical-uranium vehicles (Cameco, the Sprott trust) inherit the term-price grind and the 2028 Russian cutoff. And the hyperscalers themselves, Microsoft, Meta, Alphabet, Amazon, are now the balance sheets underwriting the gas plants, PPAs and restarts (Crane/TMI for Microsoft, Duane Arnold for Google), which is what makes the order flow durable rather than speculative.

One quiet spot worth flagging: nobody this week put an operator's voice on Vistra, Talen, Oklo or NuScale, and there was no fresh read on PJM capacity-auction clears or merchant forward spreads. If the IPPs are your book, the tape didn't feed you this week, watch for it next.

This note draws on operator and expert podcast commentary from the trailing week; it is research, not investment advice. Verify company-specific figures against primary filings before acting.