Newsletter · · Ashutosh Agarwal

Power Equipment Is Sold Out to 2028 as the AI Squeeze Tightens - Powering AI, Grid, Gas, Generation & Nuclear - Week of September 4, 2026

Powering AI for the week of September 4, 2026: transformers, turbines and switchgear are booked out to 2028 and the squeeze now has an early-2030s end date, gas is winning the near-term fight to electrify the machines, and the only real bear case on the tape is politics and power bills, not demand.

Powering AI: Grid, Gas, Generation & Nuclear

Week of September 4, 2026: Power Equipment Is Sold Out to 2028 as the AI Squeeze Tightens

If there was one message running through this week's podcasts, it's that the thing holding back the AI boom isn't chips anymore, it's the boring gray boxes that move electricity around. Transformers. Turbines. Switchgear. The stuff nobody thinks about until you can't get it. And you can't get it: the people building this infrastructure now openly say the key gear is sold out for years, and the squeeze doesn't ease until the early 2030s.

That's a longer runway than most investors have penciled in. It's also the whole trade.

TL;DR

  • The shortage got a hard timeline. Multiple guests now say power equipment is booked "to 2028" and the market stays "very, very tight" into the early 2030s, this is a multi-year condition, not a 2026 blip.
  • Gas quietly won the near-term power fight. A four-year Chevron / GE Vernova / Microsoft project to turn Permian gas into electricity for a data center, plus Appalachian operators tracking up to 20 billion cubic feet a day of new gas demand, shows how hyperscalers are actually getting powered right now.
  • The bear case showed up, but it's about politics and rates, not demand. The data-center backlash is real and now a midterm-election issue, and the buildout is pushing up ordinary people's power bills. Nobody serious argued demand is rolling over.

What's new this week

1. The people selling the gear admit it's sold out, and they're happy about it. On Energy Gang (Sept 1), Roger Martella, GE Vernova's chief corporate and sustainability officer, said that if you wrote the story of GE Vernova in 2026, "the first chapter admittedly would be AI." The company makes more than 25% of the world's electricity, employs 85,000 people, and its stock is up roughly 600% since it was spun out of General Electric in March 2024. It put 26 gigawatts of new capacity on the grid globally in 2025, and it's committing $1.3 billion and 1,800 US manufacturing jobs to build more switchgear, circuit breakers and transformers, the components everyone is short of. Martella pushed back on the idea that his company is the bottleneck, relaying a favorite line from CEO Scott Strazik: "if you can show me a turbine stand without a turbine, I will find your turbine." In plain terms: we're building as fast as anyone can, and we're not the reason your data center is late. Microsoft's energy partner Melissa Lott joined the same conversation (speaking personally, not for Microsoft), a rare case of the buyer and one of its biggest suppliers talking shop in public.

2. A serious energy analyst put a hard end-date on the squeeze. On Catalyst with Shayle Kann (Sept 3), Kann and his guest, a developer who sources power for these projects, laid out why this stays tight. "This problem is going to persist for the foreseeable future," the guest said. "Everyone who thought, oh, I just got a constraint in 2027 or 2028, you have the same constraint in '29 and '30 and possibly into the early 2030s." That's the single most useful sentence for a book this week: the constraint isn't a spike that clears, it's a plateau that holds. Why? Because the industry spent a decade with flat electricity demand and let its supply chains for transformers and turbines wither, and you can't rebuild them overnight.

3. Gas is how the data centers actually get powered, and there's a marquee deal to prove it. On Squawk on the Street (Sept 3), Jim Cramer walked through a real project: Chevron (he name-checked CEO Mike Wirth) "is in a project, a gigantic project in the Permian... working with GE Vernova... a four-year project to be able to get Permian gas into electricity for a giant Microsoft plant." Note the timeline, four years. These aren't things that inflect next quarter. Cramer's blunt read on the political noise around data centers: "It will have no impact whatsoever. The demand for these is going to go on steadily." (That's pundit bravado, not gospel, but it's the consensus mood.)

4. Appalachia is sizing the gas prize. On NGI's Hub & Flow (Sept 3), the guests noted Pennsylvania already makes about 60% of its electricity from natural gas and has saved consumers roughly $9 billion versus 2008 prices. The eye-opener: EQT is tracking 45 potential data-center projects that could add up to 20 billion cubic feet a day of new gas demand by the end of the decade, though operators realistically guide to something closer to 10 Bcf/d. For scale, that "realistic" number alone is a meaningful chunk of total US gas demand, and it's why gas-weighted producers and the pipelines that move their product keep coming up.

5. The grid is already straining before most of this gets built. On Factor This (Aug 28), the cleantech roundup flagged that PJM, the grid operator for 13 states across the mid-Atlantic and Midwest, is facing a capacity shortfall of upwards of 6 gigawatts, pushing prices up across the region as data-center growth in Virginia outruns new supply. They described an incident where about 4 gigawatts dropped offline in seconds. When the grid is this tight, every megawatt of firm, dispatchable power (gas, nuclear) gets more valuable, and every delay in new equipment costs real money.

The debate

This is usually where we steel-man the bull and the bear. Here's the honest truth about this week: the bull case was everywhere, and the bear case that showed up was about politics and power bills, not about demand disappearing.

The bull case is a durable, multi-year supercycle where grid, utilities, gas, and equipment makers re-rate together, and it got fresh support from serious people. The a16z Show's infrastructure episode (Aug 28) put it bluntly: "we've got shortages of transformers and turbines and everything that goes into that," and "these lead times are not that easy to compress" while demand grows "10x a year." Their line on supply: components are "basically all booked out to 2028," with GPUs reselling for four times their price and even reinforced concrete now one of the fastest-rising costs. A striking labor detail: only 2% of US electricians are certified to work on the high-voltage DC power these dense server racks now require, Meta is running its own training program to fix it. The episode also cited a gap of roughly 44 gigawatts of new data-center power needed by 2028 against only about 25 gigawatts of expected grid additions. The Street sees it too: on Alpha Exchange (Sept 1), UBS Wealth Management's Ulrike Hoffmann-Burchardi, CIO for the Americas and head of global equities, put the power shortfall at roughly 100 gigawatts by 2030 and named transformers and grid components the key supply-side risk to AI capex, though she frames the bottleneck as temporary (human-paced supply catching up to exponential demand). When the people writing the checks and the people allocating client money describe a shortage this broad, "supercycle" stops sounding like a slogan.

The bear case, such as it was, came in three flavors, none of them "demand is fake":

  • Politics. On Thoughts on the Market (Sept 2), Morgan Stanley reaffirmed a forecast of more than $1 trillion in hyperscaler spending next year and stayed constructive, but singled out midterm-election risk around data-center regulation as the fall catalyst to watch. Shayle Kann was more pointed: public sentiment on data centers has fallen "off a cliff" since last September, and it's now a live campaign issue that both parties are weaponizing.
  • The rate backlash. Kann's most nuanced point: a data center tends to lower the power bill of its immediate neighbors (the utility structures the deal that way), yet the national buildout is inflationary because it created shortages of transformers and gas turbines and drove up gas demand, pushing up every input to everyone's electricity price. Both things are true at once, which is exactly why the politics are so combustible.
  • The "bring your own power" trap. Texas and Pennsylvania are pushing rules that make data centers supply their own generation. Kann's guest was skeptical this helps, and shared a telling figure from a GE Vernova contact: to serve a 1-gigawatt data center off-grid, you end up building about 2.6 gigawatts of capacity across generation and batteries. Overbuilding like that just adds more demand for the scarce equipment, making the shortage worse, not better.

What was missing this week is notable: no fresh operator commentary on uranium spot or enrichment (SWU) pricing, no genset (Cummins, Caterpillar) discussion, and no new hyperscaler nuclear PPAs or plant-restart news in the pods. The nuclear thread went quiet, worth watching whether that's a lull or a signal.

The names in play

Guests actually moved theses on a handful of names this week:

  • GE Vernova (GEV) was the star, and unusually, from its own executive. The combination of a booming order book, a 600% post-spin run, and management's own framing of AI as chapter one is the clearest single-stock signal in the pods, with the caveat that a company officer is, of course, talking his own book.
  • Microsoft (MSFT) and Chevron (CVX) showed up as the demand-and-fuel side of the same Permian gas-to-power deal, a template for how hyperscalers get firm power fast when the grid can't deliver.
  • Cheniere (LNG) got a specific mention on The HC Commodities Podcast (Sept 1): it runs under a tolling model that gives it some discretionary spot volume to optimize between domestic and export markets. The bigger frame, US LNG export capacity roughly doubling to about 40 Bcf/d by 2032 from about 20 today, is the demand pull that keeps competing with data centers for the same molecules.
  • Freeport-McMoRan (FCX) anchored a copper conversation on Brew Markets (Sept 2), which flagged the stock up on the day and noted Amazon signing a two-year copper supply deal with a Rio Tinto mine in Arizona, a sign hyperscalers are now securing raw materials, not just chips and power.

Read-throughs

The whole point of this note is that a move in one link ripples through the others. This week's chain:

  • Copper is the sleeper. Three separate pods pounded the table. 10-Minute Contrarian (Aug 30) called copper supply "terminal", ore grades halved since 2000, three major mines offline (a mudslide, an explosion, a shutdown), while US data-center electrification is "only 20-30% of the way" there. On The KE Report (Sept 3), John Rubino said we'll "need more copper in the next 30 years than we mined for all of human history," with copper near $7/lb and miners throwing off record cash. The counterweight: on The David Lin Report (Aug 31), Gianni Kovacevic reminded listeners that data centers are only a small slice of total copper demand (about 75% of fabricated copper already goes into electrical uses growing about 3% to 4% a year), so this is a broad electrification story, not a pure AI bet. Both can be right: copper is tight, but don't underwrite it on data centers alone.
  • Gas producers and the pipelines that serve them read straight through from the Permian and Appalachia demand numbers above. If even the "realistic" ~10 Bcf/d of new data-center gas demand shows up, the molecules and the takeaway capacity both get bid.
  • Equipment makers beyond GE Vernova, the transformer, switchgear and turbine names, inherit the same booked-to-2028 backlog. The bottleneck is the thesis.
  • Hyperscalers (META, MSFT, AMZN) are increasingly the ones funding and de-risking the power and materials supply chain directly, the Meta electrician-training program, the Amazon and Rio Tinto copper deal, the Microsoft Permian project. The buyers are reaching upstream because they have to.

The bottom line

The most valuable thing the pods delivered this week wasn't a hot new name, it was a timeline. The power-equipment squeeze that's been the backbone of this trade now has serious people putting an early-2030s end-date on it, gas is winning the near-term fight to actually electrify the machines, and copper is quietly the tightest link nobody's watching closely enough. The risk that's growing isn't demand, it's the politics of who pays for it. Watch the midterms, watch the rate cases, and watch whether the nuclear thread wakes back up.