Newsletter · 2026-09-08 · Ashutosh Agarwal

Big Power Writes the Checks While Copper Becomes the AI Bottleneck - Powering AI: Grid, Gas, Generation & Nuclear - Week of September 8, 2026

Powering-AI podcast roundup for the week of September 8, 2026. Almost nobody reported earnings, yet NextEra, ONEOK and Cheniere committed tens of billions of dollars, gas-for-power demand math kept climbing, and copper emerged as the AI trade's most physical bottleneck.

Powering AI: Grid, Gas, Generation & Nuclear

Week of September 8, 2026: Big Power Writes the Checks While Copper Becomes the AI Bottleneck

A quiet earnings week is usually a quiet week. This one wasn't. Nobody had to report numbers to tell you what they believe: a handful of the biggest names in American power simply reached for their checkbooks. A $66.8 billion utility merger. A $13 billion gas plant. A $9 billion private-equity injection to fund a Permian deal. And running underneath it all, the same argument the sector has been having for two years: is this the early innings of a multi-year buildout, or the sound of everyone rushing to spend before the music stops?

Here's what the podcasts were actually saying.

TL;DR

• The money moved before the earnings did. NextEra, ONEOK and Cheniere all made large capital commitments in a week when almost no one reported, and the way they paid for them (stock, private equity, cash) tells you as much as the deals themselves.
• The gas-for-power number keeps getting bigger. Analysts now peg data centers as adding roughly 15 Bcf/d of new gas demand by 2030, and one tracker counts 189 gigawatts of gas plants tied to data centers, nearly double where it stood six months ago.
• Copper quietly became the AI trade's most physical bottleneck. Near record prices, an 18-year timeline for new mines, and the uncomfortable fact that China does most of the world's smelting. Nuclear, by contrast, was oddly silent this week.

What's new

The check-writing week. On Telltales, hosts Ava Cabot and Marcus Graham walked through four capital decisions made inside seven days, and the framing was sharp: "Nobody was reporting. Everybody was spending." NextEra shareholders approved a $66.8 billion all-stock merger with Dominion that would create the largest regulated electric utility in the country, and in the same week, NextEra's competitive arm committed $13 billion to a natural gas complex in Fayette County, Pennsylvania. The tell, per Telltales: NextEra paid for the merger with stock, not cash, on a balance sheet already carrying $82 billion of net debt and negative free cash flow mid-build. "Stock is the currency you use when your own cash flow can't fund the plan, and you'd rather not add to the debt load."

Someone else is buying a piece of ONEOK so ONEOK can buy Permian gas. The RBN Energy Blogcast detailed ONEOK's $4.425 billion purchase of Brazos Midstream's Midland Basin gathering-and-processing assets, a deal that more than doubles ONEOK's Midland processing capacity to about 2.3 Bcf/d and lifts its total Permian processing toward 3.5 Bcf/d. The eye-catching part is the financing: a $9 billion non-voting minority equity investment from Apollo. As Telltales put it, "That is not a loan. Somebody else is buying a piece of the company so the company can buy the assets", a rational move for a name already levered around 10x net-debt-to-free-cash-flow, but a dilutive one.

Cheniere just finished the thing and raised the number. The cleanest capital story of the week and the one that got the least attention: Cheniere reached substantial completion on Corpus Christi Stage 3, shipped its 5,000th Gulf Coast cargo, and took 2026 EBITDA guidance up to $7.9–$8.4 billion from a prior $7.25–$7.75 billion, with Equinor taking a first cargo under long-term contracts (Telltales). No merger, no PE partner, no arbitration: they built it and raised the guide.

The gas-demand math keeps climbing. On The HC Commodities Podcast, gas analyst Emily Kyne laid out the cleanest framework of the week. US gas-fired power burn averages about 36 Bcf/d today (running 50–51 Bcf/d in the current heat) and she expects the annual average to climb roughly 20 Bcf/d by 2030, about 6 Bcf/d from coal retirements and conversions (some 60 GW of coal on the schedule) and about 15 Bcf/d from data-center large loads. Geographically it clusters where the gas already is: a third in the Northeast and mid-Atlantic (Northern Virginia, the Marcellus), a third in Texas's ERCOT grid, the rest scattered across Arizona, Georgia, Ohio and Chicago. Her real worry isn't generation, it's the wires: "There are a lot of infrastructure constraints in the supply chain for the grid," with PJM and MISO looking tight over the next couple of years.

GE Vernova's own people are pushing back on the bottleneck story. This is the most useful operator commentary of the week. On Energy Gang, a GE Vernova executive acknowledged the scary lead times everyone quotes (roughly five years for a large-frame gas turbine, two-to-five years for transformers) but argued the equipment isn't the true constraint. The company has lifted US manufacturing capacity "from 18 gigawatts... to announcing 30 gigawatts," and the memorable line was the challenge: "If you can show me a turbine stand without a turbine, I will find your turbine. And no one has been able to do that yet." In plain English: the holdup is permitting, grid connections and the surrounding infrastructure, not GEV's factory. For scale, in 2025 the company put 26 GW of new capacity on the grid and energized 68 GW of transformers.

The debate

This is a rare week where the bull case and the bear case were argued by the same people, sometimes in the same sentence.

Bull: the demand is real, the buyers are rate-insensitive, and the shortfall is structural. The single best summary came from UBS's Ulrike Hoffmann-Burchardi on Alpha Exchange. UBS estimates the US will be roughly 100 gigawatts short of the electricity it needs by 2030 once you add AI to electrification and industrial demand. And crucially, the hyperscalers barely flinch at cost: she argued it would take more than 50 basis points of rate increases to slow AI capex, "not 25, 50 basis points", because these buyers are "already paying three times the salary of an electrician to build the data center" and "twice the base rate for electricity." When your customer is that price-insensitive, the whole power stack that feeds them re-rates together.

"The rate elasticity of AI CapEx is very, very low. I think we need to see a lot more."

Bear: the numbers are slippery, the politics are turning, and copper is a hard physical limit. The Shift Key hosts made the skeptic's case best. Yes, one tracker (Global Energy Monitor) now counts 189 gigawatts of gas plants tied to data centers, nearly double the ~97 GW of six months earlier, against an existing US gas fleet of about 512 GW. But they cautioned that a lot of that pipeline may be double-counted: developers request capacity on the grid, get told how long the wait is, then pivot to an off-grid project, and both requests can end up in the same tally. "Everything felt like sand slipping through my fingers," one host said of trying to pin down the ten biggest projects. Not all 189 GW gets built.

And the same demand that excites the bulls is now a political liability. On Thoughts on the Market, Morgan Stanley's Ariana Salvatore called data-center pushback "far and away" the biggest debate heading into the midterms, and stressed it's a granular, local risk, not a Washington one. Watch specific governor races, she said, because governors appoint the public utility commissioners who actually set rates; the opposition cuts across both parties. Morgan Stanley is still constructive (its internet team models over $1 trillion of hyperscaler capex next year), but the companies are clearly playing defense, signing ratepayer-protection pledges and, in Google's case, proposing a water-usage framework.

Where the bear case has the sharpest teeth is copper, more on that below.

One honest note: nuclear and small modular reactors, usually the loudest corner of this sector, were nearly silent this week. No fresh operator commentary on Vistra, Constellation or Talen; nothing new on plant restarts or SMR design-certification milestones; no uranium price call worth citing. That's not a bearish signal, just an absence, and worth watching, because it's exactly the corner that has driven the sentiment swings all year.

The names in play

GE Vernova (GEV) got the most airtime. Beyond the operator commentary above, Schwab Network's Jessica Inskip laid out the bull case on the stock: a clean balance sheet (total debt just 3.53% of assets), free cash flow up 7% quarter-on-quarter, positive earnings revisions, and Street price targets "well over $1,200," with earnings growth projected to accelerate from about 15% to 25% as the backlog converts. The catch is the chart: the stock had already pulled back roughly 25% from its early-July high toward its 200-day average near $880, which the technicians framed as a possible support level rather than a warning. A high-quality compounder that ran too hot and is now digesting.

Eaton (ETN) and Schneider Electric got a nod on Motley Fool Hidden Gems Investing as the "doesn't matter which chip wins" way to play the buildout: makers of high-voltage cables, transformers and switchgear with "favorable multi-year backlogs." The same episode dug into Bloom Energy, freshly added to the S&P 500, whose pitch is speed: on-site fuel cells deployed "in 90 days or less" versus multi-year grid interconnects, sitting on a $20 billion backlog ($6 billion product, $14 billion service). The obvious caveat is valuation: roughly 80x forward earnings after a ~500% run, and the analysts were candid that if the AI buildout slows, so does the growth story.

NextEra (NEE), ONEOK (OKE) and Cheniere (LNG) are covered above, the three names that actually moved capital this week.

Read-throughs

The most important cross-current this week wasn't power at all. It was copper, and it showed up on four separate podcasts, which is usually the sign of a theme that's about to matter.

Start with the setup, from Brew Markets: a new copper mine takes on average 18 years to go from discovery to production, ore quality is declining, Chile (the world's largest producer) just saw output fall on weather and maintenance, and the US is weighing tariffs on refined copper from 2027, which has traders front-running by moving metal into the country now. Amazon has already gotten creative, inking a two-year copper deal with a Rio Tinto mine in Arizona. Data centers themselves are only 1–5% of global copper demand, but the transmission and grid buildout around them is enormous.

On The KE Report, John Rubino put it bluntly: copper "just doesn't care anymore," sitting near all-time highs and approaching $7 a pound while big producers underdeliver. Well-run miners are throwing off cash and cleaning up their balance sheets, though he was careful to note copper is still cyclical, and a genuine recession that stalls the AI buildout would hurt.

The supply-side detail came from two insiders. On Dig Deep, Solaris Resources' Matthew Rowlinson said the world needs roughly an extra million tons of copper a year through 2035 and he doesn't see it coming near-term, while underscoring the real chokepoint: China smelts 50–60% of the world's copper (with only ~10% of the mining), and makes ~70% of copper foils and 50–60% of the rod and wiring. Owning a mine in Latin America doesn't fix Western supply security if the concentrate still has to go to China to be processed. The Columbia Energy Exchange episode with DOE's Audrey Robertson drove the same point home from Washington: the US has plenty of copper ore but only three copper smelters, while China built 42 in the last decade. "It doesn't matter if we have copper in the ground here if we can't turn it into a usable resource."

The read-through: this is bullish for Freeport-McMoRan (FCX) and the well-capitalized copper miners on price, but the more durable insight is that the AI bottleneck is migrating. Two years ago the question was whether you could get enough chips; now it's whether you can pour enough concrete, wind enough copper and pull enough permits. That broadens the trade well beyond semiconductors, into miners, electrical-equipment makers and utilities alike.

On the gas side, the read-throughs run straight through the Marcellus. On NGI's Hub & Flow, Marcellus Shale Coalition president Jim Welty made the operator's case for behind-the-meter power: Pennsylvania went from 5% to 60% of its electricity from natural gas since shale took off, saving consumers about $9 billion versus 2008 prices and becoming the nation's top electricity exporter. He argued that same gas can power data centers behind the meter without raising consumer bills, while voicing real frustration at the PA Governor's recent "180-degree" executive order tilting toward renewables. And in Canada, ARC Energy Ideas flagged hopes for final investment decisions on LNG Canada Phase 2 and Stellarwind that could reach ~6 Bcf/d of exports by the early 2030s, roughly tripling current Canadian capacity of ~1.8 Bcf/d, with data-center gas demand now an explicit input to Alberta's strategy.

What changed

The shift this week is in tone, not thesis. For months the story has been about demand forecasts and lead-time anxiety. This week it turned into signed deals and committed dollars (NextEra, ONEOK, Cheniere) and into a growing recognition that the binding constraints are moving downstream, from turbines and transformers (where GE Vernova insists it can keep up) toward permitting, transmission wires, and processed copper. Keep one eye on the midterm races Morgan Stanley flagged; if data-center opposition starts winning governorships, the ratepayer politics could do what high interest rates so far have not.