Newsletter · · Ashutosh Agarwal

Vistra Says the AI Power Panic Is 10x Overstated - Powering AI: Grid, Gas, Generation & Nuclear - Week of September 18, 2026

The Powering AI: Grid, Gas, Generation & Nuclear brief for the week of September 18, 2026, synthesizing this week's energy and infrastructure investor podcasts. This issue leads with Vistra's Stacey Doré arguing AI power-demand forecasts are overstated by 10x or more and that the fix is connecting load, not building plants, then weighs the supply-gap case from Bank of America, Microsoft and Brad Gerstner, the off-grid gas land grab powering Meta's Prometheus with Williams, the split between nuclear operators and skeptical traders on Oklo, and uranium's quiet admission that its real bid is geopolitical, not AI.

Powering AI: Grid, Gas, Generation & Nuclear

Week of September 18, 2026: Vistra Says the AI Power Panic Is 10x Overstated

For months the story writing itself on the podcasts has been one long crescendo: AI needs power, power is scarce, buy everything that makes an electron. This week the crescendo finally drew a real heckler, and it came from the last place a power bull would want one. The owner of some of the most valuable power plants in America went on a show and said the panic is overdone by a factor of ten. That's the argument worth chewing on today, because if she's even half right, a lot of the trade changes shape.

TL;DR

  • Vistra just told the power bulls to calm down. A senior executive at one of the biggest merchant generators in the country said the headline demand numbers are "overstated by 10x or more," that the existing grid can already handle 99% of the hours, and that the real bottleneck is getting fake projects out of the interconnection queue, not building a wall of new plants.
  • But the supply-gap camp came loaded too. Bank of America pegs new build at ~93 GW over five years against 230 GW of data-center applications, Microsoft says it's turning away business for lack of power, and Brad Gerstner thinks even 2026's plans slip from 43 GW to 25 GW because the equipment and permits aren't there.
  • The uranium bulls quietly divorced themselves from the AI story. With spot at $90 and enrichment at record highs, the sharpest voices in the fuel cycle are now saying out loud that data centers have "very little impact" on the uranium price: the real bid is China, Russia and Japan hoarding pounds.

The one that matters: an operator says the emperor has fewer clothes

Most of the loudest AI-power commentary comes from people who are selling something, a reactor, a turbine, a research subscription. So it's worth stopping when the pushback comes from someone whose stock literally goes up if the mania is real.

On Energy Gang, Vistra's Stacey Doré walked through the demand math the way a skeptic reads a prospectus. Yes, if you add up every utility's interconnection queue you get more than 700 gigawatts of new load by 2030. Her response was blunt:

"That is not going to happen. There are not enough chips in the world to supply data centers to build that amount."

Vistra's own numbers, which she said the company has held for two years: about 17 to 18 GW of real load growth in PJM (the mid-Atlantic grid) by 2030, 2 to 3 percent a year, and 20 to 30 GW in ERCOT (Texas), of which only 10 to 15 GW is actually data centers. Put next to the headlines, she argued, "the forecast people are seeing every day are overstated by 10x or more."

Her tell for how much of the queue is vapor: when AEP tightened its data-center rules to require real site control and security deposits, its interconnection queue collapsed from 38 GW to 13 GW. Most of the rest, in her telling, is real-estate speculators reserving the same phantom project across three markets at once.

And here's the part that should make a utility-capex bull sit up: she doesn't think we need to build one-for-one to meet even the real load. Vistra's gas plants run at a 50-60% capacity factor; the grid as a whole sits below 50% utilization for 99% of the year. The problem isn't average power, it's the handful of super-peak hours, and those can be shaved with demand response and the backup generators data centers already wheel in. Her one-line prescription wasn't "build," it was "get load connected as fast as possible" so the hyperscalers start sharing the fixed costs of the grid instead of leaving them on residential ratepayers.

She's not bearish on Vistra, to be clear, she reminded everyone the stock is up more than 600% over five years, even after flatlining lately. But this is the first time this year a big generator has publicly argued the shortage narrative is more marketing than physics. That deserves a place at the top of the file.

The other side showed up too, and it's not a strawman

The nice thing about this week is that the bull case didn't stay silent. It just came from different rooms.

The supply gap is real, and quantified. On Clean Power Hour, host Tim Montague cited a Bank of America estimate that the US builds about 93 GW of new power plants over the next five years against 230 GW of data-center interconnection applications, 2.5 times more demand than expected supply. Even haircut for cancellations, that's a gap, not a glut.

The people who buy the power say power is the ceiling. On Bloomberg Tech, Microsoft's plan to triple data-center capacity to 38 GW from about 12 GW today came with the detail that matters: Microsoft is turning away AI and cloud business because it can't get power, and Tencent defected to Oracle when Microsoft couldn't supply. When Cleaning Up's guests noted that Google's AI token usage grew 330-fold in a year, pushing Google's monthly AI power from 16 megawatt-hours to 5.3 terawatt-hours, and that's only a fifth of global AI compute, the scale of the demand curve stops feeling like hype.

And the smart-money bull actually agrees with Vistra on the near term, for the opposite reason. On All-In, Altimeter's Brad Gerstner said 2026 has ~43 GW of AI compute additions planned (roughly all the compute the US has today) but that only about 25 GW will actually get built, throttled by permitting, interconnection and equipment shortages. Vistra says the demand is overstated; Gerstner says the demand is real but the build is capped by the same physical bottlenecks. Both roads lead to the same place for an investor: whoever sells the scarce equipment and the fast electrons wins, and whoever's valuation assumes the full 43 GW shows up is going to be disappointed.

The honest read of the week: this isn't bull-versus-bear on whether AI needs power. It's a fight over how much, how fast, and who pays, and for the first time the "less than you think, and connect it don't build it" case has a credible operator behind it.

Off-grid is the release valve, and it's a gas story with a name attached

While the grid argues with itself, the hyperscalers stopped waiting. The best single explainer this week was Michael Thomas on Inevitable, walking through what he calls the "Mad Max phase" of power development.

The template is xAI's Colossus: told by utilities it would take years to connect, Musk's team wheeled in a couple dozen gas turbines (the kind meant for oil-and-gas fields) on the backs of semi trucks and had a gigawatt-scale facility running in months. (Thomas's useful reminder: a gigawatt is a nuclear plant's worth of power; New York City's entire peak load is ten of them.) Since then, he's tracking more than 100 GW of off-grid, behind-the-meter data centers announced, roughly 2 GW actually operating, and 5-10 GW under construction.

Here's the part with a ticker on it. Meta's Prometheus campus in New Albany, Ohio (a cluster of more than a dozen rapid-deploy "tent" buildings that cut construction from 24 months to 12) is being powered by islanded natural gas built with Williams (WMB): 200 MW live, another 200 MW coming, on the way to more than a gigawatt. That's not a slide; that's a signed, physical, gas-fired deal feeding one of the largest AI clusters in the country. He echoed the same point on Growth Stories With IBD: Amazon and Meta are signing 10-year gas contracts as bridge power, because a five-year grid wait in Texas is a five-year revenue hole.

And the revenue math is why they'll pay almost anything for speed. Thomas said data-center economics have gone from about $10 billion of revenue per gigawatt to $30-50 billion per gigawatt in the recent SpaceX deals with Google and Anthropic. When each year online is worth tens of billions, a truck full of gas turbines is a rounding error. The catch he flagged: the politics are turning, with a genuinely bipartisan backlash (Bernie Sanders and Ron DeSantis on the same side) and water, not carbon, as voters' top complaint.

Nuclear: the operators are building, the traders are skeptical, and the two aren't talking

This was a heavy week for atoms, and the interesting thing is how split the room was between the people building reactors and the people pricing the stocks.

The builder's view. Oklo's CFO Craig Bellmer, on Energy Evolution, described demand from hyperscalers like Meta as "unconstrained" and laid out real milestones, not vibes: Oklo took its first pilot reactor critical in August, targets its 75-MW Idaho Aurora powerhouse operational in 2028, and is advancing a 1.2 GW deployment in Ohio with Meta as anchor customer for the early 2030s. He's raised $2.5 billion since the IPO (about 30% of all the money that's gone into small modular reactor (SMR) developers) and signed Siemens Energy as a preferred supplier for the steam-turbine side. Aalo Atomics' CEO made the parallel pitch on Modern CTO: Microsoft, Google and Amazon can't wait 10-15 years for a big reactor, so bolt on 10-MW SMR units next to the data center and skip the grid entirely.

The trader's view. On Motley Fool Hidden Gems, the reminder that Oklo is down about 88% from its 2025 peak near $200 to roughly $40: the business hasn't changed, the hype has. The show's advice was to prefer "boring" established names over pure-play SMR lottery tickets. Both things are true at once: the operators are making genuine progress, and the equities front-ran it by years. That gap is the trade.

Uranium's quiet confession: this isn't really an AI story

If you own uranium because of AI, the uranium bulls would like a word. On Going Nuclear, Justin Huhn laid out a fuel cycle at or near record pricing: spot at $90, term at $97, enrichment (SWU) at an all-time high, conversion just shy of one, and called Bank of America's $130 target for late next year "reasonable," with the industry treating $150 as a "foregone conclusion" at the recent World Nuclear Association conference.

But then came the honest part. The AI/data-center narrative, he and his co-host agreed, has had "very little impact on the price of uranium": it moves the equities up and down on sentiment, but the actual pounds are being bid by something else entirely: China (sitting on ~700 million pounds and signing $20 billion-plus contracts), Russia (buying uranium for the first time ever, on top of controlling 43% of world enrichment), and Japan restarting. Kazatomprom is openly telling Western utilities to "step up or we're not going to have anything left for you." As the co-host put it, after two years of AI-nuclear headlines, "I don't see anything tangible yet to hang our hat on," and he thinks divorcing uranium from the AI story is actually healthier for the thesis.

Rick Rule made the same point from a different chair on The Competent Investor, waving off data-center demand as "six years out, seven years out" and pointing to Japanese restarts and global refueling as the near-term drivers. The takeaway for a book: the uranium trade may be excellent, but if you're underwriting it on AI, you're right for the wrong reason, and you'll be caught off guard when the fuel-cycle catalysts turn out to be geopolitical, not silicon.

The read-throughs

  • Gas midstream is the clearest winner of the off-grid pivot. Williams (WMB) is literally the named partner powering Meta's Prometheus. ARC Energy Ideas is bullish North American gas on data-center plus LNG demand, and NGI's Hub & Flow put a number on the pipe: up to 25,000 miles of new gas lines needed by 2050. Neither named Kinder Morgan (KMI) or ONEOK (OKE), but the demand pull they describe reads straight through to anyone moving molecules to a turbine.
  • Copper and conductors are the tax on all of it. On Odd Lots, Robert Friedland made his "monumental shortage" case tied to data centers and electrification; DHUnplugged pegged copper at a record $14,000+/ton and estimated 30,000-40,000 tons of copper per AI data center. That's the read-through to Freeport (FCX) and the cable and transformer supply chain, every off-grid megawatt still needs wire.
  • The efficiency lever is the sleeper bear case. If you don't need to build as much because you can use what's there better, the arms dealers sell less. Volts profiled a design (Verus) hitting 77% utilization of provisioned utility capacity versus 49% for conventional builds; Interchange Recharged showed reconductoring can double a line's capacity in 18 months instead of a 5-7 year rebuild, a direct tailwind for Quanta (PWR) and the conductor names, and a quiet headwind to the "we must build 100 GW of new plants" story.
  • Demand response is becoming a real revenue line. The Dynamo Show noted DR payments have gone from ~$10/kW to over $100/kW in ERCOT and PJM as data centers proved they can defer training runs, which is exactly the "flexibility beats new steel" argument Vistra was making, now with a price tag.
  • Behind-the-meter generation has its own winners. Catalyst flagged Bloom Energy (BE) taking "massive orders from Oracle and others" for on-site fuel cells, another way to get online without waiting for the grid.

What changed this week

The debate got two-sided. For most of the year the podcasts sang from one hymnal: buy the shortage. This week, for the first time, a major generator (Vistra), a smart-money bull (Gerstner) and the uranium complex's own analysts all independently argued some version of "the number is smaller, slower, or driven by something other than what you think." Nobody's calling the supercycle over. But the easy part of the trade (where you could buy anything with a turbine and be right) is starting to require an actual opinion about which electron, how fast, and who pays. Into next week, watch for whether any regulated utility pushes back on Vistra's "don't rate-base it, connect it" framing, and whether a single gas-turbine or transformer OEM confirms Gerstner's equipment-shortage thesis with a hard lead-time number.