Newsletter · · Ashutosh Agarwal

AI Data Centers Go Off Grid as Power Supply Falls Behind - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 14, 2026

Power and infrastructure newsletter for the week of July 14, 2026. AI data-center builders have largely stopped waiting for the grid and are bringing their own generation behind the meter, driving sold-out turbine backlogs at GE Vernova, new domestic transformer capacity at Quanta, and a record-high uranium term price, even as grid veterans argue much of the spend is overdue maintenance and politics is starting to slow projects.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 14, 2026: AI Data Centers Go Off Grid as Power Supply Falls Behind

The single loudest message from this week's podcasts is a quiet one: the people actually building AI data centers have largely stopped waiting for the grid. When a utility tells you it can give you a gigawatt in 2032 instead of 2027, and your entire business runs on electricity, you go find your own power. That decision, repeated across the country, is rippling straight into the order books of turbine makers, transformer plants, gas pipelines and uranium miners. Here's what people who work in this world were saying over the past week.

TL;DR

  • Supply isn't close to keeping up. Data-center demand is growing something like 50% a year while the grid adds only single-digit gigawatts of new firm power annually. The fix, increasingly, is "bring your own generation" behind the meter: gas turbines, engines, fuel cells.
  • The equipment makers are sold out. GE Vernova's backlog is $163B and heading toward $200B, with turbines booked out nearly a decade; a Quanta executive described building brand-new domestic transformer and 765 kV switchgear capacity because the grid's aging iron simply can't handle the new loads.
  • Uranium's floor is rising. A nuclear-fuel buyer says the term price just hit an all-time high (~$95/lb) even as spot sits near $85, and the term price sets the floor. Bulls love the commodity; several are wary of the miners.

What's new

AI labs, not hyperscalers, are now the demand you should watch, and they're going off-grid. The sharpest conversation of the week was on Catalyst with Shayle Kann, "Inside the AI power wars". Kann's guest argued that if you want to understand power today, "you don't have to ask about hyperscalers. You need to ask about AI labs", because a big share of what Amazon and Microsoft are building is really for OpenAI and Anthropic. The math is the whole story: data-center buildout is running at tens of gigawatts a year and rising roughly 50% annually with "no signs whatsoever of a slowdown," while the grid adds maybe 5-6 GW of new gas and 20-25 GW of solar-and-battery nameplate a year, "just not enough." (Nameplate is the sticker-rating of a plant; the usable firm amount is much less.) A single gigawatt of data center runs about $50 billion of capex, or $60-70 billion if someone else funds it, and, as the guest put it, "you cannot make that kind of investment decision if there's uncertainty on the timeline." So Anthropic, targeting 1.5 GW at the end of 2025 and 10+ GW by 2027, "building a Google in two years", brings its own power. Why it matters: this is the demand signal underneath every equipment and gas order below.

GE Vernova is sold out for a decade, and even the skeptics are buying the story. On Motley Fool Hidden Gems Investing, "Meta Has Been Busy", the hosts walked through why GE Vernova (GEV) may be "the best positioned power stock for the AI infrastructure buildout": a $163 billion backlog expected to reach $200 billion by next year, an electrification segment that "booked more data center orders in the first quarter than it did in all of 2025," and turbine production "essentially sold out for almost a decade." The clever part is the margin shape: turbines earn maybe 10%, but the aftermarket parts-and-service that follows for decades earns nearly triple. With ~400 GW of turbines installed today and ~200 GW more planned over five to seven years, today's low-margin build phase is planting a high-margin annuity. The honest caveat, in their words: the stock is up over 700%, "most of the gains recently have been valuation," and the bull case "assumes the demand cycle we're seeing is going to last for years into the future."

An operator's view of the transformer choke point. The most useful insider color came from two Quanta Services (PWR) executives on Powerline Podcast, "3 Myths About Manufacturing Transformers in America". Their framing: a 1 GW campus used to be a landmark, "that's one nuclear plant pretty much full output," historically enough to serve 800,000 residents, and now sits on a single 500-to-1,000-acre parcel, with announcements of "three gigawatts, five gigawatts, 10 gigawatts." Meanwhile the existing grid runs on transformers "80 years old, 60 years old," made of paper and oil. So Quanta is spending to build domestic capacity for transformers, shunt reactors (which regulate voltage on long lines) and switchgear, including a new extra-high-voltage (765 kV) facility near Pittsburgh coming online "the end of 27 and the early 28," and a plant mothballed in 2014 being revived for switchgear "all the way up to 765." They're hiring 300-500 people at that one site alone. When an operator is pouring concrete for new high-voltage capacity, that's a real lead-time and pricing-power signal, not a talking point.

The macro case, with numbers, from someone running the trade. On Mission Matters, "Powering the Infrastructure Behind the AI Revolution", Kineo Capital co-founder Jim Struger laid out the supercycle thesis his fund is built on. Best estimates put the US power gap at 60-90 GW by 2030, in reactor terms, "60 to 90 nuclear reactors short," and the last US build (Vogtle, near Augusta) "took 15 years and was $15 billion over budget." He leans on Nvidia's Jensen Huang forecasting $3-4 trillion of annual AI-infrastructure capex by 2030 and $8.5 trillion cumulatively through 2040, reverse-engineered, a 48% annual growth rate off a ~$500 billion 2025 base. His policy read matters for the whole complex: the Department of Energy invoked the Defense Production Act in late April, formally classifying AI compute and the energy stack as national security; it's offering ~$18 billion in low-interest loans for 10 Westinghouse AP1000 reactors; and FERC now effectively tells hyperscalers to bring their own power (he points to a Microsoft-Chevron plan for ~3 GW of off-grid nuclear-and-gas capacity in East Texas). He invests "from the GPU out to the utility" and deliberately owns no software.

Uranium's term price just hit an all-time high, from the person who buys the fuel. On Sprott Radio, "Uranium Update from Per's Cabin – The Sequel", WMC's director of nuclear fuel, Per Jandier, gave the cleanest primer of the week. Spot uranium is around $85/lb, but the term price (UX $94, TradeTech $95) is "the highest it's ever been", higher even than the 2007 spike, and rising every month. Since roughly 85% of utility volume is bought through long-term term contracts and only ~15% on spot, and the market is in contango (term above spot), "there is a limit to how low the spot price can go", if the gap tops ~$15 a pound, a trader or utility just buys spot and carries it. He calls this "the best environment for nuclear I've ever seen in my 25 years," with demand "coming out of everywhere" and supply struggling (Kazatomprom hinting it may miss targets, Cameco back on track after a Saskatchewan bridge washed out in spring flooding, Lotus short ~1M lbs it owes over six months). Why it matters: a rising term price is the leading indicator; spot follows.

The debate

This week the bull and bear cases weren't shouted across a table, they came from different rooms. Worth steel-manning both.

The bull case: a genuine, multi-year supercycle that lifts the whole stack together. The through-line from Kineo's Struger, the Motley Fool hosts, and Jerry Robinson on Follow the Money, "AI's Power Problem" is that this is picks-and-shovels investing at its purest. As Robinson put it, "AI may be the gold rush, but power infrastructure may be the picks and the shovels", and the beauty is "it doesn't require us to know which AI model wins." Demand growth of ~50% a year, backlogs booked out a decade, a term uranium price at record highs, and Washington now treating the energy stack as national security all point the same direction. The supply side simply can't respond quickly: on Earthlings 2.0, "Cracking the Code on Building Energy Costs", an Edgecom Energy executive noted grids that historically grew "1%, 2%, 3% per year" are now growing "8% to 10%," and the real bottleneck is people, not technology, "even if you set up a natural gas turbine manufacturing plant tomorrow, you won't have the talent." Scarcity that takes a decade to fix is exactly what sustains pricing power.

The bear case: an aging-grid story wearing an AI costume, and a political ceiling. The most sober voice was regulatory and grid-planning veteran Alice Yake on Columbia Energy Exchange, "Planning for a Reliable, Cleaner Grid". Her blunt line: "it's not at all about the data centers." The average US grid asset is about 55 years old, so a giant replacement wave, and higher bills, was coming regardless of AI; against that spend, what data centers put into the grid is "a rounding error." That cuts both ways for the bull thesis: some of the "AI capex" is really overdue maintenance that would have happened anyway. She also flagged a real cost most people miss: data centers raise the system's "load factor" (how constantly the grid is used), which makes generators run harder, raising maintenance and shortening asset life.

"The amount of money that needs to be spent on the grid as compared to the amount of money that data centers are spending on the infrastructure... it's a rounding error." Alice Yake, Columbia Energy Exchange

The other half of the bear case is political, and it got louder this week. On Thoughts on the Market, "AI's Next Stress Test", Morgan Stanley's Ariana Salvatore and Stephen Byrd cited Data Center Watch data that 75 data-center projects worth $130 billion were blocked or delayed in the first quarter of 2026 alone, as many as in all of 2025, across blue states weighing moratoria (New York, Michigan, Illinois, Minnesota) and red-leaning ones (Pennsylvania, Arizona, Ohio, parts of Texas) pulling tax incentives. Their base case isn't a ban but a "conditional buildout" that loads grid-modernization and community give-backs onto every project, and a wave of off-grid projects to sidestep the fight. You could see that resistance up close on Closed! NYC's Real Estate, "New York's Data Center Moratorium", where State Senator Kristen Gonzalez explained her one-year "yellow light" pause: 28-plus proposals totaling ~9,000 MW, "about a third of the energy New York State uses today," and a demand-supply gap NYISO says exists by 2030 "if we added nothing to the grid." She wants a separate rate class so those loads don't land on household bills.

Where the week actually landed: the supply-shortage bull case dominated the mics, and the two thoughtful counters weren't really "there's no boom", they were "some of this is overdue maintenance, not new growth" and "politics will slow and reshape it, not stop it."

The names in play

GE Vernova (GEV) is the cleanest expression of the bull case and got the most airtime: sold-out turbines, a $200 billion backlog on the horizon, and a service annuity building underneath. The debate isn't demand; it's whether a stock up 700% has already priced a decade of it.

Quanta Services (PWR) came through not as a stock pitch but as an operator quietly committing capital to 765 kV transformers and switchgear, the physical bottleneck the Catalyst guest called out (a high-voltage transformer "takes 3 years to order"). If you believe transmission and distribution is the real constraint, this is where it shows up.

Eaton (ETN) and Vertiv (VRT) were framed by Follow the Money as the power-management and data-center-cooling picks, Vertiv's angle being that "as AI chips get more powerful, they begin to run hotter," so cooling demand scales with compute.

Cameco (CCJ) is the consensus quality name in uranium, but the equity-vs-commodity split is worth noting. On Money of Mine, "Why Small Caps Are Trading Too Cheap", fund manager Matt Griffin said uranium has been "structurally undersupplied for 15 years" and that Cameco is "locking in deals above $100 a pound... ceilings up to $140, $150", yet he avoids developers and juniors because "it's hard to think of a Western world uranium project that's actually delivered anywhere near on time and on budget," and he flagged the brutal working-capital cycle (Lotus may not see first cash until roughly a year after commissioning). On Palisades Gold Radio, Dr. Nomi Prins made the mirror-image point: Western-hemisphere uranium miners "have significantly underperformed" a stable-to-rising commodity, a lag she reads as opportunity (she named UEC).

Cheniere (LNG) and the gas-turbine suppliers. On Power Lunch, Baker Hughes CEO Lorenzo Simonelli described "two deals in two days", a Cheniere award to expand Sabine Pass Train 7 (adding 6 million tons a year of LNG) and a multi-year deal to supply gas turbines for data centers, calling it a "demand cycle decade" that is "not slowing down." His mention of "power generation needs off the grid, behind the meter" is the same behind-the-meter theme, straight from a supplier.

Read-throughs

  • From behind-the-meter to the engine makers. The Catalyst discussion is a near-explicit read-through to on-site generation: XAI kicked it off with 20-30 MW aeroderivative turbines (jet-engine-derived units that start fast), then OpenAI/Crusoe/Oracle in Abilene and Meta in Columbus followed, and OpenAI's Oracle deal in Shackleford County, Texas involves 2.3 GW of reciprocating engines (think very large piston engines). That's a direct pull on Cummins (CMI) and Caterpillar (CAT) engine franchises and on turbine OEMs, and on Bloom Energy's fuel cells, described as "sold out till 2031" and the "ultimate play on power constraints" for islanded sites.
  • From nuclear demand to the fuel cycle. If utilities are signing term contracts at record prices, the read-through runs past the reactor operators (Constellation, Vistra) to converters and enrichers and physical-uranium vehicles, the SWU/enrichment bottleneck Follow the Money flagged via Centrus (LEU), and the miners that ultimately have to fill those contracts.
  • From data-center gas demand to pipelines and storage. On NGI's Hub & Flow, an East Daley analyst pegged ~3.5 Bcf/d of incremental Texas data-center gas demand ("playing with house money") on top of ~16 Bcf/d of new LNG capacity by 2030, and warned that US gas-storage withdrawal capacity is slipping below 100% of demand coverage (it was 108% in 2022). Because data-center demand is "non-intermittent" (not weather-driven), any winter disruption gets amplified, bullish gas-weighted E&Ps and volatility. The RBN Energy piece on the proposed Eddystone LNG terminal near Philadelphia (7.2 mtpa, ~1.1 Bcf/d of feed gas via Williams' Transco and Enbridge's TETCO) is the same story from the takeaway side, more calls on Marcellus-Utica gas and the pipes that move it, even if first gas is 2032 at the earliest.
  • The odd read-through of the week: Bitcoin miners as AI power landlords. Also on Power Lunch, Morgan Stanley's Stephen Byrd said the TerraWolf deal implied about $19 of net value per watt of power, versus miner stocks trading around $3-6/watt (and ~$1/watt two years ago). The read-through: companies sitting on grid-connected power are being repriced as AI landlords, "all AI stories are really just energy stories in disguise."