Newsletter · 2026-07-10 · Ashutosh Agarwal

Power, Not Chips, Is the AI Bottleneck Now - Powering AI: Grid, Gas, Generation & Nuclear - Week of July 10, 2026

Powering AI newsletter for the week of July 10, 2026. Grid operators, equipment makers, and nuclear developers made the case that electricity, not silicon, is now the binding constraint on the AI build-out, with PJM pinning 94 percent of its peak-load growth on data centers and GE Vernova, Quanta, Vertiv, and Eaton as the names guests moved on.

Powering AI: Grid, Gas, Generation & Nuclear

Week of July 10, 2026: Power, Not Chips, Is the AI Bottleneck Now

A year ago, on a 90-degree afternoon in Baltimore, the country's largest grid operator did something it had never done in that region: it ordered utilities to cut power to rotating pockets of Maryland. Not a storm. Not a downed line. Just too much demand for the wires to carry. This week, the people who run that grid, dig the copper, forge the transformers, and split the atoms all showed up on podcasts to say a version of the same thing: the AI boom no longer has a chip problem. It has a power problem. And the power problem is now the whole trade.

Here's the line that reframes it. For two years the AI story was about who makes the smartest model and the fastest chip. This week, one operator after another argued the next chapter is simpler and more physical: who can actually get the electrons to the building. That's good news if you own the picks and shovels. It's a warning if you're paying up for a boom that has to last a decade to make the math work.

TL;DR

• The grid, not the GPU, is the choke point. PJM now expects data centers to drive 94% of its peak-load growth through 2030, and a rare "five-alarm" alert from the reliability regulator says clusters of data centers are literally destabilizing the grid. The bottleneck is where the pricing power lives.
• A real debate broke out this week: build your own power behind the fence (gas engines, fuel cells, small reactors) versus fix the grid and make data centers flexible. Both roads run straight through the same equipment makers, transformers, switchgear, turbines, cooling.
• The names guests actually moved on: GE Vernova (a backlog now bigger than most countries' power plans), Quanta Services (which just doubled-and-then-some its own market estimate), Vertiv, Eaton, plus a fresh nuclear milestone from a startup that just went critical next to a data center.

What's New

PJM says data centers now own 94% of its growth, and the grid is buckling. On TED Tech, Asim Haque, a senior vice president at PJM Interconnection (the operator for 67 million people across 13 states), laid out the math plainly: PJM is a roughly 180,000-megawatt system heading to 220,000-plus within a decade, and data centers account for 94% of that projected peak-load growth. His killer detail on why this gets expensive and messy: "You can build a data center in two years. The power plant to fuel it takes seven." He also flagged that PJM has already studied over 50,000 megawatts of mostly-renewable projects that could plug in now but are stuck, the number-one reason being permitting and local opposition. And in one recent capacity auction, the clearing price jumped from $28 to about $270 per megawatt-day in a single year. That is the sound of scarcity showing up in a market price, and it flows to merchant generators and, eventually, everyone's bill.

The reliability regulator issued only its third-ever alert, and used the words "five-alarm fire." On InvestTalk, the hosts walked through a genuinely new problem: NERC (the North American grid-reliability body) warned that clusters of data centers have been spontaneously disconnecting from the grid during minor line faults and flipping to backup power, and when a group does that at once, it can trigger cascading failures. Data-center power draw can swing "50 megawatts or more within minutes," the equivalent of switching 50 Walmarts on and off. PJM has even secured emergency federal orders letting it curtail data centers as a last resort. The blunt investor takeaway from the show: "The AI trade can't exist without the power trade."

The "AI power wars" have a clear front line: behind-the-meter vs. the grid. Catalyst with Shayle Kann hosted the sharpest debate of the week. One side: data-center demand is growing ~50% a year, tens of gigawatts annually, while the grid is adding maybe 5-6 GW of gas and 20-25 GW of nameplate solar/battery, "rookie numbers" against that demand. So the AI labs (the guest argues Anthropic wants to go from ~1.5 GW at end-2025 to over 10 GW by 2027, "you're building a Google in two years") simply bring their own generation. Shayle Kann pushed back: more generation is coming, and the real gum-up is transmission and distribution, "you need a high-voltage transformer, and that takes three years to order." Either way, the money lands on the same doorstep: turbines, transformers, switchgear. A striking financing detail: Google has reportedly stacked up ~$50 billion of backstop obligations to enable Anthropic's buildout, and OpenAI signed a deal to deploy 2.3 gigawatts of reciprocating engines in Texas. On-site gas engines are having a moment.

GE Vernova's backlog is now bigger than most countries' entire power plans. On Motley Fool Hidden Gems Investing, the hosts made the cleanest single-stock case of the week: GE Vernova (GEV) sits on a $163 billion backlog expected to hit $200 billion next year, its electrification unit "booked more data center orders in the first quarter than it did in all of 2025," and its turbines are "essentially sold out for almost a decade." The nuance a casual read misses: GE Vernova makes only ~10% margins selling the turbine, but the aftermarket service on that turbine runs at "nearly triple the margins" and lasts for decades. With ~400 GW of turbines installed and ~200 GW more coming over five to seven years, today's low-margin equipment surge is really seeding a long, high-margin annuity. The catch, which the hosts were honest about: at ~700% since the spin-off, "most of the gains recently have been valuation," and the price only holds if the AI capex cycle really does run a decade.

A startup went critical next to a data center, the first co-built reactor-plus-data-center. On Nuclear Barbarians, Matt Loszak, CEO of Aalo Atomics, described hitting cold (zero-power) criticality on a 10-megawatt reactor, part of a Department of Energy pilot that challenged 3 of 10 companies to reach criticality by July 4, 2026. Full power and a co-located Crusoe data center follow shortly, which he calls the world's first plant and data center built side by side. The investor-relevant strategy choice: Aalo is using ordinary, off-the-shelf low-enriched uranium fuel (the kind today's reactors run on, with a ~40 GW/year supply chain) rather than the exotic TRISO/HALEU fuels most rivals need, and is targeting 14-15 cents per kilowatt-hour to compete head-on with gas. He confirmed Oklo (OKLO) is running two projects in the same program, including one for isotope production. This is the clearest sign yet that small reactors are moving from slide decks to physical milestones, still years from scale, but no longer theoretical.

The Debate

Bull: this is one supercycle, and it re-rates the whole stack together. The most confident version came from Stephanie Link, Hightower's chief investment strategist, on Money Rehab with Nicole Lapin. Her case: the big four tech companies will spend ~$800 billion this year, up ~75%, heading toward ~$1.1 trillion in 2027, and the beneficiaries fan out far beyond chips into "the food chain" of grid, power and cooling. Her anchor stat is the one to remember: at its recent analyst day, Quanta Services (PWR) raised its 2030 addressable market from $960 billion to $2.4 trillion, "I almost fell off my chair." She notes 75% of the US grid is over 25 years old, GE Vernova is sold out in power to 2028 and supplies ~30% of global electricity, and, crucially, the picks-and-shovels names (Vertiv, GE Vernova, Quanta, Eaton, Vistra) are showing ~34% average backlog growth versus a normal ~5%. Her one honest fear: "Where I'm wrong is if they start to cut back on the capex." So far, she says, it's the opposite.

PIMCO's Greg Sharenow, on PIMCO Accrued Interest, sharpened the macro version: AI is one of three simultaneous, commodity-hungry capex cycles (alongside military-industrial and the energy transition), and because permitting and local opposition make it "hard to supply all the energy that is needed," he's bullish near-term power pricing. He also flagged the supply-side kicker for metals: capital spending per dollar of free cash flow across miners is at 40-year lows, so the raw materials pipeline is thin just as demand inflects.

Bear: the load is real, but the "it's all AI" story is oversold, and much of the fix isn't a new plant at all. The most useful pushback came from grid-planning veteran Alice Yake on Columbia Energy Exchange. Her contrarian line: rising bills are "not at all about the data centers." The US grid averages ~55 years old, and simply replacing that aging base means prices go up "even without load growth." Her most quotable jab at the bull thesis: compared with the grid capex that has to happen anyway, what data centers spend directly on grid infrastructure is "a rounding error." She's not dismissing the strain, data centers do overload specific regions and change how hard assets get worked, but she's warning that a lot of this week's demand headlines describe a problem that predates AI.

The other half of the bear case is that the demand can simply be managed away. On Next in Tech, Emerald AI CEO Varun Sivaram made the flexibility argument that should worry anyone modeling a straight-line buildout: the US wants 50+ GW of AI data centers by 2028 but only ~25 GW can connect, yet the grid sits at "99% spare capacity most of the year," so if data centers flex during the rare peak hours, you could fit ~100 GW on today's infrastructure and avoid ~$4 trillion of grid buildout. He points to a June 18, 2026 FERC action and a new Texas rule creating a "fast lane" for flexible loads, and demos with NVIDIA and Oracle (including a 100 MW site in Manassas, Virginia). If flexibility scales, some of the panic-buying of hard equipment gets deferred.

BloombergNEF put numbers on that on Switched On: in ERCOT, data-center demand nearly doubles from ~8 GW to ~14 GW by 2035, and full data-center flexibility could hold 2035 wholesale prices roughly 180% below the no-flexibility "Chicken Little" case, a ~$259/MWh spread. Note the interconnection queue is a jaw-dropping ~500 GW against a peak load under 90 GW, so a lot of "demand" is speculative. Texas's Senate Bill 6 now requires large loads over 75 MW to demonstrate flexibility before connecting. The bull and bear actually agree on the load; they disagree on how much new steel it forces into the ground.

The Names in Play

GE Vernova (GEV) was the week's most-discussed single stock, and the read is consistent across shows: the demand clearly justifies today's numbers, and the sold-out turbine book plus tripling aftermarket margins give unusual visibility, but the valuation only works if AI capex holds for years. Bull and bear here is really one question: how long does the cycle run.

Quanta Services (PWR) got two independent, credible mentions. Stephanie Link's $960B-to-$2.4T TAM jump is the headline, but the ground-level confirmation came from operators at Pennsylvania Transformer (a Quanta business) on Powerline Podcast. Their color: data-center campuses have jumped from 1 GW ("one nuclear plant's full output") to announced 3, 5, even 10 GW, and they're rebuilding a site to make the biggest EHV/UHV transformers (online late 2027/early 2028), reviving a mothballed line for switchgear up to 765 kV, and hiring 300-500 people. They also flagged a quieter tailwind: transformers 60 to 80 years old carrying real load, feeding a replacement supercycle that "trickles all the way down" to neighborhood distribution. This is operator testimony, not a pundit's model, the most grounded evidence of the equipment squeeze all week.

Vertiv (VRT) and Eaton (ETN) anchored the cooling-and-electrical layer. On Money On Tap, the hosts called Vertiv "the primary name" in liquid cooling with "a lot of life left" and no real challenger emerging, and named Eaton, Hubbell and Powell as "highly engaged" in switches and breakers, with a reminder that 70-80% of AI data centers are still to be built.

A small-cap worth knowing about the constraint, not necessarily the stock: on WTR Small-Cap Spotlight, AIB Data Centers CFO Jolene Haliski put the whole thesis in one sentence: "The bottleneck in our AI infrastructure industry today isn't chips. It's power delivery." Her firm locks executed utility power agreements before breaking ground, she calls uncommitted capacity "fake power," because interconnection queues now run years and "some markets are effectively closed to new large load customers." Whatever you think of a ~40-megawatt developer with a ~395-megawatt pipeline, that's an operator confirming the core view from the front line.

Read-Throughs

Gas is the near-term winner nobody wants to admit. On NGI's Hub & Flow, the analyst tracks ~3.5 Bcf/d of incremental Texas data-center gas demand, "house money" on top of ~16 Bcf/d of new LNG capacity by 2030 (Golden Pass, Corpus Christi trains, Rio Grande, Commonwealth). The key point for power investors: this demand is non-intermittent, it doesn't need a cold winter to show up, and US gas storage withdrawal capacity is slipping below 100% of demand coverage, which means more price volatility on any disruption. He floated a three-handle for Waha gas in 2027. That reads straight through to gas-for-power midstream (Williams, Kinder Morgan, ONEOK, Cheniere) and gas-weighted E&Ps (EQT, Coterra, Antero, Range), several of which Jerry Robinson mapped out by name on Follow the Money, where he also cited the IEA projecting global data-center electricity use more than doubling to ~945 TWh by 2030 and the EIA calling for record US power in 2026-27. His most vivid detail: an Ohio brick maker whose monthly grid capacity charge jumped from $1,600 to $12,000.

Copper is the pick-and-shovel under the picks and shovels. On Palisades Gold Radio, macro strategist Tavi Costa argued copper is "structurally bullish" on data-center, electrification and onshoring demand into thin supply, likening it to silver right before its big move and expecting it to "go crazy." (Every transformer, every substation, every foot of cable is copper, so this is the same trade one layer down, and it reads through to Freeport-McMoRan and the broader conductor complex. Note: most of this week's copper commentary came from junior-mining shows, so weight it accordingly.) Costa also made the blunt gas point: "You can't get nuclear quickly… the only solution in the near term is going to be natural gas."

Uranium's quiet re-rate setup. On Palisades Gold Radio, Dr. Nomi Prins noted uranium spot has held near $85 while utilities lock long-term supply at $95-$100, so the term price is effectively pulling spot up. Her wrinkle: Western-hemisphere miners have underperformed the commodity because uranium's paper/futures market is shallow and ETF selling hits the equities, setting up what she sees as a lag-then-catch-up trade in names like UEC. This reads through to Cameco (CCJ), the enrichers and converters, and physical-uranium vehicles as the nuclear-PPA and restart narrative builds.

The hyperscalers are financing the whole chain. The connective tissue across every episode: Meta, Microsoft, Amazon and Google (plus AI labs OpenAI and Anthropic) are the customers whose ~$800B-and-rising capex funds the turbines, transformers, gas engines and reactors above. Watch their capex guidance the way you'd watch a commodity price, because on this week's evidence, it is the commodity price for everything downstream. Stephanie Link's tell is the right one to track: backlog, not orders. Orders can be cancelled; backlog is contracted. When that ~34% backlog growth rolls over, the supercycle thesis rolls with it.