Newsletter · · Ashutosh Agarwal

Data Centers Now Own 94 Percent of PJM Peak Load Growth - Powering AI Infrastructure - Week of July 7, 2026

Powering AI Infrastructure for the week of July 7, 2026. PJM's own SVP says data centers now drive 94 percent of the grid's peak-load growth through 2030, and the week's podcasts split between operators pricing the scarcity trade and strategists questioning the multiples on the buildout.

Powering AI Infrastructure

Week of July 7, 2026: Data Centers Now Own 94 Percent of PJM Peak Load Growth

Here's the number that reframes the whole debate: 94%. That's the share of PJM's projected peak-load growth through 2030 that data centers now account for, straight from the grid operator's own SVP. Not a share. Nearly all of it. And the week's tape kept circling the same uncomfortable gap between how fast compute wants power and how slowly the grid can hand it over.

If you run a book across grid, gas, generation, or nuclear, this was a week to lean in.

TL;DR

  • PJM is the tell. Data centers = 94% of peak-load growth through 2030; a heat wave sent day-ahead power up ~900% to $436/MWh, with a one-hour print above $1,200. IPPs are the direct beneficiaries.
  • Behind-the-meter went vertical. The average proposed North American on-site project hit ~2 gigawatts in Q1, 10-100x historical norms, and it's gas. Read that straight through to engines, turbines and gas demand.
  • The bull-bear line is now about multiples, not demand. Nobody's arguing the buildout is fake. The argument is whether 40x on a cyclical capex input survives the second derivative.

What's New

The grid operator said the quiet part out loud. On TED Tech, PJM's Asim Haque laid out the math with unusual candor: his 180,000 MW system is heading to 220,000+ MW within a decade, data centers are 94% of that growth, and over 50,000 MW of mostly-renewable supply that's already cleared PJM's process is stuck, the number-one reason being state permitting, siting and NIMBYism. His one-liner is the whole thesis: "You can build a data center in two years. The power plant to fuel it takes seven." That seven-year wall is why merchant power and long-lead equipment have pricing power.

A heat wave stress-tested the bull case in real time. Reporting on Squawk on the Street, Pippa Stevens walked through PJM's "imminent electricity reliability emergency" warning: record demand of 166 GW, day-ahead prices up from $44/MWh the prior week to $436 (a ~900% jump), and a one-hour peak from 7-8 p.m. north of $1,200. Her read on who gets paid: "The independent power producers, though, like Constellation, Vistra, and Talen, could be beneficiaries here since they're getting higher prices for the power they sell into competitive markets." Note the tell, Henry Hub actually fell on the week. This is scarcity value in the electron, not the fuel.

Behind-the-meter stopped being a footnote. The single most actionable operator data point came from EnergyCents, where the guest, who tracks these projects quarter by quarter, said the average proposed North American behind-the-meter project reached roughly 2 GW in Q1 2026, versus historical norms of single-digit to tens of megawatts, and stressed it's broad-based, "just seeing bigger projects across the board." It's gas: reciprocating engines from Caterpillar, Wärtsilä and Cummins, smaller industrial turbines, and fuel cells. The "speed-to-power premium" now runs ~80% above retail, about $140-150/MWh vs. ~$80. And the punchline for anyone modeling turbine backlogs: Baker Hughes is "sold out through 2028," Siemens is quoting 24-36 months, and GE's 2029 capacity is slipping toward 2030 on EPC bottlenecks. Chevron just signed a 20-year offtake with Microsoft in Texas, one of the longest such deals yet, and a sign some of this "bridge" power isn't a bridge at all.

The grid itself is old, and someone has to pay to replace it. On Columbia Energy Exchange, Columbia's Doug Arent and Robin Millican put hard numbers on the rate-base story: distribution capex rose ~160% from 2003 to 2023 to nearly $51 billion, now 43% of investor-owned utility capex, with 70% of power transformers over 25 years old, 60% of breakers over 30, and wires and cables up 150% since 2019. That's the CAGR under utility EPS. But they're the skeptics of the group: they note load growth is not the dominant national driver of prices (fuel is), and where states demand financial commitment, "expected demand is down by a third." Useful ballast against the hype.

The Debate

This is the sharpest part of the week, because the bull and bear cases finally stopped talking past each other.

Bull: The demand is contracted, not speculative. On Money Rehab, Hightower's chief investment strategist, 35 years covering industrials, said the picks-and-shovels names (Quanta, GE Vernova, Vertiv, Eaton, Rockwell, Vistra) averaged 34% backlog growth year over year against a 5% historical norm. "Backlog is stickier stuff. It's contracted stuff." She flagged Quanta lifting its 2030 TAM from $960B to $2.4T at a recent analyst day, and GE Vernova "sold out in power until 2028." The steel-man: this is a multi-year re-rating of grid, IPPs, gas and equipment together, funded by hyperscaler capex of ~$800B this year (up 75%) heading to ~$1.1T.

"You're putting very high multiples on what is basically the front end of the investment."

Bear: That quote is from Erik on Monetary Matters, and it's the cleanest bear framing of the year. His worry isn't that demand is fake, he thinks capex is "headed to a trillion." It's that "you wouldn't be trading 40 times on connectors" if the market believed this was cyclical. If the second derivative rolls, $400B, $600B, $800B, a trillion, then back to $750B, the highest-multiple supply-chain names get repriced hardest. Even the bull agreed with the mechanism: if capex slips from $800B toward $500B, "this whole thing kind of starts to unravel." So the debate isn't up or down. It's what multiple you pay for the front end of someone else's capex cycle.

One honest caveat: this week's tape leaned bullish on the physical demand and only the two market-strategy shows voiced the valuation bear case. Nobody made a serious "renewables and efficiency cap the whole move" argument, worth watching whether that surfaces next week.

The Names in Play

IPPs (VST, CEG, TLN) are the purest expression of the scarcity trade, the heat-wave prints show exactly how merchant spreads behave when reserves thin out. GE Vernova got name-checked as a bull and a bear this week: sold out through 2028 on one show, the poster child for cyclical-multiple risk on another. Vertiv got its own love letter on Money On Tap, framed as the dominant liquid-cooling provider with only 20-30% of AI data centers built. Quanta (PWR) is the grid-labor pure-play behind that $51B distribution capex wave.

Read-Throughs

The threads connect in ways the single-stock desks sometimes miss:

  • Gensets (CMI, CAT): The 2 GW behind-the-meter number is a direct order book for reciprocating engines. When Baker Hughes and Siemens turbines are sold out to 2028-2030, engines and fuel cells absorb the overflow.
  • Copper and switchgear (FCX): On Interchange Recharged, Siemens' Nick Chaset explained the shift to 800-volt DC as racks blow past 1 MW, a design that cuts copper up to 45% and needs new solid-state breakers. Watch that: DC architectures are a modest headwind to copper-per-rack even as total demand explodes, and a tailwind to next-gen switchgear.
  • Hyperscalers funding the grid: On Let's Talk Energy, Microsoft's Per Christian Honningsvaag reaffirmed the pledge to pay for transmission and substation upgrades directly, and named GE Vernova and Schneider as digitization partners. Electric Perspectives showed the mechanism live: Alliant got a five-year rate stay-out in Iowa by having QTS pay the marginal cost of 500 MW-1 GW of load, the customer funds the buildout, existing ratepayers stay flat. That's the template that keeps the political permission slip valid.

What Changed

The reliability story got a harder edge. InvestTalk flagged that NERC issued only the third rare alert in its history, warning that data-center clusters are spontaneously dropping to backup power during minor disturbances, with loads swinging 50+ MW in minutes. And on the supply side, Factor This noted PJM's median interconnection queue is still 60 months even after FERC Order 2023, with developers needing "10 projects to get 2 across the finish line." Both cut the same way: the constraint is getting worse, which is bullish for anyone who already has capacity, backlog, or a cleared queue position, and brutal for anyone who doesn't.

Watch into next week: any PJM capacity-auction commentary, more 20-year behind-the-meter offtakes like the Chevron/Microsoft deal, and whether a credible efficiency/renewables bear finally shows up to argue the other side.