Newsletter · · Ashutosh Agarwal

Data Center REITs Get Their Cleanest Bull Case as Physics and Debt Set Limits - Powering AI Infrastructure - Week of September 24, 2026

The powering-AI-infrastructure weekly for the week of September 24, 2026, synthesizing REIT, credit, and power podcasts. A Jefferies analyst laid out the lease-reset bull case for data-center landlords Digital Realty and Equinix, Goldman sized hyperscaler capex above $6 trillion through 2030, and NRG's CEO capped the Texas demand hype at 40 gigawatts.

Powering AI Infrastructure

Week of September 24, 2026: Data Center REITs Get Their Cleanest Bull Case as Physics and Debt Set Limits

Something shifted this week. For two weeks running, the loudest voices in this corner of the market were power engineers and grid nerds, while the landlords themselves stayed quiet. This week the landlords got a proper hearing, a Jefferies real estate analyst laid out, in plain terms, exactly why the boring old data center and warehouse REITs might be the smarter way to play the AI buildout than the flashier stuff. And right on cue, the people who actually finance and power these buildings showed up to explain what it's going to cost, and where physics says the whole thing hits a wall.

TL;DR

  • The mark-to-market trade is finally being spelled out. Digital Realty and Equinix locked in leases at the bottom of the market in the late 2010s; those roll over at the end of this decade, and a Jefferies analyst thinks that flips these landlords from ~high-single-digit to low-to-mid-teens earnings growth.
  • The buildout is now a financing story, not just a construction story. Goldman's chief credit strategist pegs hyperscaler capex at more than $6 trillion through 2030 and sees their bond issuance jumping from ~$250B this year to $400B in 2027, with data centers being funded "like a utility," off balance sheet.
  • Physics is the governor. NRG's CEO says the headline 500-gigawatt demand number for Texas "is not going to happen", 40 GW is realistic, and the real bottlenecks are turbines, switchgear, and welders, not ambition.

What's new

A real estate analyst just made the cleanest bull case yet for the "boring" data center REITs. This week on WSJ's Take On the Week, Jefferies equity analyst John Peterson (a pundit/analyst, not an operator) drew the distinction that trips up most investors. The gigawatt-scale "as big as Manhattan" AI campuses you read about get built in Louisiana or West Texas. But the publicly traded landlords, Digital Realty and Equinix, own the plumbing in the big interconnection hubs: "Northern Virginia… Atlanta, Chicago, Dallas, Phoenix, Portland, Oregon, Silicon Valley." That's where enterprise data lives and where the cloud "availability zones from AWS and Microsoft Azure… are all located." As AI shifts from training brand-new models to inference (running the finished models for everyday queries), Peterson argues that demand flows back into exactly these markets. Why it matters: he thinks the leases these landlords signed "at the bottom of the market… the late 20-teens" come up for renewal "at the end of the decade" and reprice sharply higher, enough to move the group "from a high single digit earnings growth business to a kind of low to mid-teens earnings growth business."

The data center boom is quietly renting warehouses, too. In the same conversation, Peterson flagged a demand source that "sort of came out of nowhere": in the first half of this year, data center supply chains drove "a huge wave of demand for new logistics real estate." Schneider Electric leased warehouse space in Nashville to build switchgear; Corning needs space to store fiber-optic cable; and the hyperscalers themselves are renting logistics capacity near their campuses to stage GPUs before installation. Why it matters: it's a fresh, non-e-commerce demand leg for industrial landlords, but see the flip side below, because the same analyst is cautious on where warehouse rents go next.

Goldman put a number on the debt wall, and it's staggering. On Alpha Exchange, Amanda Lynam, Goldman Sachs' chief credit strategist (pundit/analyst), sized the financing behind all this concrete. Hyperscaler capex from 2026 through 2030 is "upwards of $6 trillion", and the hyperscalers are "only 40% of total AI-related gross issuance," so the true figure is much bigger. Their investment-grade bond issuance went from "$108 billion… in 2025" to "$229 billion" already this year, heading to a full-year "$250 billion," then jumping "a 60% increase… $400 billion" in 2027, plus "another $300 billion" just for data center and chip financing. Why it matters: the market can absorb the dollars, in her view; the real constraint is investors not wanting too much of one name or one theme. That's a subtle but important shift, the bottleneck is becoming appetite, not capacity.

Meta is being financed like a pipeline company. On The Investor's Podcast, hosts Daniel Mahncke and Shawn O'Malley (pundits/analysts) walked through the mechanics of Meta's Louisiana mega data center. Meta's 2026 capex guidance is now "between 130 and $145 billion," with 2027 estimates "close to $200 billion." But the Louisiana project sits off Meta's balance sheet inside a special purpose vehicle (SPV), a standalone company whose only job is to own that one building. Private-credit giant Blue Owl holds the majority equity ("something like 80-20"), the SPV "sold roughly $27 billion of bonds," and Meta signs a long lease and guarantees the building's residual value. Why it matters: this taps insurance companies and private credit funds "desperate for… long-dated bond-like exposure to AI infrastructure", a completely different money pool than equity markets. As the hosts put it, hyperscalers are now "financed like a utility." Meta's liabilities are "close to $700 billion" and have "grown about 800%" in a year; Alphabet carries "about a trillion dollars in off-balance sheet liabilities."

Alphabet stopped buying back stock to pay for data centers. On The Intrinsic Value Podcast, hosts Kyle Grieve and Shawn O'Malley (pundits/analysts) noted Alphabet now "plans to spend around $200 billion in a single year on data centers", roughly six times its 2023 level, "has stopped buying back its own stock entirely," and "raised over $100 billion in capital." Why it matters: a company famous for gushing free cash flow is now behaving like a heavy-industry capital sink, with return on invested capital at its "second-lowest levels since IPO." Whoever supplies the buildings and the power to that spend is the beneficiary; the question is what Alphabet earns back on it.

The debate

For the first time in a few weeks, both sides actually showed up.

The bull frame, scarcity re-rates the landlords. Peterson's case is the cleanest version: no one can build new data centers in Northern Virginia, which only makes the existing ones and the land banks more valuable, and the coming lease resets hand landlords real pricing power. Layer on a genuine power "demand super cycle" (NRG's CEO, below, doesn't dispute that part) and years of almost no new construction across property types, and you get a sector Peterson thinks can compound at ~7% earnings growth plus a 3.5% dividend, with the lowest leverage "in a very long time." His kicker: after the dot-com bubble burst, REITs outperformed the market "seven straight years" as investors fled tech for hard assets.

The bear frame, physics, and a mountain of locked-in obligations. The most quotable pushback came from an operator. On Let's Talk Energy, NRG CEO Robert Gaudette (operator/insider) took a blowtorch to the demand hype:

"In ERCOT, there's a number out there that says 500 gigawatts. Now that's on a system that has 91… That's not going to happen. Physics is going to win… 40 gigawatts is probably the right amount."

He ticks off the real constraints, "a turbine constraint… high load switch gear problems… definitely labor constraints," with the firms that build things "completely booked" and not enough "electricians and welders." That's a ceiling on how fast anyone can actually energize these buildings, hype notwithstanding.

The financial bear case is just as sobering. The Meta hosts warned that all this off-balance-sheet debt locks hyperscalers into contracts they'll "pay… whether the AI vision is working out or not," and that with "everyone and their mother building a data center right now," we could be talking about "so much more supply than demand" in about two years. And Peterson himself supplied the warehouse-side caution: logistics rents "peaked out in 2022," and because those are typically five-year leases, they're now "resetting", he expects earnings growth at Prologis and Stag Industrial to "moderate over the next few years." Even the data center landlords face an eventual "day of reckoning" when today's 15- and 20-year leases mark back to market, though, as he dryly noted, "that's 15 years from now."

Net: this week the bull and bear weren't arguing about direction, both agree demand is real and enormous. They're arguing about pace and price: how fast physics and the workforce let it happen, and how much debt gets buried in SPVs along the way.

Names in play

  • Digital Realty (DLR) / Equinix (EQIX), Peterson's clearest longs. The whole thesis rests on land banks in constrained tier-one markets and late-2010s leases repricing higher at end of decade. Next catalyst: watch transmission capacity coming online in Northern Virginia/Atlanta/Dallas, which he says "unlocks development potential."
  • Prologis (PLD) / Stag Industrial (STAG), more nuanced. Real near-term help from data center supply-chain tenants (switchgear, fiber, GPU staging), but rents that peaked in 2022 are rolling to lower marks. Bull and bear in one name.
  • Meta (META) / Alphabet (GOOGL), the tenants whose capex underwrites everyone else. The read-through for landlords and power suppliers is bullish; the read-through for the tenants' own returns is the open question (Alphabet's ROIC at post-IPO lows; Meta's liabilities +800% YoY).
  • NRG, Gaudette is pitching "Bring Your Own Power" as the model for large loads: NRG's product delivers "1.2 gigawatts of supply for every one gigawatt of load." A way to play data center power without betting on the most speculative demand forecasts.

Read-throughs

  • Power & grid. Beyond NRG, energy VC Dr. Carly Anderson on The Contrarian Capitalist Podcast argued the near-term fix isn't giant reactors but micro-reactors in the "one to 20 megawatt range," mass-manufactured to get on the grid "before 2030" without needing a "$100 per megawatt hour, 20 year… power purchase agreement." She flagged transformer lead times of "two, three, five year[s]" as a real chokepoint, good for anyone selling grid interconnection gear. A useful reality check buried in there: only "about 30 to 40%" of the past year's US electricity demand growth was data-center driven; the rest is reshored manufacturing and fleet charging.
  • Electrical equipment. The switchgear and transformer bottleneck keeps surfacing (Schneider's Nashville switchgear plant lease per Peterson; multi-year transformer lead times per Anderson). The picks-and-shovels demand is if anything more supply-constrained than the buildings.
  • Freight & logistics. The demand signal is split. On FreightWaves Today | September 22, RxO's VP of pricing Corey Kluza (operator) said "flatbed has been very positive… effectively the whole year," explicitly driven by "data center dynamics and reshoring… as well as… power industry dynamics", i.e., hauling heavy equipment to construction sites. But van and reefer softness is "a capacity story, not a demand story," with consumer freight "surprisingly soft" after Labor Day. On FreightWaves Today | September 18, the team noted Norfolk Southern intermodal (rail-plus-truck containers) volume "rose 13.7 percent year-over-year," and that weight-loss drugs are visibly denting food-and-beverage freight, Walmart's data shows households with a male on a GLP-1 buying "10% less" in calories. Translation for industrial landlords: the AI-adjacent freight is booming, the consumer-goods freight that fills e-commerce warehouses is not.

What changed

  • The industrial and data center REITs got a direct advocate again. After two weeks where single-name REIT commentary went missing and everything came through the power/grid side door, an equity analyst covered Digital Realty, Equinix, Prologis and Stag by name this week, with a specific rent-reset thesis. That's a genuine reappearance of the landlord-level narrative.
  • Still silent: the towers. For a third straight week, there was no dedicated discussion of the tower REITs, American Tower, Crown Castle, SBA, or of carrier capex and the Crown Castle fiber/small-cell review. No one voiced a tower thesis this week, bull or bear; we're not going to invent one.