Newsletter · · Ashutosh Agarwal

AI Capex & The Bubble Debate - Week of June 11, 2026: Oracle's Capex Bill Scares the Tape, Power Still Wins

AI capex and the bubble debate newsletter for the week of June 11, 2026. Oracle's capex blowout finally spooked the tape, but the operators pouring concrete and signing power contracts still sound supply-starved.

AI Capex & The Bubble Debate

Week of June 11, 2026: Oracle's Capex Bill Scares the Tape, Power Still Wins

Funny week. The same buildout that's supposed to re-rate every landlord with a power cord finally sent a chill through the tape, Oracle spending more than 100% of revenue on capex will do that. But peel back the panic and the operators who actually pour the concrete and sign the power contracts sound as supply-starved as ever. The gap between the screen reaction and the ground truth is where our edge lives this week.

TL;DR

  • Oracle's capex blowout ($90–95B against ~$90B of revenue) became the whole market's Rorschach test, bears called it the top, operators called it Tuesday. Either way it's the loudest "is AI overbuilt?" debate we've heard all year.
  • Power is still the binding constraint, full stop. Microsoft's own infra chief says you can't find a spare gigawatt on any grid; the response is to network smaller data centers together. That's the bull case for behind-the-meter power and the merchant nuclear fleet, said out loud by the buyer.
  • Industrial is quietly bifurcating: record-low new warehouse supply (a tailwind), but the first give-back of pandemic-era rent gains and 8%+ distribution vacancy (a yellow flag). Don't sleep on the rent line.

What's New

Oracle turned the AI-capex question into a market-wide gut check. On Squawk on the Street (Jun 11), the desk walked through FY27 net cash outlay of "about $70 billion," an RPO that "was gigantic, it's $850 billion," and the Michigan Stargate build: "$16 billion to build it... $30, $35 billion to outfit it... almost $50 billion for a gigawatt." For a landlord, that last number is the thesis, replacement cost is exploding, which is bullish for stabilized assets and brutal for anyone underwriting new development math.

The bear translation showed up fast, and it's about tokens, not just dollars. Narwhal Capital's PMs, on The Morning Market Briefing (Jun 11), flagged that Oracle is "about ready to spend over 100% of revenue on CapEx," then made the point that should worry every DC bull:

"If tokens is what you measure GPU throughput and AI usage in, it's the first indication to me that maybe we're overbuilding... we have too much supply of tokens and not enough demand."

That ties to OpenAI reportedly weighing "drastic" price cuts, The Rundown (Jun 11) also noted Oracle is now "the biggest corporate borrower in America outside of banks, with about $117 billion in bond debt" and "more than half of the $638 billion backlog... from a single customer, which is OpenAI." Single-tenant concentration, financed with debt. File that.

But the buyer of all this capacity says the constraint is still physics, not demand. Microsoft's Alistair Spears (GM, Azure Infrastructure, 19 years there) gave the cleanest operator read of the week on Tech Disruptors (Jun 9):

"It's very hard right now to find a spare gigawatt or five gigawatts on any single energy grid. And so the approach now... is really distributed training... network smaller data centers together."

He added that long-term land and energy contracts are "measured in decades... 10 or 20-year investments." This is a hyperscaler telling you the bottleneck is power and steel, not a shortage of tenants.

The private-capital crowd quantified the wall, and the failure rate. On Tech Disruptors (Jun 11), Apollo's Rob Bittencourt sized five-hyperscaler capex going "from below $100 billion" in 2019 to "over $750 billion," with a total AI infrastructure need of "five to six trillion dollars," addressable "only... through investment grade rated financing." Same episode, Helix Digital CEO Adam Selitsky (ex-AWS CEO, now partnered with KKR, Nvidia and Vistra) dropped the stat the bears will quote for months:

"Over 25% of data center projects which are announced now are not actually delivering, and that number seems to be increasing."

Translation: the announced pipeline is not the real pipeline. That's bullish for existing, powered, leased assets and bearish for the speculative-supply narrative.

Industrial isn't the clean "record-low supply" story everyone repeats. Two operators, two angles. Link Logistics CEO Luke Petherbridge (Blackstone's ~400M sq ft platform) on Inside the ICE House (Jun 8) called new supply "record low... like a decade low," with demand driven by e-commerce and "this infrastructure build out, whether it's data centres, whether it's power," even spillover land "transitioning to a data center." Bullish. But Moody's Analytics' Ermengarde Jabir, on America's Commercial Real Estate Show (Jun 10), flagged the crack: distribution vacancy "push above 8%," cap rates backing up to "6.4%, 6.5%," and, the line that matters,

"This is the first time in the entire period of growth... post 2020 for the sector where the sector has given back its pandemic era gains in terms of rent growth."

Low supply protects occupancy; it doesn't guarantee the mark-to-market spreads PLD/REXR bulls are modeling.

The Debate

Bull frame (well-voiced this week): Power is genuinely scarce, a hyperscaler infra chief, a former AWS CEO, and the merchant-nuclear story all said it independently. Constellation gets framed as "the largest supplier of 24-7 carbon-free power to hyperscale AI data centers via long-term PPAs" (InvestTalk, Jun 11), and Jensen Huang's "the choke point here is on the power side... we need something like a thousand times more power" got requoted on Market News with Rodney Lake (Jun 11). Add record-low industrial supply, and you have landlords sitting on irreplaceable, power-adjacent dirt.

Bear frame (also well-voiced): Capex is lapping revenue, financed with record debt and dilutive equity, against demand that may be softening at the token level. The Trepp team, on The TreppWire Podcast (Jun 5), cited "a JP Morgan analysis [that] found that more than 60% of data center capacity planned for completion in 2027 is not yet even under construction," and asked whether the $4T hyperscaler spend "gets pushed out to 2035, 2037." Pair that with the 25%-don't-deliver stat and the OpenAI price-cut chatter, and the bear case isn't "no demand," it's "the new-build economics break before the demand shows up."

Net: a rare week where both sides are argued by people who move the numbers. The bull case comes from buyers and builders of power; the bear case from allocators staring at the financing stack. That tells you where to point diligence: power offtake and balance-sheet capacity, not headline demand.

Read-throughs

  • Utilities / merchant nuclear: CEG is the podcast favorite as the AI-power proxy, but note it's "down 31.41%" YTD with a ~$90B cap and a Three Mile Island restart "delayed by transmission interconnection challenges" (InvestTalk, Jun 11). Vistra surfaced only as Helix's ~50 GW power partner (Tech Disruptors, Jun 11).
  • Electrical / cooling gear: thin tape, Vertiv got a passing nod as an "obvious beneficiary" of "hundreds of billions... in new data centers" (Motley Fool Hidden Gems Investing, Jun 9); no backlog color on GE Vernova, Eaton, or ABB.
  • Carriers / towers: AT&T's Shawn Hakl (Telco in 20, Jun 9) framed AT&T as "the connectivity layer for AI workloads via fiber, fixed wireless, and last-mile," capex tilting to fiber/edge, not macro-tower density. Quietly unhelpful for the tower-leasing thesis.
  • Freight (a warehouse-demand tell): FTR's Avery Weiss (FTR | State of Freight, Jun 9) has broker spot rates at "another all-time high... risen for 20 straight weeks," dry-van "55% higher than... the same week last year," loads "around 57% higher." Tightening freight usually leads tightening warehouse demand, a quiet positive for industrial occupancy.

What Changed

The tone, not the numbers. For a year the AI-capex story was unalloyed "they'll spend whatever it takes." This week, for the first time, insiders started pricing failure: Selitsky's 25%-don't-deliver, Narwhal's token-oversupply tell, the IG-financing-is-the-only-way framing. The demand bulls didn't blink; the financing bulls did. That's the subtle shift to watch into next quarter.