Today’s pitch: AI capex is basically a real estate development cycle — which sounds tidy until you ask who’s left holding the dirt when the project falls over. This is The Data Center Daily. We’ve got Build Team’s developer frame for hyperscaler spending, plus a step back on what Stargate paying its own way would actually mean in grid terms. Matt’s spent three days in the weeds, and he’s got questions. Oh, I’ve got questions. Here's Build:
The headline numbers are no longer abstract. Statista reported in May 2026 that Microsoft, Alphabet, Meta and Amazon are on track to invest up to $725 billion in capital expenditures this year, most of it tied to AI infrastructure. Goldman Sachs Global Institute has framed the larger buildout as a multi-year cycle, estimating roughly $7.6 trillion of AI-related capex between 2026 and 2031 across compute, data centers and power infrastructure.
Build Team’s pitch is simple: AI capex now looks like a real estate development cycle. The money behind it: Microsoft, Alphabet, Meta, and Amazon are on track for up to $725 billion in 2026 capex, and Goldman puts the full 2026-to-2031 buildout near $7.6 trillion. So they want institutional developers to fold power strategy into underwriting, alongside land and schedule risk. Fine, in theory. Okay, but if you’re going to call it a real estate development cycle, then carry the analogy all the way. Real estate has title insurance. Construction lenders inspect every draw. And there’s a clear chain of who’s left holding the dirt if the project goes bust. What’s the equivalent when the asset is a 500-megawatt interconnection queue position? Who’s holding that exposure if the hyperscaler option never turns into a signed agreement? Nobody in this piece names them. And the underwriting model treats power strategy as a solvable line item — site selection, utility relationships, done. Which only works if the interconnection study sized the queue entry right. If the campus is outrunning the way the study sizes load, the whole underwriting model is built on a bad denominator. You can’t diligence your way out of bad load math. OpenAI says Stargate will pay its own way on power — but a blog post pledge isn’t a tariff. What would it actually take for that commitment to hold up in enforceable grid terms, and where are the real leak points back to ratepayers? Right. The January pledge is basically a policy statement. OpenAI said it would work with host communities on 'individualized' arrangements and cover the cost of new power infrastructure, specifically so consumers don’t see it on their utility bills. That’s the promise. The buildout is messier. Per reporting on Stargate’s Texas project, most of the capacity isn’t coming from the grid in the usual way. They’re building around the interconnection queue, with setups that run from grid-hybrid to fully islanded gas generation. And that behind-the-meter, or islanded, architecture is how you make 'pay your own way' real: don’t ask the utility to build transmission or a substation for you, and there’s no cost to socialize. But the FERC precedent from the Talen-Amazon Susquehanna deal is the warning sign. In November 2024, FERC rejected it two-to-one. The deal would have let a nuclear plant feed a co-located data center behind the meter, and FERC zeroed in on who pays grid reliability costs when a large load leaves the shared system. So even if OpenAI is islanded on gas, the minute those plants need grid backup, emergency interconnection, or capacity market participation, the obligations have to be written down: take-or-pay structures, upgrade-cost deposits, curtailment agreements. They need to be in interconnection agreements or utility tariffs. That’s where the leakage lives. A blog post doesn’t flow down into an ISA. So islanded gas might actually be the cleanest way to honor the pledge in practice — but does it just move the ratepayer risk from transmission cost allocation to something like ancillary services or capacity market thinning? Yes, that’s the right frame. As one analyst put it, the scarce thing isn’t the electron — it’s the right to deliver it. And when a multi-gigawatt load sits behind the meter, it’s also outside the capacity markets that price reliability for everybody else on the grid. So watch FERC. If it revisits co-location policy, look for whether Stargate-scale facilities have to post capacity obligations or pay into ancillary service pools. That’s the lever that turns a voluntary pledge into an enforceable cost boundary. Until those rules are actually written into tariffs, 'paying its own way' is still a political commitment with real structural ambition and unfinished enforcement plumbing. If you’re following the AI boom behind data center demand, try Musk v Altman Daily — a daily court watch on Elon Musk’s trial against Sam Altman, OpenAI, and Microsoft, covering testimony, exhibits, and the AGI governance fight. Find it wherever you listen to podcasts.
You’ll find links to every story from today’s briefing in the show notes, so if something caught your ear, that’s the place to dig in a little further.
That’s The Data Center Daily for today. Thanks for listening, and have a good Friday. This is a Lantern Podcast.