Anthropic's Half-Trillion Dollar Question

Today on Concrete Compute: five hundred seventeen billion dollars. That's the price tag one outlet has put on every compute deal Anthropic has signed in less than a year — and the question we're chasing is whether any company's revenue, even one growing as fast as this one, can ever catch up to a number that size. Before we get there, in the headlines: ERCOT, the operator of the Texas grid, says it will finish Governor Abbott's audit of nearly three hundred data center projects by December — a deadline the grid operator itself admits could still slip. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Tuesday, September 8, 2026. Let's get into it.

Let's start in Texas, where the grid operator ERCOT says it's on track to finish an audit Governor Greg Abbott ordered back in early August. DCD reports ERCOT intends to complete its review of roughly three hundred planned data center projects seeking a grid connection by December tenth. That audit is the gate — until it's done, ERCOT's paused 'Batch Zero' interconnection study, the queue that decides who actually gets wired into the grid, can't resume. And here's why that queue matters so much: DCD reports ERCOT is currently sitting on more than four hundred seventy-four gigawatts of connection requests, over five times the state's record peak demand, and about ninety percent of that is data centers. Abbott's audit requires every one of those projects to disclose its power and water use, its community impact, and who actually owns and finances it, before it can move forward. DCD reports the delay could hold up as much as twenty percent of the entire US data center pipeline, and cost projects roughly eight billion dollars by the first quarter of 2027 if it drags on, with nearly fifty gigawatts of projects caught in the freeze. Now, ERCOT's own senior vice president, Chad Seely, says requests for information will keep going out through October and November, which tells you December tenth is a target, not a promise — ERCOT has already told Texas regulators it expects to miss an April 2027 deadline for the Batch Zero results themselves. That date matters. If it slips again, you're looking at real money and real megawatts sitting in limbo while a state that wants to be the center of this buildout waits on its own paperwork.

Our main story today: the compute bill nobody's revenue can pay yet. Five hundred seventeen billion dollars — that's the value DCD reports The Information has assigned to every compute contract Anthropic has signed since October of last year. And for comparison, that's nearly triple the one hundred eighty billion dollars Anthropic told its own investors, just last December, it expected to spend on renting servers through twenty twenty-nine. Sit with that gap for a second — what the company told investors and what it's actually signed up for moved by hundreds of billions of dollars in under a year. DCD reports the total works out to fourteen point eight gigawatts of compute capacity — gigawatts here meaning the raw electrical power needed to run all those chips — that Anthropic has locked in access to since October, on top of an estimated one to two gigawatts it had already secured before that, according to an unnamed source cited in the reporting. Now, none of this is exactly a surprise in isolation. DCD reports Anthropic signed a forty-five billion dollar compute capacity agreement with Nscale toward the end of August, on top of more than a dozen letters of intent for data center leases with US developers, a one hundred ninety-one megawatt agreement with Riot Platforms for a Texas site, and a two hundred billion dollar deal with Google for TPUs — Google's own AI chips — that Anthropic could deploy itself. Beyond leasing data center space outright, the company has also gone shopping among cloud providers: DCD reports deals with Akamai, with Amazon Web Services including a large cluster of AWS's custom Trainium chips, with CoreWeave, and a fifty billion dollar partnership with the AI cloud firm Fluidstack. The list keeps going — SpaceX and xAI, Google Cloud, Lambda, Volta Infra, AMD, and Microsoft all show up as counterparties. Of all that capacity, DCD reports Google and AWS combined will supply eleven gigawatts of it — by far the largest share. And this is landing at a specific moment: Anthropic confidentially filed for an initial public offering with the SEC back in June, meaning these commitments are about to get read by public market analysts instead of just private investors and their own boards. The company's revenue has been moving fast too — its annualized revenue reportedly topped sixty-five billion dollars, ahead of OpenAI's reported forty billion dollars plus as of July. But even at that pace, sixty-five billion in revenue sitting against five hundred seventeen billion in contracted spend is a gap that doesn't close on its own.

So how do you actually think about a number like that? Start with what the five hundred seventeen billion dollar figure is and isn't. It's not something Anthropic disclosed — it's DCD reporting on The Information's own reconstruction, built by adding up public announcements and additional reporting, and it spans contracts of wildly different lengths, some running years past twenty thirty. That's not a single bill coming due on one date; it's more like a stack of leases with different maturities, some of which the company could renegotiate, delay, or walk away from before they ever get expensive. It's also worth being precise about what fourteen point eight gigawatts actually means. That's contracted access to capacity — a legal right to use power and chips someday — not fourteen point eight gigawatts of servers humming today. The gap between signing a compute deal and having it energized is exactly where the execution risk lives, and it's a gap neither Anthropic nor, frankly, OpenAI can currently cover out of revenue alone. So who eats the difference if AI demand growth even just slows down, rather than stops? That's the real question sitting underneath this story, and it's one the reporting doesn't answer, because nobody involved — not Anthropic, not Google, not AWS — has published what happens to these contracts if the revenue curve bends. My own read, and I'll own this as opinion: a lot of that five hundred seventeen billion is doing double duty. Some meaningful slice of it is very likely capacity that Google and AWS had already built or already planned to build as part of their own capex, that Anthropic is now claiming credit for locking in — which would mean the headline number overstates how much genuinely NEW construction this is triggering, even though the dollar commitment is real. That's a distinction worth holding onto, because it changes who's actually taking the risk. If Google is building a facility it would have built anyway and just leasing slices of it to Anthropic, Google's balance sheet carries that risk regardless of what Anthropic's revenue does. If instead these are purpose-built, dedicated sites financed against Anthropic's specific promise to pay, then Anthropic — and by extension whoever's IPO prospectus this ends up in — is the one holding the bag if the growth curve disappoints. The reporting doesn't tell us the split, and that's the single most important number missing from this story. There's also a timing wrinkle that I think matters more than it's getting credit for: this is landing right as Anthropic has confidentially filed to go public. Once that filing becomes real and public, five hundred seventeen billion dollars in disclosed and reconstructed obligations is exactly the kind of figure that turns from a trade-press headline into a line item institutional investors will fight over in the prospectus. Private backers can take a growth story on faith for a while longer. Public markets tend to ask sharper questions about contracted liabilities versus contracted revenue, and that scrutiny is coming whether Anthropic's growth holds up or not. Time for the Hype Check. I'm putting this one at a five. The dollar figure and the gigawatt figure are both real signs of an extraordinary, genuine buildout — Anthropic's revenue growth is legitimately fast, and these are largely named, reportable deals with real counterparties, not vague MOUs. But the five hundred seventeen billion number is an outside analyst's reconstruction wearing the clothes of a disclosed liability, it mixes contract lengths that aren't directly comparable, and it tells us almost nothing about who's actually exposed if the revenue curve doesn't keep pace. Real steel, real deals — but a headline number doing a lot of unearned work.

Nobody's published what happens to five hundred seventeen billion dollars in contracts if AI revenue growth ever actually slows down — and until somebody does, that's the assumption this entire buildout is quietly resting on. If you want the daily version of that question asked story by story, follow Concrete Compute wherever you're listening. This has been Concrete Compute, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!