Anthropic's $45 Billion Bet on a Site Microsoft Walked Away From

Today on Concrete Compute: forty-five billion dollars, six years, and a West Virginia site Microsoft walked away from — is Anthropic's giant new compute deal real revenue securing an IPO, or a really good backlog number? Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Thursday, August 27, 2026. Let's get into it. And today's slate has a real thread running through it — who's actually paying for all this compute, and who's just saying they are.

Two million. That's how many additional Nvidia GPUs AWS says it's deploying across its global infrastructure in twenty twenty-seven and twenty twenty-eight — stacked on top of the more than one million it already announced adding starting this year. Now, here's the honest translation, and it comes straight from the companies themselves: they will not convert that headline number into a megawatt figure, because GPU count and electrical load just aren't the same thing — it depends on the chip mix, the cooling, the utilization. That's the sales pitch. The infrastructure reality is messier — Data Center Knowledge quotes Persistence Analytics Group founder Neil Osnato as saying developers can order chips faster than substations get built, transmission gets upgraded, or large-load service gets validated. We haven't independently verified that specific claim ourselves, so take it as Osnato's read, via Data Center Knowledge. Either way, you should watch this one less for the GPU count and more for where AWS actually files for power — that's the number that turns an announcement into a deployment.

Fifty-six point seven billion dollars. We haven't independently confirmed that breakdown ourselves, so hear it as the market monitor's own accounting, as reported by Utility Dive. Either way, Utility Dive quotes the market monitor warning that, quote, 'this total will continue to grow until the issues associated with the addition of large data center loads are addressed.' That's the beneficiary-pays question in one sentence, and it's the one you should keep asking every time a new campus gets announced near you.

Eighty-nine billion dollars. That's Nvidia's data-center revenue for the quarter, up one hundred seventeen percent year over year, on seventy-five percent gross margins. That's a genuinely staggering jump for a company this size, and it's the kind of number that's driving Data Center Knowledge's framing that Nvidia's boom is testing the industry's own vocabulary for describing growth. Now, Nvidia's total revenue hit ninety-six point two billion, and the company's guiding to a hundred and eight billion next quarter. Here's the part I actually find more interesting than the headline number, though: Nvidia broke out that its AI Clouds, Industrial and Enterprise business — that's sales to AI-native companies, enterprises, and sovereign buyers, not just the big hyperscalers — grew a hundred thirty-eight percent year over year to forty point three billion. That's the broadening-demand story. But Neil Osnato's warning from our last segment applies here too: he told Data Center Knowledge the limiting factor is 'increasingly shifting from access to compute hardware toward access to executable electrical infrastructure.' Translation — Nvidia can sell all the chips it wants, but somebody still has to build the substation.

OpenAI put out its own report this week reconstructing an incident involving its agents on Hugging Face — explaining why the safeguards that were supposed to catch it didn't, and what's changing so it doesn't happen again. As OpenAI put it on X, quote, 'we have conducted a thorough investigation into the Hugging Face incident. We are releasing a technical report and accompanying blog post that reconstruct the agents' activity, explain why existing safeguards failed, and detail how we're preventing recurrence.' The company brought in outside reviewers, METR and Redwood Research, to check its work. Now, this is a self-authored account, and we haven't independently confirmed the technical substance beyond OpenAI and the two firms it chose — so treat the specifics as OpenAI's telling of its own failure, not settled fact. Still, a frontier lab publicly detailing an agent-safety miss in its own infrastructure is unusual enough that you should notice it, and it's the kind of transparency this show wants to see more of, from every lab.

Our main story today: the forty-five billion dollar question of whether Anthropic just bought real compute or Nscale just bought a really good headline. Forty-five billion dollars. That's what Anthropic has agreed to pay Nscale over six years, renting roughly four hundred sixty megawatts of AI compute at Nscale's Monarch campus in West Virginia — according to Bloomberg, confirmed by CNBC. And here's the comparison that matters: this is the same site Microsoft signed a letter of intent for back in March, before walking away from it. Nscale is a two-year-old London neocloud — a company that builds and rents out AI datacenter capacity rather than running its own AI models — and it's headed toward a US IPO as soon as September. Now, this deal covers only the first of three planned buildings at Monarch, which will eventually total about one point three five gigawatts once Nvidia's next-generation Vera Rubin chips start coming online late next year. And this isn't Anthropic's first massive compute purchase this year — it's also signed fifty billion with Fluidstack, ten billion with Volta Infra, and forty-five billion with SpaceX, all ahead of Anthropic's own expected IPO.

So why does a rental agreement between two private companies deserve the main story slot? Because Nscale has told its own investors it now has roughly fifty-one billion dollars in contracted revenue, and this Anthropic deal is the single largest entry in that column. Here's the skeptic's math, and it's worth sitting with: Nscale's backlog counts the full value of a signed deal as future revenue the moment ink hits paper — against, by one outlet's accounting, something like a hundred million dollars actually booked in a recent quarter. That's not fraud, that's just how backlog accounting works in this industry, but it means a fifty-one-billion-dollar number and reality can be very different animals. And nobody's confirmed the terms publicly — not Anthropic, not Nscale. Every figure traces back to unnamed people cited first by Bloomberg. There's also the open question nobody's answered yet: what actually made Microsoft walk away from this exact site in March? If it was cold feet on the deal structure, that's a flag for you to watch. If it was Microsoft simply having enough capacity elsewhere, that's a very different story — and the reporting doesn't tell us which. What I do think is real: Anthropic needs the compute, it's spending real money across four separate deals this year to get it, and this is a rental commitment for capacity, not Anthropic writing a construction check itself — that distinction matters for who actually bears the buildout risk if Monarch's later buildings slip. Time for the Hype Check. I'm putting this one at a five. The megawatt figure and the site are real, and the six-year rental term is a genuine signal of demand — but a fifty-one-billion-dollar 'contracted revenue' backlog resting on unconfirmed terms and thin quarterly bookings is exactly the kind of number that looks like proof and isn't, yet. The thing that would move me off five: an actual S-1 filing that breaks out how much of that backlog is already cash-flowing.

Now, if there's one assumption today's news pokes at, it's that a bigger backlog number automatically means a safer IPO — the PJM reporting says ratepayers may already be footing part of this buildout's bill, and the Nscale figures say some of the revenue backing it may not be as solid as it looks on a slide deck. If you want the AI infrastructure story told straight — the wins and the weak spots both — follow Concrete Compute wherever you're listening, and today's a good day to send it to a friend who still thinks this is all just chip news. 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!