Fifteen billion for one campus, with Google's credit rating quietly co-signing the loan. Today, the compute buildout comes with a harder question: who's exposed? If you're just joining us, Meta's AI buildout now has a capacity-allocation problem on top of the model problem. The company has been weighing whether to sell some excess compute through a cloud-style business while keeping enough headroom for its own models. Anthropic has reportedly held preliminary talks to lease that compute, which could turn Meta's capex-heavy buildout into an outside infrastructure revenue line. This is the AI Daily Briefing. Amazon's at $220 billion while Meta has a quarter-trillion in leases. One utility CEO is already admitting his grid is oversubscribed. Stick around—we're tracing who's actually on the hook. We're staying with this story: Meta AI compute monetization dilemma. Follow the show and you won't miss what comes next. Ashley Capoot, Hugh Son, writing in CNBC:
Morgan Stanley is the lead banker engaging in advanced talks to lend $15 billion to Nexus Data Centers to build out a large artificial intelligence infrastructure project for Anthropic, CNBC confirmed on Thursday. Nexus Data Centers will use the capital to develop a data center campus in Hubbard, Texas, according to a source familiar with the discussions who asked not to be named because the details are confidential.
The key detail in this $15 billion Nexus deal is Google's agreement to backstop Anthropic with its own investment-grade credit rating. Morgan Stanley is lending against Google's balance sheet, with Anthropic's name on the door. Right. Anthropic doesn't have investment-grade paper on its own. You want to build a campus in Hubbard, Texas, you need someone whose credit the bond desk actually trusts. And that's a different kind of dependency from a chip order or a deployment deal. Google already committed up to $40 billion to Anthropic in April. Now it's lending its rating too. Whatever the press releases say about independence, they're structurally stitched together. Nvidia guaranteed OpenAI's lease last week. Now Google's guaranteeing Anthropic's credit. The guarantor's becoming a standard part of the stack, same as the GPUs. So I keep coming back to this: if demand for these campuses comes in soft, Google's holding that Texas exposure. That's a lot of faith riding on a rating for a campus that hasn't been poured yet. From Converge Digest:
SEATTLE — Amazon Web Services delivered its strongest quarter in more than four years, with revenue surging 37% year-over-year to $42.2 billion—its fastest growth in 18 quarters—as demand for AI infrastructure accelerated across cloud computing, custom silicon, and foundation model services.
AWS AI has cleared a $25 billion run rate. Custom silicon—Trainium and Graviton—has cleared another $25 billion, and both are growing at triple digits. That's real revenue, not a demo. I'll give Jassy that one. And then he funds it with a $220 billion capex plan. The revenue's real. But nobody has independently verified the demand curve behind that plan. Right, and at a $25 billion run rate, Jassy's still saying capacity won't meet demand. Wednesday's BEACON preprint asks whether long-horizon agent workloads justify pouring in compute at that scale. Today's number only makes that question louder. The multi-gigawatt Trainium commitments from Anthropic and OpenAI tell you where this is headed. These deals are being measured in power now: gigawatts. And we just heard Nexus scrambling for $15 billion on the Anthropic side. Everybody's writing checks against megawatts that aren't interconnected yet. A serverless product built specifically for persistent agentic workloads is a very optimistic thing to ship in the same quarter that a paper says those chains fall apart at step seven. Natalie Musumeci, writing in Business Insider:
Meta has more than a quarter trillion dollars in future lease obligations — largely tied to AI data centers— as the tech giant races to build out its artificial intelligence infrastructure, a new regulatory filing shows. In its quarterly securities filing released a day after its earnings report, Meta said it had approximately $278.99 billion in operating and finance leases that have not yet begun and were not yet included on its balance sheet.
So here's the update on Meta's compute buildout: the lease tab is finally visible. About $279 billion in future lease obligations, and it hasn't hit the balance sheet yet. The quarterly securities filing puts it at $278.99 billion, one day after earnings. You treat a filed number differently from something a CFO says out loud on a call. Right, and those leases start rolling in from now through 2036. That's a decade-long commitment that outlasts basically every product cycle these leases are supposed to serve. And it lands the same morning Amazon guides to $220 billion in capex. In the same 24 hours, two of the industry's biggest infrastructure bets are now in filings, with no independent check on the demand curve behind either one. Sit with the phrase off-balance-sheet. The commitment is real, and it runs through 2036. You just don't have to stare at it in the same place every quarter. From Converge Digest:
AMD and Core Scientific announced a strategic infrastructure partnership under which AMD will secure access to more than 500 megawatts of U.S. AI-ready data center capacity beginning in 2027, with the potential to expand to as much as 2.5 gigawatts over time.
So AMD's deal here is for power, not chips: more than 500 megawatts starting in 2027, scaling to 2.5 gigawatts, locked in with Core Scientific before a single rack goes in. Core Scientific can bring already-connected land to the table, thanks to converted crypto-mining sites. That interconnection queue everyone else is stuck in? They skipped it. That tells you exactly where the pain is. When a chip company signs a capacity deal this size, the bottleneck has moved from silicon to substations. The warrant structure is the part I'd underline: AMD gets market-priced warrants to buy Core Scientific stock. That gives AMD an equity hook into its own landlord alongside the colo contract. It echoes the Google-Anthropic backstop we talked about earlier: wire the counterparty into your balance sheet so the dependency runs both ways. A 2027 start date, with 2.5 gigawatts coming "over time." Nobody's independently priced that demand yet, but at least AMD's hedging with equity instead of simply eating the lease. Here's Kristi E. Swartz at Atlanta Journal-Constitution:
The company had inked deals for roughly 8,000 megawatts of electricity out of the 10,000 megawatts that the Georgia Public Service Commission approved last year before OpenAI’s project was announced. “ … Our OpenAI contract in Georgia pushes us beyond our recently approved capacity by right around 1 gigawatt,” said Chris Womack, chairman, president and chief executive of Southern. A gigawatt is 1,000 megawatts.
Less than a year ago, Georgia's commission blessed a doubling of the grid—10,000 megawatts—specifically for data centers. On Thursday's analyst call, Womack said it was already gone, and OpenAI's Savannah project pushes them a full gigawatt past that. Twenty billion dollars, 3,200 megawatts—that's three Vogtle reactors' worth of power for one campus. Next to the interconnection queue, capital looks easy. Here's the part no earnings call names: they'd already inked deals for 8,000 of the 10,000 approved megawatts before OpenAI even showed up. The Amazon and Meta capex numbers from earlier are bets on a demand curve. Womack's the operator saying the curve is outrunning the grid—out loud, on the record. They've signed 17,000 megawatts of large-load deals through the mid-2030s, and sales are up 55% in a year. To cover the gap, they're proposing 2,000 to 6,000 megawatts of dispatchable capacity by 2032. The answer for today's demand is a maybe, seven years out. That sends another expansion back to the same commission that just approved this one. Every one of these campuses is now waiting on a utility docket rather than a chip shipment. That's where the exposure sits if the projections miss. If AI Daily Briefing is part of your routine, check out Anthropic Pentagon Watch, a daily briefing on Anthropic’s fight with the DoD, military AI use, autonomous weapons, and AI procurement blacklisting. Find it wherever you listen to podcasts.
One thing we’re watching: AMD’s reserved access to more than 500 megawatts of AI-ready U.S. capacity from Core Scientific is due to begin in 2027.
You’ll find links to every story in today’s show notes if you want to dig into anything that caught your attention.
That’s AI Daily Briefing for today. This is a Lantern Podcast.