Meta's building enough compute to power a small country, and Zuckerberg's now wondering out loud whether to sell some of it. That's new. Here's how we got here. NAVER, NVIDIA, and Brookfield were already tied to a reported $10 billion push to take South Korea's national AI factory to 200 megawatts. The question all along was whether the financing, chips, and power would turn that announced scale into capacity people could actually deploy. Until now, 200 megawatts was the near-term benchmark for the whole project. This is AI Daily Briefing. Gigawatt campuses are going up on three continents, and one financier's name keeps showing up. A preprint dropped yesterday that quietly undercuts the whole spending spree. Let's start with Meta. From Jonathan Vanian at CNBC:
Among the four major U.S. hyperscalers, Meta is the only one that doesn't have a business selling cloud infrastructure and services even though its capital expenditures rival that of its peers. But in recent months, Zuckerberg has been floating the possibility of launching a cloud business as a way to take advantage of its capacity stockpile in a resource-constrained market.
Meta's the odd one out here. It's the only one of the four hyperscalers without a cloud business to justify capex that rivals Amazon and Microsoft. Now Zuckerberg's openly floating one on the earnings call. The honest part is Zuckerberg's line: "How much do you monetize something today versus develop future assets." Every hyperscaler has that argument in private. Zuckerberg just said it out loud. It's the same pressure that pushed Nvidia from chip supplier toward landlord. Anyone sitting on compute ends up with the same choice: sell it or hoard it. But look at the price of that dilemma: free cash flow down ninety percent year over year, the stock off seven after hours. Zuckerberg says he's getting offers at a "significant premium" over what he paid. Fine, Anthropic's reportedly in line to lease some of that capacity. But he can't sell the racks he needs to keep Llama competitive, and that's the trade-off he can't hedge. Right. He's never sold a day of cloud service, and the market just docked him eleven percent for the year while he figures out whether to start. Here's Construction Review Online:
A coalition led by Brookfield and NextEra Energy has unveiled plans to develop a $100 billion artificial intelligence data center campus at the U.S. Department of Energy’s Paducah Site in western Kentucky, transforming part of the former uranium enrichment complex into one of the largest AI infrastructure developments ever proposed in the United States.
Okay, here's the number that grounds this one: 4.6 gigawatts of new generation built exclusively for the campus—2 gigs of gas and 2.6 gigs of battery storage. NextEra owns and builds the power; Brookfield runs the compute. There's an actual site and a real generation plan, with partners attached. That's more than we've gotten from most announcements this week. And it sharpens the question I keep circling. This is private money on a Department of Energy site—a former uranium enrichment complex on federal land. So if it goes sideways in 2032, whose exposure is it, exactly? And look who's leading the coalition: Brookfield. It's the same financial firm that surfaces whenever somebody announces a continent-sized build this week. It's quietly becoming the counterparty to an awful lot of concrete. Zuckerberg just spent his earnings call agonizing over how much capacity to sell and how much to keep. Meanwhile, here's a $100 billion campus promising 1.2 gigawatts of compute by 2032. Everyone's pouring foundations against projections nobody's stress-tested. And they made a point of saying costs won't hit residential ratepayers. Which is the kind of sentence you only write when someone's already asked the question. Converge Digest writes:
NAVER, NVIDIA and Brookfield announced plans to expand Korea’s sovereign AI infrastructure by growing the initial NVIDIA DSX AI factory deployment at NAVER’s GAK Sejong hyperscale data center from 55 MW to 200 MW by 2028. The project represents the next phase of a broader roadmap in which NAVER intends to deploy up to 1 GW of AI infrastructure to support enterprises, government agencies, startups, and AI cloud customers in Korea and internationally.
Today, we get the number for NAVER and Nvidia in Korea: the 55-megawatt AI factory scales to 200 by 2028, inside a stated one-gigawatt roadmap. Nvidia's putting in a billion, Brookfield up to nine billion. Brookfield keeps popping up. It backs the Korea build and sits in the Paducah coalition we just covered. It even threads through the Nebius story coming up. One balance sheet, three continents of AI concrete. The date's the tell for me: 200 megawatts by 2028, on Vera Rubin and Blackwell systems. NAVER is standing in the exact same silicon line as everyone else announcing gigawatts this week. And the announcement came during President Lee's San Francisco visit. This has government-level sovereign AI backing, well beyond a vendor press release. Right, but Nvidia still owns the stack, from multi-tenant operations and lifecycle software to token throughput per megawatt. The sovereignty lives in the data center. HyperCLOVA X is even rebasing on Nemotron, so a lot rides on one supplier's roadmap actually shipping. Here's DataMagz:
AI cloud company Nebius has revealed additional details about its planned hyperscale data center campus in Pennsylvania, outlining a phased development that could ultimately deliver 1.2 gigawatts (GW) of computing capacity. The Nasdaq-listed company is developing the facility in Butler Township, Schuylkill County, within Highridge Business Park, approximately 55 miles (89 kilometers) west of Allentown.
So after Paducah and NAVER, here's Nebius in Butler Township. The site covers 600 acres and six single-story buildings, with the first 260 megawatts live by October 2027. Unlike the coalition deals we just walked through, this one's fully owned and operated by Nebius. No Brookfield in the cap table on this one. Which actually makes it the cleanest data point in the batch. A Nasdaq-listed cloud company is putting its own name on 1.2 gigawatts, built in phases: 260 in the first phase, 660 by 2029, then more than a gigawatt at full build. And notice what the phase-one date does. October 2027 for the first 260. That's the honest number—everything after that is a projection nobody's stress-tested against the silicon supply. That's what I've been watching all week. Everyone's announcing 2029 full-build capacity on paper. If it doesn't show up on schedule, who takes the hit? Here, at least, the answer's simple: Nebius. And this is a company that just pulled a 775 million dollar debt facility to fund the cloud side. So a lot's riding on those phases landing when they say they will. Yukiyasu Kamitani, Ken Shirakawa, writing in Arxiv:
Generative AI removes that difficulty, and an old error returns on a scale and at a speed it never had before. We locate the problem not in evidence growing weaker but in how surprise is counted.
Okay, after a week of gigawatts and hundred-billion-dollar campuses, here's the paper nobody at Paducah is reading. Kamitani and Shirakawa submitted "Phantom Evidence" to arXiv two days ago. Their claim is blunt: higher resolution and greater fluency add zero evidence. A prettier, more convincing output still has a hard ceiling on what it can prove. Letting the model grade its own homework doesn't raise that ceiling. They trace it back to Francis Bacon. Four centuries ago, he warned that looking convincing shouldn't count as evidence. Generative AI has industrialized that old error. The timing is almost too on-the-nose. We just spent this segment on Meta guiding to a hundred-thirty-plus billion in capex and Nebius lighting up 1.2 gigawatts in Pennsylvania, all justified by projections that more compute means better outputs. This preprint quietly says: measure whether the convincing answer still shows up when the target is absent. Bacon's table of absence. That's the part that keeps me up. Nobody signing a 4.6-gigawatt power deal is asking what the extra fluency actually buys in true findings. The authors say the fraction of published results that hold up drops right back to where it was before you ran anything. Building the fastest possible machine for manufacturing false positives. At scale, on schedule, with a PPL energy agreement attached. If AI Daily Briefing helps you stay informed, subscribe and leave a review wherever you're listening. It helps other people find the show, and we really appreciate the support.
We're watching for NAVER's AI agent platform launch in Korea in the second half of this year. The first 260-megawatt phase of Nebius's Pennsylvania campus is due by October 2027. Paducah's full AI campus is slated for 2032, with up to 1.8 gigawatts of utility capacity and more than 1.2 gigawatts of computing capacity.
Links to every story are in the show notes if you'd like to dig deeper.
That's AI Daily Briefing for today. This is a Lantern Podcast.