The compute race just stopped being about models. Today it's power, cloud capacity, and who's allowed to ship a model across a border. If you're just joining: South Korea's AI buildout stopped being just a chip-policy story a while back. It's now a planned 18.4-gigawatt nationwide data-center buildout by 2035, tied to domestic AI chips and physical AI — with public-private projects pulling in SK Group, GS Group, Naver, and the semiconductor incumbents. This is the AI Daily Briefing. Today — a utility buying its way into data-center power, a Korean pledge you have to read twice, and the day a government gate froze a model release. Let's start where the money actually has skin in the game. South Korea AI infrastructure megaprojects isn't over. Follow us wherever you're listening, and the next chapter comes to you. PR Newswire writes:
Rising demand from data centers and other power-intensive industries is driving the need for dedicated, scalable power — often outpacing the timelines of traditional grid connections. This investment is consistent with National Grid's ambition to pursue selective critical infrastructure growth opportunities emerging from the transforming energy landscape, and marks NGV's entry into one of the fastest-growing areas of electricity demand.
National Grid Ventures is paying $1.75 billion for 35% of Joulent LLC. This is the first infrastructure line this week I can actually map — an actual company, a real stake, a real dollar figure, not another capex slide. And it's a regulated utility arm taking a direct equity position in data-center power contracting. A utility is deciding large-load power delivery is its next core business, the kind of deal it doesn't just book and forget. Right, but 35% of what? Yujnovich's quote leans hard on 'contracted critical infrastructure' and 'predictable cash flows' — which is code for 'we want signed customer load before we pour concrete.' So the question I've got is what Joulent's contracted book actually looks like today. Thirty-five percent of nothing is still nothing. Chris James basically admits the model in his own quote — 'independent company designed for speed, scale and execution.' Independent is the tell. The play is outside capital locking in the power layer underneath the hyperscalers. This one's from Ars Technica:
The US has lifted export curbs on Anthropic’s newest Claude models, Fable 5 and Mythos 5, about three weeks after the Trump administration flagged the models as national security risks. As of today, Anthropic confirmed in a blog post, Fable 5 will be available globally, and US organizations have had access restored to Mythos 5 since June 26.
So the export curbs on Anthropic's Fable 5 and Mythos 5 came off yesterday — three weeks after the same administration flagged both as national security risks. Fable 5 goes global today, and US orgs got Mythos 5 back on June 26th. And Lutnick's letter to Anthropic says the risks were addressed 'in close coordination with the US government.' So there was a government review gate, it tripped, and then it cleared. Operationally, that's new. That's the part I'm actually chewing on. If you're an enterprise that built a deployment plan on Fable 5 and it can vanish behind an export flag for three weeks — availability now depends on geopolitics as much as it depends on pricing. You can hedge on cost. You can't hedge on 'Commerce froze my model on a Tuesday.' And notice the asymmetry — Fable 5 is now the widest-distributed Claude they've ever shipped, while Mythos stays gated internationally except through Glasswing, which is just cyber researchers at trusted shops doing defensive work. Right, so 'global release' has an asterisk. Fable's open to the world, Mythos is a members-only club with a government approval line at the door. From Allen Lee at TechTimes:
SK Group Chairman Chey Tae-won laid out one of the boldest pieces of South Korea's June 29 mega-project announcement: roughly 1,000 trillion won (about $654 billion) for AI data centers and some 1,100 trillion won ($719 billion) to expand semiconductor supply — including about 400 trillion won ($261 billion) for a new chip cluster in the southwest.
Okay, so this is the update on Korea — SK is finally putting dollar figures on the buildout. And they're big enough to break my usual sense of scale: $654 billion for data centers, $261 billion for a new chip cluster in the southwest. And now we can do the math we couldn't do yesterday. We already had Korea's 18.4-gigawatt power target — now we've got Chey's compute number to stack against it. Right, and the nearly $1.2 trillion total is two-thirds of the country's entire GDP, per TechTimes. At that point, I stop treating it like a normal utilization bet. No company puts two-thirds of GDP on the line. A country does. And Chey's framing tells you why. He wants Korea to move from consuming AI to exporting intelligence — he calls the data center an 'intelligence factory.' That phrasing is the pitch. It lets him sell $654 billion and $261 billion as one plan instead of two. That's the part that makes me nervous, honestly. Peak-profit capacity bets are exactly the ones that age worst — and this is the peak-profit moment for the memory cycle, dressed up as national strategy. Rebecca Bellan, writing in TechCrunch:
Meta has spent billions of dollars developing AI and building out data centers to support it. But now, the company may be preparing to put those data centers to a more immediately profitable purpose. On Wednesday, Bloomberg reported that Meta is developing plans for a cloud infrastructure business, selling access to both AI compute power and models.
Meta selling surplus compute — Bloomberg had it Wednesday, and the stock jumped more than 10%, about $149 billion in market cap in a day. Wall Street basically just found out what the $125 to $145 billion capex year was for. And that lands right on CoreWeave and Nebius — the GPU-rental layer. If Meta dumps excess capacity at a price that undercuts them, that's a deflationary event for the whole neocloud sector. This is the hyperscaler gravity well doing what it always does — the middle tier gets collapsed from above. SpaceX did the same thing weeks ago through xAI, buying out Colossus 1 for Anthropic. Here's my question though — Meta selling surplus assumes there's a stable surplus to sell. That capacity exists because Meta over-provisioned for its own training runs. The day Llama's next generation needs it back, the outside customer is the one who gets bumped. Which is a very different SLA than AWS is offering. 'Cloud, but only until we need it.' Right — that's the inference cost curve story hiding inside a cloud revenue headline. Cheaper compute is great, until the pricing that made it cheap was structural excess that evaporates. This one's from TelecomTV:
The agreement covers up to 170,000 NVIDIA AI accelerators across Grace-Blackwell, Vera-Rubin, and Vera platforms through 2027 and 2028, positioning the Batam campus among the largest AI infrastructure developments in Asia-Pacific. It marks substantial growth in Firmus' deployed AI Factory capacity in the region.
Finally — a number I can actually count. 170,000 GPUs, 360 megawatts, a contract running through 2034. That's the closest thing to a shovel-count anyone's put on the table this week. And it's Batam, Indonesia — not Virginia, not Korea. Firmus is dropping the campus where the power's available, not where the demand narrative is loudest. But read the structure, Cass. NVIDIA earns standard product revenue and a share of Firmus's cloud revenue on the supported capacity. NVIDIA's selling the picks and taking a cut of the gold. Revenue-sharing plus credit support. So NVIDIA's underwriting the demand for its own chips. That's the vendor financing the whole neocloud tier runs on now. And here's what nags me — 'a pace that outruns the capital market's ability to support their contracting.' Translation: the AI-natives can't finance this themselves, so NVIDIA's backstopping the credit. Circular, much? Stack it against the National Grid piece we just hit — a utility taking equity in power contracting, NVIDIA taking a revenue share in compute. Everybody's trying to own more than one layer at once. If you're following the power struggles shaping AI, 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 we covered today in the show notes, so if one caught your ear, you can head there to read more. That’s AI Daily Briefing for today. This is a Lantern Podcast.