All week, we've watched this move from gigawatt press releases to plans with actual names on them — and today Korea, Nvidia, and Crusoe all put real numbers on the table. If you're just joining, here's where we were: South Korea's push was already broad — a government plan for 18.4 gigawatts of nationwide AI data-center capacity by 2035, plus domestic AI chips and physical AI. SK Telecom had layered on a 140 trillion won Yeongnam plan, starting with a 100-megawatt Ulsan site and expanding toward gigawatt-scale data centers from there. This is where it gets interesting. A pension fund, a named fab, and a $30 billion valuation with no revenue attached — all in the same week. Start in Korea, because for once the target comes with a specific facility. If you want to keep up with South Korea AI infrastructure megaprojects, tap follow so the next episode lands in your feed. Lee Gyu-lee, writing in The Korea Times:
SK Telecom on Sunday announced its plan to develop up to 15 gigawatts of artificial intelligence (AI) data center capacity nationwide, aiming to position Korea as a leading AI infrastructure hub in Asia. The initiative comes as surging demand for AI model training and inference is driving an unprecedented need for high-performance computing infrastructure.
So here's the update on Korea's megaprojects: SK Telecom is now talking about up to 15 gigawatts of data-center capacity by 2035. CEO Jung Jai-hun unveiled it Friday in Jinju. Monday, this was a government gigawatt target with no company attached. Now there's a name and a date on it. Fifteen gigawatts by 2035. That's a ten-year runway to fill halls I'd love to see the contracted load behind. And they're framing it as a follow-up to the government's megaproject push, so you have the carrier volunteering to be the private balance sheet behind a national target. That's the part that gets me. All week I've been trying to keep sovereign bets separate from corporate ones to figure out who eats the utilization risk. SKT just blurred that line — a listed telco doing what an industry ministry would do. This one's from SK hynix:
He also stated, “SK hynix will invest a total of KRW 100 trillion in Cheongju—KRW 80 trillion in the M17 NAND production fab and KRW 20 trillion in the P&T7 advanced packaging facility.” He emphasized, “We will re-establish Cheongju as a key hub for driving the competitiveness of the memory semiconductor industry in Korea.”
Eighty trillion won for the M17 NAND fab, twenty for the P&T7 packaging facility. We get an actual site and product mix, with a real number attached — which, after the SKT 15GW plan we just hit, means the carrier and the memory maker are doubling down in the same country, in the same week. And that's the disclosure bar I keep wanting. Cheongju, M17, P&T7 — a specific city and specific facilities. Compare that to a certain hyperscaler's European land options with no size, no capacity, no location. This is the version you get when a company actually means it. Kwak's framing is the part that gets me — he basically says: supply is short, demand is surging, so expand. These are the same two companies that held back on memory capex for basically a decade because they got burned by the boom-bust cycle. Right, so that's the tell: the reversal. When the historically cautious ones commit a hundred trillion, the demand read is real — or they've all convinced themselves of the same thing at once. And a private company doing what a ministry would do collapses my whole sovereign-versus-corporate distinction. SK hynix has moved past waiting for a subsidy — at this point, they're carrying the industrial policy themselves. Here's Canada Newswire:
TORONTO, July 3, 2026 /CNW/ - Canada Pension Plan Investment Board (CPP Investments) announced today it will invest US$1.75 billion (C$2.4 billion) to support EQT's strategy to build AI Infrastructure, led by global data centre developer and operator EdgeConneX.
So CPP Investments — the Canada Pension Plan — is putting US$1.75 billion into EQT's AI infrastructure strategy, flowing through EdgeConneX. That's retirees' money contracted into data-center buildout. And that changes the shape of the week for me. All week, the compute layer's been telcos and hyperscalers — now institutional pension capital is locking into that same tier. Third capital type in the pool. Right, and CPP's own guy — Max Biagosch — calls it "durable, long-term demand." Durable. On a data center platform that leans on hyperscale customer relationships they don't name. EdgeConneX is in 20-plus countries and was founded in 2009, so it's not a paper company. But the number I want is the contracted book against the capital going in. Here's Niklas Berg at LavX News:
The first two named operators give the proposal real scale. Sharon AI, an Australia-based sovereign AI cloud provider that started in 2024, plans to deploy up to 40,000 Nvidia Grace Blackwell GB300 GPUs. Firmus plans up to 170,000 Nvidia GPUs at a 360-megawatt AI factory campus in Batam, Indonesia, with a design based on Nvidia's DSX platform.
Okay, so Nvidia's revenue-share model has a name and a number now — 210,000 GPUs across Sharon AI and Firmus. Sharon takes up to 40,000 GB300s, Firmus goes for 170,000 at a 360-megawatt campus in Batam. Nvidia collects the hardware sale up front, then takes a cut of the cloud revenue on top. Great deal — for Nvidia. But if a partner's utilization drops below breakeven, that revenue slice just... stops? And notice what's missing. No financing terms, no lender structure, no share rate, no duration. Nvidia won't even say whether it's lending off its own balance sheet or brokering through banks and private credit. That's the tell. Announce the GPU count and skip the credit structure, and you're putting out the flattering half. All week, the buildout story has been moving from press-release gigawatts to named facilities — this is the financing layer of that move, and it's the part they'd rather we didn't scrutinize. Here's SiliconANGLE:
Data center builder Crusoe Inc. is reportedly in talks to raise a $3 billion funding round. Bloomberg on Thursday cited sources as saying that the deal could value the company at $30 billion. That’s about three times what Crusoe was worth last year. The report didn’t name the investors that may participate in the new raise, but late-stage rounds often draw contributions from existing backers.
Crusoe, three billion at a thirty billion valuation, per Bloomberg. That's triple what they were worth last year, on a total of two-point-seven-seven billion raised to date. And the number I can't find anywhere in that report? Revenue. Twelve months, three times the price — what changed operationally, not narratively? Here's what changed: two of eight buildings on the Abilene Stargate site are live, and Meta signed for two more in Texas and Missouri. That moves it from a deck to poured concrete with tenants. Fine — but they build the halls. The demand risk sits with OpenAI and Meta filling them. Crusoe gets repriced on customers who are themselves unproven at that scale. And that's the week in one number. Pension money, a named fab, and now a tripled private valuation — the infrastructure tier is repricing in real time, and none of it comes off a leaderboard. If AI Daily Briefing is part of your routine, try The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. 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 attention, you can follow it there and read further.
That’s AI Daily Briefing for today. This is a Lantern Podcast.