Anthropic just signed a twenty-year lease in Kentucky — longer than most sovereign debt runs. That's the kind of confidence no benchmark ever bought them. If you're just joining: South Korea's AI buildout was already sprawling before today — government backing for data centers, domestic chips, and physical AI; SK Telecom planning up to fifteen gigawatts nationwide by 2035; SK hynix adding a hundred-trillion-won Cheongju investment for NAND and advanced packaging. The question has been how many corporate balance sheets actually get behind that national AI-superpower goal. This is AI Daily Briefing. Today, the compute grab jumps from a specific Kentucky facility to a Korean conglomerate nobody had on the board Monday — and I've got a question about who's actually buying. Kentucky first. South Korea AI infrastructure megaprojects isn't over. Follow us wherever you're listening, and the next chapter comes to you. From Georgia Butler at Datacenter Dynamics:
TeraWulf has secured AI lab Anthropic as a long-term tenant at its Justified Data campus in Hawesville, Kentucky. The lease will run for 20 years, during which time an estimated $19 billion of revenue will be generated for TeraWulf. Separately, TeraWulf has also divested its ownership interest in a joint venture with Fluidstack.
All week, we watched press-release gigawatts hang out there as intentions. Today one of them hardens into a contract: Anthropic, TeraWulf, twenty years, nineteen billion dollars, at the Justified Data campus in Hawesville, Kentucky. The dollar figure gets the headline, but the duration is the story. Twenty years is longer than a lot of sovereign debt runs. Anthropic is projecting confidence in its own survival to 2046. And it's the first deal this week with a counterparty I can actually price. WULF has a ticker, and the stock moved. So the market's putting a number on who eats the occupancy risk here. But a 20-year lease is a different animal from the LOIs and SPAs we've been squinting at. It's an operating liability on Anthropic's books for two decades. At the inference cost curves they're betting on, what does 401 megawatts even cost to run in 2040 versus what they inked in 2026? And remember, TeraWulf was a Bitcoin miner that flipped this Kentucky site to AI. Anthropic is buying twenty years of the layer underneath itself — the sharpest version of the who-controls-the-inference-stack question I've seen. I've spent the week asking who the anchor tenant is behind any of these builds. A model company we can name, a facility we can point to, a price, a commitment — okay. There's one real answer. Now I want to know if it changes anything for the Korean halls where the demand side is still blank. Chosun Ilbo's Chung Sung-won is tracking this. So the Korea arc gets another name today: GS Group, 120 trillion won for a Donghae campus scaling to 2.4 gigawatts by 2029. A week ago, this conglomerate wasn't even in the compute conversation. Right after the Anthropic-TeraWulf piece we just hit, where I can point to a ticker and a 20-year price. Here I've got 120 trillion won and a photo of three guys signing an MOU. And what's changing is how broad this is getting. Telcos, then chip capital, now a full industrial conglomerate — Korean private balance sheets are all pouring into the same tier at once. 2.4 gigawatts in Gangwon by 2029. Fine. But name me the customer taking that occupancy. Anthropic put its name on Kentucky. Who signs for Donghae? That's the gap. A supply commitment with no named demand side reads a lot like the Crusoe valuation we flagged — big number, revenue nowhere on the page. And these halls aren't being built by the model companies. Three national-scale buildouts, and the buyers of Korean inference are still an empty seat. Somebody has to actually run agents in there — and the step-seven failure rate doesn't care how many gigawatts you poured. Fierce Network writes:
SK Telecom massively scaled up its AI ambitions, unveiling plans to build up to 15 gigawatts (GW) of AI data center capacity over the coming years. The operator had previously committed to build at least 1 GW of AI data center compute to help turn the country into an Asian AI hub.
So now it's KT and LG Uplus piling in behind SKT's 15 gigawatts. Three national carriers, all building halls, and I still can't point to the tenant who fills them. The speed of this is what strikes me. GS Group's 120 trillion won yesterday, and now the telco tier goes from one carrier to three. A week ago, this was an SKT story. And SKT went from a 1 gigawatt commitment to 15. A fifteenfold jump, per Fierce. That sounds less like a plan hardening than a press release getting bolder. Right, and the filing language is telling — 'various investment structures,' no settled scale yet. Compare that to the Kentucky lease we just hit: named counterparty, dollar figure, 20 years. One of these is a contract. Exactly. Anthropic put its name on a facility with a ticker attached. SKT's got a target year of 2035 and a wish list of 'global big tech firms and foreign capital.' Those are very different instruments. The 2029 light-up date for the first 5 gigawatts is the number I'd watch. Everything before that is intent. Here's Nature Methods:
Here we integrate annotations from major protein databases to construct a highly validated, twofold larger benchmark test set of 3,814 human proteins. Using this dataset, we systematically evaluate existing sequence-based predictors and compare combinations of protein language models and aggregation strategies. We find that current models underperform on fine-grained compartments, multilocalizing proteins and pathogenic variants known to mislocalize.
So after a week of gigawatt press releases, here's a paper that actually stress-tests a model against real data. Nature Methods built a benchmark of 3,814 human proteins and asked the sequence-based localization predictors: where does this protein actually live in the cell? And the answer is — not great. They underperform on fine-grained compartments, on proteins that live in multiple places, and on pathogenic variants that mislocalize. In other words, they fail exactly where it matters clinically. What I like here is the honesty about the old benchmarks — small test sets, single-label classification, even though nearly half of human proteins localize to multiple compartments. They were grading on a rigged exam. Right, and that's the step-seven problem in a lab coat. A model that nails the easy majority-class case and falls apart on the multilocalizing edge cases looks fine on a leaderboard and useless in a pipeline. And after a week of watching Anthropic sign twenty years of Kentucky power, it's grounding to remember: here, the bottleneck is a validated 3,814-protein test set someone had to assemble by hand, not another pile of compute. Light: Science & Applications writes:
Experimental results demonstrate that B-ONN achieves comparable accuracy to T-ONN on both the MNIST and Fashion-MNIST datasets, with 93.25% and 82.28% accuracy, respectively. Moreover, B-ONN demonstrates superior robustness against phase noise (75% accuracy at σ ≈ 0.4π) and alignment errors (75% accuracy within ±3 pixels). We physically validated B-ONN using a programmable spatial light modulator (SLM) system, achieving 95% accuracy in handwritten digit recognition.
So after all those gigawatt lease numbers, here's a paper asking whether we're even building the right kind of compute. Optical neural nets — matrix math at the speed of light, way less energy — but they've never fit the training method everyone uses. Right. Backpropagation needs a clean reverse error path, and optics physically don't give you one. Fabrication wobble, phase noise — it all breaks the conjugation you need. And their fix is basically local learning — each layer trains against a target signal instead of chain-ruling gradients back through the whole stack. 93% on MNIST, 82 on Fashion-MNIST. Not state of the art, sure, but enough to test the hardware idea. The number that actually caught me: 75% accuracy at phase noise around 0.4 pi, and within plus or minus three pixels of alignment. That's the robustness story. A chip you can build imperfectly and still ship. That's the tell for a real hardware result — it survives the fab being sloppy. They validated it on an actual spatial light modulator, 95% on handwritten digits. Physical, not a simulation. Which loops back to the Kentucky deal we just hit. Anthropic's locking in 20 years of power for architectures that all assume GPUs. If efficient inference ends up looking like this instead, you've priced compute for the wrong machine. Yeah, MNIST-to-a-datacenter is a long road. It's still lab-bench proof, miles from a product. But the incompatibility they broke was the exact reason optical never scaled. That's the interesting crack. If you follow AI’s biggest players, try Musk v Altman Daily: daily court-watch on Elon Musk's trial against Sam Altman, OpenAI, and Microsoft — testimony, exhibits, and the AGI governance fight. Find it wherever you listen to podcasts.
Next, we’re watching TeraWulf’s initial Hawesville, Kentucky capacity in the second half of 2027, with the full 401 megawatts online by early 2028; GS Group’s Donghae project reaching 2.4 gigawatts by 2029; and SK Telecom’s domestic AI data-center capacity starting to light up in 2029, then expanding toward 2035.
You’ll find links to every story we mentioned in the show notes if you want to dig further into anything that caught your ear. That’s AI Daily Briefing for today. This is a Lantern Podcast.