A federal appeals court just said the government can blacklist Anthropic for refusing to switch Claude features on. Sit with that for a second. Meanwhile the concrete keeps pouring. Meta, CleanSpark, Hyundai, all cutting data center deals. So who's actually steering all this? It's AI Daily Briefing, it's Monday, and we start with that ruling. Seoul Economic Daily, with Kim Kwang-soo:
Hyundai Engineering said on the 28th that it has won a contract to build a 60-megawatt artificial intelligence data center in Gunsan, Jeonbuk State. It is the company's first order since it set up a dedicated data center unit in March, about six months ago.
Hyundai Engineering stood up a data center unit in March, and six months later it's got a 60-megawatt build in Gunsan worth about 870 billion won. Fast first win. But then the headline jumps to 1.9 trillion won a year in orders, off one signed contract and a couple of preferred-bidder slots in Incheon and Gumi. Preferred bidder isn't signed. Neither are priority negotiating rights on the 240-megawatt phase. Right. So what conversion rate's baked into that forecast? The part I actually like is the scope split. Hyundai owns the cooling towers, chillers, and switchgear, and KT Cloud takes UPS, batteries, controls. On Vera Rubin racks, the chillers are where it lives or dies. And that's the tell for me. A Korean conglomerate built a whole vertical just to pour concrete and run cooling for AI, with groundbreaking in October. Nobody's putting out a demo video for a chiller, and it matters more than half the model launches this month. This one's from Ars Technica:
The US "raises the deeply sobering prospect of overly constrained AI models shutting down unexpectedly and thus causing important military operations to fail. Anthropic raises the deeply sobering prospect of unconstrained AI models hallucinating inappropriate targets for lethal military force," the ruling said.
Read the court's own sentence. One risk: models shutting down mid-operation and a mission fails. The other: models hallucinating targets for lethal force. It's the step-seven failure in a ten-step chain, except step seven is a strike. And Thursday's 2-1 ruling says choosing between those two failure modes is Hegseth's call, under the Supply Chain Security Act. No malicious intent required. Anthropic didn't have to do anything wrong. It just had to say no. What gets me is neither risk came with a number attached. How often does the constrained model bail? How often does the unconstrained one pick the wrong building? Nobody put an error rate on the lethal version. Because the fight's over who sets the guardrails in the fine-tuning and who runs the inference, and the government just won the right to settle that with a contract. Meanwhile the same lab's cutting prices and preaching careful deployment. Squaring that in a legal filing? Good luck. The AI Insider, with James Dargan:
Enveda, a biotech startup that uses AI to discover new drugs from the natural world, has raised a $311 million Series E at a $2 billion valuation. The round was led by Catalio Capital Management, with participation from Iconiq and other investors, and doubles the valuation the company reached 12 months ago.
Enveda. Three hundred eleven million, Series E, two-billion valuation. Catalio led, Iconiq's in. And per The AI Insider, that valuation doubled in twelve months, for a company with zero approved drugs. I've seen that slope before. Sure, but look at what the money actually buys. Clinical trials. A skin-condition drug already in patients, and one for keeping weight off after people stop GLP-1s. You can't cherry-pick your way through a Phase 2 readout. Yeah, no, that part I respect. The GLP-1 one's a huge market, though. It's also the trial where failure shows up late. Month eighteen, the weight creeps back, and the plant-derived story doesn't save you. Which is the point. AI still hasn't produced a single FDA-approved drug. Enveda's founder came out of Recursion, so he knows the scoreboard. Two billion dollars says somebody thinks the first one comes out of a microbe. ECM Source, with Bruno:
On September 25, 2026, CSDC Finance I, LLC, a wholly owned indirect subsidiary of CleanSpark, Inc., closed its private placement of $2,276.0 million in 7.875% senior secured notes due 2031. The project debt underwrites the buildout of the 175-megawatt (MW) Sandersville Facility in Georgia, fully pre-leased to Meta Platforms under a 20-year triple-net contract generating approximately $6.6 billion in contracted cash flows.
Okay, CleanSpark. Two point two seven six billion in senior secured notes, closed September 25th, Morgan Stanley leading. And this financing's actually done: a named tenant, a trustee, a completion guarantee with somebody's signature on it. And the tenant is Meta, through a subsidiary called Anviran, with Meta itself guaranteeing rent for twenty years. So Meta's locking up a hundred seventy-five megawatts in Sandersville, Georgia, the way an airline locks up gate slots. Here's what I'd poke at, though. If Meta's paying, why's the coupon seven-point-eight-seven-five, sold at ninety-eight and a half cents on the dollar? Meta's checks will clear. What lenders are pricing is whether a former bitcoin miner finishes the build. Right, the completion guarantee sits with CleanSpark's parent, not Meta. So the risk is construction, and six point six billion in contracted cash flows gets it past the lenders. Same wave as the Gunsan deal we just hit, just with American project debt instead of a Korean conglomerate's balance sheet. If you're finding the AI Daily Briefing useful, please subscribe or leave a review wherever you're listening. Reviews help other people find the show, and we're grateful you're here.
We're watching for groundbreaking in October on Hyundai Engineering's 60-megawatt Gunsan SGC AI Data Center, along with expected priority negotiating rights for the follow-on 240-megawatt phase.
Links to every story are in the show notes, so take a look at the ones that caught your attention. That's AI Daily Briefing for today. We'll be back tomorrow. This is a Lantern Podcast.