A presidential memo orders AI into live weapons systems, a Chinese lab wants thirty billion dollars, and Nvidia signs another national-scale cloud deal — the hard-power week showed up on a Tuesday. This is AI Daily Briefing. Cassidy, this was supposed to be a spending week; it turned into policy and money, and honestly the numbers got worse. They did. We've got NSPM-11, Moonshot's valuation math, SK Telecom and Nvidia, and Newsom's worker order — start with the memo, because it lands on top of everything we tracked last week. Yeah, let's do it. The directive's doing the editorializing for us today. NSPM-11 bars vendors from disabling AI systems warfighters depend on. Read that again — that's a procurement clause, and it slams straight into every responsible-deployment policy OpenAI and Anthropic have ever published. Right, because all week I've been asking what happens when a chain breaks midway through. Here's the government version — except the broken step isn't a failed API call, it's a weapon that won't stand down. And nobody's saying that tension out loud. A lab can write 'we reserve the right to shut this off' all it wants — the federal directive says you don't get to. Moonshot. Four billion in December, thirty billion now, and an ARR that topped two hundred million in April. That's a 150x revenue multiple. And yet I'm almost grateful — that two hundred million is the first actual denominator anyone's offered in the big-valuation conversation all week. Sure, there's a number under it. But 150x earns the same red pen I'd use on any American leaderboard claim — the IPO-window timing explains the rush, it doesn't excuse the multiple. Agreed. A revenue figure isn't a defense, it's just the first time I could even do the division. Then SK Telecom and Nvidia — an infrastructure alliance, first AI factory in Korea next year. It's the Naver story again, just with a new actor making the same pitch. 'Next year.' That goes in the announced-not-running column with everything else we logged this week. Third deal in seven days, same Asia's-leading-cloud language. At this point, that sounds like a sales script more than a coincidence. And Newsom's order is the only one on the page pricing in the downside — severance standards, employment insurance. Every founder I know models the upside spreadsheet; Sacramento just opened the other one. Those two ledgers don't reconcile, and that gap's going to be the next fight. Quietly, though, there's a real npj paper today, an agent pipeline turning raw audio into structure for medical LLMs. Actual work that survived actual data, while everyone else was signing memos. A technical report on a hard-power Tuesday. We'll take it. More after the break. Here's The Next Web:
President Donald Trump signed National Security Presidential Memorandum 11 on Friday, directing the US military and intelligence agencies to accelerate their adoption of advanced AI while protecting frontier models from theft by foreign adversaries. The directive replaces the Biden administration’s NSM-25, which had governed AI in national security since 2024, and adds a provision that no commercial vendor can disable, degrade, or modify an AI system that American warfighters depend on without prior government approval.
NSPM-11. Trump signs it Friday, it replaces Biden's NSM-25, and buried in there is the line everyone's skating past — no vendor can disable, degrade, or modify a system warfighters depend on without prior government approval. That's a procurement clause wearing a safety badge. And it runs straight into every responsible-deployment policy OpenAI and Anthropic have published — the kill-switch they keep promising they'd pull? Now they need a permission slip. Right, and think about what that clause does — it turns production reliability into a national security requirement. I've spent all week asking what happens when a multi-step AI chain fails mid-task. Except here the failure isn't a dropped API call. It's a live weapons system, and the vendor who built it legally can't hit the brakes without Washington signing off first. 120 days to rewrite procurement for that. The Next Web writes:
Moonshot AI, the Beijing-based developer of the Kimi chatbot, is seeking as much as $2 billion in a new funding round that would value the company at $30 billion, Bloomberg reported on Sunday. If the round closes at that target, it would mark a seven-fold increase in capitalisation since December, when the startup was valued at just over $4 billion.
Moonshot — $4 billion in December, $30 billion now. That's a 7x markup in six months on the Kimi chatbot. And yet — here's the part I actually want to sit with — they put a revenue number on the table. $200 million ARR as of April. All week I've been staring at valuations with no denominator. This one's got one. $200 million ARR against a $30 billion ask. That's a 150x multiple, Bill. I'd put the same red pen on that as I would on any leaderboard chart out of an American lab. The timing tell is the IPO window. DeepSeek's raising at up to $59 billion, Zhipu's pegged near $80 billion — everybody's sprinting to get capitalized before the listing doors open. The numbers are racing each other, not the fundamentals. Look, 150x is eye-watering, no argument. But a real ARR figure beats the pre-revenue hand-waving I've watched all week. At least Moonshot's multiple has a floor under it. Here's Roh Kyungjo at The Asia Business Daily:
SK Telecom is joining forces with Nvidia to build a gigawatt (GW)-scale artificial intelligence (AI) factory. The goal is to establish an AI ecosystem and meet the rapidly growing demand for AI infrastructure.
So Chey Tae-won and Jensen Huang stood at the SK Seorin Building yesterday and announced a gigawatt-scale AI factory. And the same phrase that's been following Nvidia around all week — national competitiveness — gets another turn. Here's the number I want: the headline is gigawatt, but the deliverable is one AI factory launching in Korea next year. So what's the phase-one number with actual hardware behind it? Right — "next year" is the tell. After the Naver deal opened at fifty-five megawatts while everybody yelled gigawatts, this goes straight into the announced-but-not-drawing-power column. What's actually moved past press-briefing stage is the DSX full-stack tie-in and the SK hynix memory relationship underneath it. That part's real. The gigawatt framing on top is the sales deck. And it's a sequel, right? Naver-Nvidia three days ago made the national-competitiveness pitch, now SKT makes the same move with a new chairman at the podium. Third data point this week means the pattern's pretty hard to miss. This one's from Insurance Journal:
California Gov. Gavin Newsom issued an executive order to attempt to deal with the impacts of artificial intelligence on workers and businesses. The order directs the state to explore policies that include severance standards, employment insurance and transition support for displaced workers, worker ownership models, universal basic capital concepts, expanded workforce training and better tracking of hiring and payroll trends to help respond to potential layoffs and economic disruption.
Newsom's order has a line in it I've never seen in an AI policy doc — severance standards. Actual compensation rules for displaced workers, plus employment insurance, and this 'universal basic capital' idea. Every founder I talk to models the upside of AI adoption. Sacramento just wrote a memo about the cost side. Those two spreadsheets do not reconcile yet. And notice the timing, Bill — this lands the same day as NSPM-11, which we just hit. One arm of government is rushing AI into live systems on a deadline, another's drafting a safety net for the people it displaces. Worker ownership models, equity in severance packages — that's interesting on paper. It's also a stack of agencies told to 'explore' and 'evaluate.' I'd like a date attached to any of it. 'Create an AI playbook to modernize job training.' That's a lot of verbs and not one displaced worker getting a check yet. npj Digital Medicine writes:
Here, we present an agent-based transcription framework that autonomously converts raw unstructured conversation transcriptions (RUCT) into structured conversation transcriptions (SCT) suitable for LLM fine-tuning. The system integrates three coordinated modules—Planner, Memory, and Executor—to orchestrate noise removal, content correction, speaker identification, and dialogue segmentation within a self-correcting workflow.
Okay, this one's actually my kind of paper. 7197 minutes of Chinese clinical recordings, eight departments, and a three-module agent — Planner, Memory, Executor — that turns garbage audio into something you can fine-tune on. And here's why I trust it more than the demo reels: they published the failure surface. Denoising 94.7, content correction 96.9, but speaker identification drops to 88.6. That's the number I'd be sweating in a clinic. Right — 88.6 on speaker ID means roughly one in nine times the system isn't sure who said the symptom and who said the diagnosis. In a doctor-patient transcript, that's not a rounding error. Exactly. And the 3.6x-faster-than-manual line is nice, but notice the English test was 240 minutes against 7197 in Chinese. That belongs in the portability footnotes — and to their credit, they treat it that way. What I appreciate is npj Digital Medicine ran controlled comparisons against a non-agent sequential pipeline. The agent framing usually gets a free pass; here they actually made it earn the orchestration overhead. Have feedback on today’s briefing, an AI story we should watch, or a correction? Send us a note anytime at aidailybriefing at lantern podcasts dot com. We read what you send.
You’ll find links to all the stories we covered today in the show notes, so if something caught your ear, you can dig into the original reporting and sources there.
That’s AI Daily Briefing for today. This is a Lantern Podcast.