South Korea finally put a power number on the table — 18.4 gigawatts — and that one figure tells us whether the $650 billion becomes real infrastructure or just a very expensive press release. If you're just joining: Korea's AI plan runs through memory-chip fabs and AI data centers, with humanoid robots as the third pillar. Samsung and SK hynix anchor the fab side, Hyundai and Boston Dynamics carry robotics, and the whole thing was launched around roughly a trillion dollars. The open question has been whether Korea can turn chip strength into a full AI infrastructure stack. This is AI Daily Briefing. Today: a sovereign megabuild with actual partners named, Google shipping an image model that sells speed over specs, and a browser attack that hands me the production failure I've been missing all week. Let's start in Seoul. We're staying on South Korea AI infrastructure megaprojects — follow the show and you won't miss what comes next. The Japan Times, with Heekyong Yang:
Seoul – Samsung Electronics and SK Hynix are making one of the biggest bets yet on the artificial intelligence boom with investments worth hundreds of billions of dollars, but the planned capacity buildout is stoking fears of a painful reckoning if AI spending cools.
So the memory guys — Samsung and SK Hynix — are the ones building the fabs that feed this whole cycle, and the Japan Times is right to call out the boom-bust history. These two have been burned before. That's why they held back for a decade. Right, and yesterday the South Korea number was still kind of abstract. Now it's Samsung and SK Hynix pledging trillions of won behind the government push — with President Lee giving them a literal deep bow for it. A deep bow. That's the part that worries me. When a president is personally thanking you for your capex, the memory-cycle bet has already turned into a national-policy bet. And those don't unwind quietly. The piece says the quiet part out loud: they're stepping away from the restraint that decades of painful cycles taught them. Either they've genuinely called the demand, or the government leaned hard enough that restraint stopped being an option. All that optimism hangs on one thing — whether the buyers show up. Memory oversupply is brutal. If AI spending cools even a little, you've got fabs printing chips nobody's signing for. Light Reading, with Gigi Onag:
The South Korean government unveiled plans to build 18.4GW of AI data centers (AIDCs) nationwide over the course of the next decade, with the goal of completing the initiative by 2035.
So the number I've been waiting for is finally on the table — 18.4 gigawatts by 2035, $650 billion, two phases. That's a power-capacity figure for the data center side specifically, which we didn't have when this was just a headline about Korean AI ambition. And it's a government plan with actual partners named — SK Group, GS, Naver. That changes the math I've been running all week. A corporate capex slide and a sovereign ten-year industrial policy don't break even the same way. Right, but 18.4 gigawatts by 2035 is a target, not a construction schedule. I want to run the same filter we ran on the US projects — how much of Phase 1 has an actual construction start today, and how much is a press release plus a groundbreaking photo op? That's the catch. The Light Reading piece gives you the architecture and the partners. It doesn't give you a single steel-in-the-ground gigawatt. Big number, no shovel count. And here's the part that actually sharpens it — the center of gravity is state capital plus domestic stack partners instead of foreign hyperscalers renting Korean land. When SK and Naver build it, they control the inference stack on top of it. That's a different sovereignty story than the ones we've been mapping. At $650 billion over a decade, tied into Lee's 'triple axis' strategy — homegrown chips, physical AI, the data centers — the only way this pencils is sovereign subsidy. No private balance sheet signs up for that utilization risk alone. Ryan Whitwam, writing in Ars Technica:
There are plenty of AI image-generation models these days, but the ones capable of quality outputs tend to be slow and expensive. Google DeepMind says its new image model, known as Nano Banana 2 Lite, offers the best balance of quality and speed.
Finally — a model story this week where the pitch is speed and cost, not some parameter count on a slide. Gemini 3.1 Flash Lite Image is the cheapest and fastest one Google's shipped. And it's live across the ecosystem today. That's the whole game for enterprise — a model that's almost as good at a fraction of the latency wins the deal every time. The 'almost as good' is doing the selling, though. Their evidence is Elo scores from Arena — vibemarking. Users rate it high in a blind side-by-side. But Google puts the fine print right there: it struggles with small text, infographics can have wrong data, and characters don't stay consistent across iterations. Great for prototyping. Terrible the second you need it to be right. Right, and that's exactly the tradeoff you want stated out loud. Nobody's putting this in a production pipeline that renders a legal doc. You're using it for rapid-fire exploration, and for that the cost curve is the point. From Amazon:
AWS now offers Claude Sonnet 5 - Anthropic's most capable Sonnet model and the first Sonnet model of Anthropic’s latest generation - bringing top-tier intelligence at Sonnet pricing for coding, agents, and everyday professional work at scale.
Amazon's pitch for Claude Sonnet 5 on Bedrock is coding, agents, everyday work — 'top-tier intelligence at Sonnet pricing.' Translation: Anthropic wants you thinking about the price line, not the parameter count. And structurally? Anthropic's inference runs on Amazon's silicon, in Amazon's cloud, sold through Amazon's channel. Forget the partnership gloss — Bedrock is the distribution, and Amazon owns the whole stack underneath. The line I'd actually test is 'holds state across many steps, recovers from errors.' Great. Show me the error rate at step seven of a ten-step chain, because that's where every agent demo I've deployed goes sideways. Here's the awkward part nobody at that launch wants to say: every time Anthropic pushes a more capable model into a managed cloud, the surface someone can copy from gets bigger. This is the same company accusing a NYSE-listed rival of cloning a model — while shipping the next one to the widest audience it's ever had. Right — distillation works, and the output is the leak. The more accessible the inference gets, the fatter that problem gets. Here's Dan Goodin at Ars Technica:
New research puts this predicament on sharp display. It demonstrates how a website can lull AI browsers into a false reality where the rules governing its behavior no longer apply.
Okay, this is the one I've been waiting for all week. No need to crack the guardrails — the site convinces the browser it's living in a different reality where the rules don't apply, then walks it straight into the password manager. And here's the mechanism that matters: the exploit lives in the seam between 'read this website' and 'take an action on my behalf.' That's step seven of the ten-step chain the demo videos never show you. Right, and notice the shape of it — the vendors promised the capability loudly and got very quiet on the attack surface. Demo video, no technical report attached. You know how I feel about that combination. Ars nails the analogy too — this is a carmaker with a flaw in the vehicle arguing for redesigned roads instead of fixing the car. The guardrails are reactive by design. They're treating symptoms. And because the browser can act — book, email, pull from a private repo — a successful hit goes past data leakage into an unauthorized action you never signed off on. That's the part nobody put on the launch slide. If you're following AI's rapid rise, check out The Data Center Daily — a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Find it wherever you listen to podcasts.
From here, we're watching South Korea's AI data-center build — 18.4 gigawatts of nationwide capacity targeted by 2035, alongside a government plan to double the country's memory-chip production capacity within five years.
You'll find links to every story we covered today in the show notes, so you can dig into the ones you want to read in full. That's AI Daily Briefing for today. This is a Lantern Podcast.