SK hynix's $26.5B ADR sale is now the largest foreign IPO in US history

Good morning 👋 A South Korean memory chipmaker just pulled off the biggest foreign IPO Wall Street has ever seen, and the 7x oversubscription tells you exactly what investors think next year's real AI bottleneck is.

In today's issue:

• 🔭 SK hynix's $26.5B ADR sale is now the largest foreign IPO in US history
• 🧠 Mistral ships prompt version control, Meta opens an agentic API, Anthropic scales chip validation
• 🔬 Memory agents, research swarms, and a warning about swapping your LLM judge
• 🏛️ Brussels finds Meta's feeds addictive by design
• 🛠️ Audit your MCP tool definitions before they tank your agent
• 🔗 Bernanke joins Anthropic's trust, OpenAI's bio bug bounty, and more

🔭 THE ONE THING

💾 SK hynix's $26.5B ADR sale is now the largest foreign IPO in US history

SK hynix priced 177.9 million ADRs at $149 apiece, raising $26.5 billion in an offering that came in more than 7x oversubscribed. That tops Alibaba's roughly $25 billion listing from 2014, making it the biggest US market debut ever by a non-US company, and second globally only to SpaceX's Nasdaq listing last month. The demand isn't mysterious: SK hynix is the HBM supplier Nvidia's whole downstream depends on, and global capital has had almost no clean way to buy that exposure until now. Pepperstone's Dilin Wu put it plainly: "the AI memory cycle is real, the earnings are real." Read the 7x cross the other way, too: it's the market betting HBM supply, not GPU supply, is next year's actual bottleneck.

🧠 MODELS & RELEASES

• 📝 Mistral AI gave Studio's prompts and skills real version control: immutable versions, rollback, audit trails, and lineage that traces a production output back to the exact asset behind it. If you've been managing prompts in a shared doc, this is the upgrade you didn't know you needed. Mistral AI
• 🎨 Meta Superintelligence Labs put Muse Spark 1.1 behind a public Meta Model API preview, OpenAI-compatible, with a million-token context and parallel tool calling. Early partners are calling it a full agentic foundation, not a rebadged image model. Meta AI
• 🏭 Anthropic and UST are pointing Claude Code at chip schematics to auto-generate regression tests, and the case study claims validation cycles that used to take four days now close in 48 hours. UST is also training 20,000 of its own people on Claude, which tells you where this partnership is headed next. Anthropic
• ⚙️ AWS shipped five HyperPod inference upgrades in one pass: multi-tier data capture, direct Hugging Face Hub deploys, NVMe-based cold starts, automated Route 53 DNS, and pod-level IAM. The NVMe change is the one that matters if you're running scale-to-zero endpoints. AWS
• 🪞 Anthropic also launched a Reflect dashboard that breaks down how you actually use Claude, plus quiet-hours controls if you'd rather not know. Handy for auditing your own habits. It's also Anthropic training you toward the usage patterns it wants to see. Anthropic

🔬 RESEARCH HIGHLIGHTS

• Proactive memory agent splits remembering from acting: a separate agent watches a long trajectory and decides what's worth resurfacing and when, instead of hoping the agent doing the work notices before it's buried in context or pushed out entirely. For anyone building multi-step coding or browsing agents, this is the wall you actually hit around step 50, not model quality. Remember When It Matters

• WebSwarm drops the single-agent, single-trajectory approach to deep research. It recursively spins up sub-agent teams to cover a search space in parallel, then folds the results back up the tree. A lone ReAct loop runs out of context and patience on genuinely deep-and-wide queries. A swarm doesn't. WebSwarm

• LLM judge swap Researchers held candidate responses fixed across four judgment datasets and only changed which model graded them. The scores moved anyway, no answers touched. If your eval pipeline shows a quality jump after a model upgrade, check whether you upgraded the judge too before you believe it. When the Judge Changes

🏛️ POLICY & REGULATION

• European Commission preliminarily found Meta in breach of the Digital Services Act over Instagram and Facebook's addictive design. The Commission's beef: infinite scroll, autoplay, push notifications, and the recommender systems, plus a finding that Meta's own risk assessments underrated the harm to minors and vulnerable users. Meta gets to respond before anything's final, but this is the DSA's addictive-design theory getting its first real test case against a platform this size. Builder read: if you're shipping recommender-driven feeds into the EU, the compliance bar just got a concrete definition. Commission notice

🛠️ TRY THIS

Run a token audit on your MCP tool definitions. AWS's ML team broke down why agents get dumber the more tools you hand them: descriptions eat context on every call whether or not the model needs them, and bloated context makes the model pick the wrong tool, which triggers retries, which bloats context further. (AWS ML)

1. Pull every tool description your agent loads by default and count the tokens. Cut anything that isn't load-bearing for the first call.

2. Rename parameters to match how the model actually talks about the domain, not your database schema. resource_class, not content_bucket.

3. Cap parameters at eight or fewer per tool. Past that, split into narrower tools instead of one do-everything call.

4. Replace vague errors ("no results") with actionable ones ("search requires 2 or more terms"). Anthropic's own tests found trimming responses to essential fields cut response tokens by roughly two-thirds.

Prompt: List every MCP tool definition loaded in this session. For each one, estimate its description length in tokens, flag any with more than 8 parameters, and flag any error messages that don't tell the model what to do next.

🔗 QUICK LINKS

• Anthropic put Ben Bernanke, the Nobel-winning former Fed chair, on its Long-Term Benefit Trust, the outside body meant to keep the mission honest. Anthropic
• OpenAI published the rules for its GPT-5.5 bio bug bounty, paying out to anyone who can get the model to help with biological harm. OpenAI
• Hugging Face ran part 3 of its PyTorch profiling series, on why Flash attention wins despite occupancy numbers that look wrong at 13%. Hugging Face
• NVIDIA cast national AI build-outs as a strategic race, spotlighting sovereign AI pushes underway from France to India. NVIDIA
• Netflix started putting video datasets and models on Hugging Face. Small move, but media companies don't usually open the vault. Clem Delangue

That 7x oversubscription number won't feel unusual for long.

Pradeep Perugu