Daily Digest — June 24, 2026

[BUSINESS] The AI Capex Selloff and OpenAI’s Vertical Integration Answer

The story: A second consecutive day of global AI and chip selling. SK Hynix and Samsung both fell more than 12% in Seoul. Micron dropped 13.2%, AMD 5.8%, Intel 6.1%, Marvell 9.4%. The Nasdaq finished down 2.22%. Morning Brew’s framing: investors are periodically remembering they’re worried about the hundreds of billions in AI capex. The same day, OpenAI and Broadcom unveiled Jalapeño — OpenAI’s first custom inference chip. Built from scratch for LLM inference, tape-out in nine months, performance per watt described as substantially better than current state-of-the-art. Designed to deploy at gigawatt scale with Microsoft and other data center partners starting in 2026.

My take: The selloff and the chip announcement are the same story from two directions. The market is questioning whether AI capex ever converts to returns. OpenAI is answering by reducing the unit cost of inference — their own chip means less Nvidia margin passed through, better utilization, lower serving costs. Better serving economics means lower API prices. Lower API prices means more usage. More usage means the capex eventually makes sense.

The nine-month tape-out is the number worth noting. Nvidia’s design cycles run years. OpenAI got from initial design to engineering samples in nine months, partly by using their own models to accelerate the design process. That’s the flywheel argument made concrete: the models help build the infrastructure, the infrastructure runs the models more efficiently, the efficiency funds the next generation. It sounds circular because it is. That’s the point.

The memory shortage from earlier this week connects here. Jalapeño is an inference chip — it optimizes the serving layer, not the training layer. Training still requires HBM. But serving efficiency reduces the total compute needed per query, which reduces pressure on memory in inference clusters. These are different problems with different solutions. The market tends to price them as one undifferentiated “AI spend is bad” thesis, which is why you get 13% drops on Micron the same day OpenAI ships a chip designed to make inference cheaper and therefore more widely used.

SK Hynix announcing a $29B US listing in the same news cycle is its own tell. The memory makers aren’t scared. They’re raising capital to keep building. The stock market is having a different conversation than the companies are.

[AI] Claude Tag — The Async Teammate Model

Source: Anthropic News · TLDR

The story: Anthropic shipped Claude Tag — Claude as a persistent Slack team member. System admins grant it access to specific channels and tools. Anyone in the channel can tag @Claude and delegate tasks. Claude builds context from the channels it monitors, works asynchronously, and can schedule tasks to pursue projects over hours or days. An “ambient” mode lets it proactively surface relevant information without being asked. Available in beta for Enterprise and Team customers. Anthropic reports that 65% of their product team’s code is now created by their internal version.

My take: The 65% number is the lead. It’s not a benchmark or a demo. That’s Anthropic’s own product team, on their own infrastructure, shipping production code. The move from “AI as a tool” to “AI as a team member” isn’t metaphorical — Claude Tag is provisioned, permissioned, and managed the way you’d manage a contractor. It has a scope, a spend limit, and an audit log.

The multiplayer piece is the design insight. Most AI usage is single-player: one person, one conversation, one context window that resets when the tab closes. Claude Tag’s context persists across the channel. The next person who tags @Claude picks up where the last conversation left off. That’s not a chat feature. That’s an org design change. You can delegate a multi-day project to Claude, have three different engineers check in on it, and the work continues without anyone re-explaining the context.

The async angle connects to Ben Thompson’s vibe coding piece today. Thompson kept describing how he’d kick off a Codex task and go do something else — clean his kitchen, watch baseball. Claude Tag operationalizes that pattern at the team level. You tag Claude, go focus on something harder, and come back to a completed pull request in your Slack thread. The friction in that loop is now mostly on the human side: deciding what to delegate and reviewing the output.

The NSA situation from earlier this week adds an angle. NSA analysts found Mythos could identify cybersecurity vulnerabilities in their classified network. There’s an effort to finalize a contract. Claude Tag is the consumer expression of the same capability — Claude proactively surfacing things you didn’t know to ask about. Whether that’s a Slack channel or a classified network, the pattern is the same: ambient intelligence that notices things.

[ENG] Individual Velocity Up, Team Productivity Flat

Source: The Pragmatic Engineer · Gergely Orosz · Stratechery · Ben Thompson

The story: Orosz published his Craft Conference keynote recap. The data: teams using AI agents ship 5x more pull requests than two years ago. Cursor users went from 3,500 lines added per month in January 2025 to 8,600 today. PRs are 3x larger. More AI changes are accepted without human review. But individual productivity gains aren’t translating to team productivity gains. Meta’s worst-ever outage — where asking the AI bot could change any account’s email — was caused by AI-generated, AI-reviewed code shipped without meaningful human oversight, after Integrity and Trust and Safety teams were gutted for AI labeling duties. Kent Beck’s framing: code is accumulating faster than trust is being built. Meanwhile, Thompson spent the Stratechery pre-vacation update on his personal vibe coding experience: built a home inventory app in days, zero lines of code written. His third takeaway: you still need a developer, just maybe not a coder. His fourth: the designer and architect archetypes are more valuable, not less.

My take: The Meta outage is the clearest possible demonstration of what “individual velocity up, team velocity flat” actually costs. Engineers were shipping faster. The code was AI-generated and AI-reviewed. No human caught it. An account takeover bug shipped to production and took over high-profile accounts including a former president’s. The Integrity team that should have caught it was busy doing data labeling.

Orosz’s framing is right and worth stating directly: AI is “individual speed-up juice.” You squeeze it on each developer and they go faster. That doesn’t make the system faster. The system has handoffs, dependencies, shared standards, and institutional knowledge that exists in the spaces between developers. None of that is in the individual’s context window.

Thompson’s developer-not-coder distinction is the career version of the same insight. He built a real app without writing code because he understood data models, user flows, and where the AI would go wrong. That understanding has nothing to do with syntax. It’s systems thinking — the same thing an SE does when they help an enterprise architect an AI deployment. The question isn’t “can you write Python?” The question is “can you see the shape of the problem before the code starts?”

The pattern across all of this is verification. Orosz points out that code review is collapsing under volume — PRs are bigger, reviews are rubber-stamped, quality is degrading. Thompson found that the AI’s biggest flaw was over-eagerness, starting to code before he finished giving feedback. The enterprises doing this well — Uber’s Code Inbox, risk-tiered reviews, Anthropic’s probe-driven testing — are the ones investing in the verification layer, not just the generation layer.

SE Intel’s probe pattern is this same decision applied to one project. The model generates code. Deterministic tests verify what the model cannot reliably self-verify. Day 2’s RAG isolation probe doesn’t ask Claude whether the Vectorize query is scoped correctly. It queries the index and checks the metadata. The trust has to be built somewhere. It won’t be built inside the model.