Daily Briefing — June 30, 2026

Samsung and SK Hynix said they will spend a combined 800 trillion won, about $518 billion, building a new chipmaking hub in the southwest of South Korea. President Lee Jae Myung stood with both companies' chairs to announce it. Each company plans to build two fabrication plants in the region, expanding well beyond their existing complexes near Seoul. Put the number in perspective: these two firms already make roughly two-thirds of the world's memory chips, and they are betting half a trillion dollars that AI demand for that memory keeps climbing.

The location was a deliberate choice. The southwest has historically lagged the rest of the country economically and has long been a political base for President Lee's party, so the plan doubles as regional development policy. Neither company gave a completion date, and SK Hynix's chair was upfront that a project this size needs vast land plus enough power, water, and skilled workers to run it, which are exactly the constraints holding up data center and fab buildouts everywhere right now.

This is the supply side of the same story that has Apple raising laptop prices. Memory makers are pouring money into capacity because AI chipmakers are buying everything they can produce. Spending on this scale tells you the companies closest to the supply chain expect the shortage to last for years, not months.

A consultant describes a new morning routine: before coffee, he checks what his AI agents finished while he slept. One night it was a 2,000-word briefing on how helium shortages threaten Asian semiconductor firms, covering a dozen companies and second-order effects. Another night, a security audit caught and fixed a small bug in his software, a customer segmentation analysis simulated 40 people reviewing a product, and his daughter's French reading practice got handled. He did none of it himself. A coordinating agent he calls RMA runs a small team, an editor, a researcher, an investor-lens reviewer, and a data specialist, with up to ten coding agents joining on busy nights.

This reads like one person's account rather than a product demo, which is what makes it useful. He describes how the unit of work shifts from "tasks I do" to "tasks I queue and review." The skill on display is delegation, quality control, and knowing which problems are worth assigning in the first place, which is closer to managing a team than using a tool.

A developer who had given up on local AI models tried Qwen 3.6 and changed his mind. The write-up argues it is the first open model you can run on your own machine that actually feels like general intelligence rather than a toy. It comes in two flavors: a mixture-of-experts version (35B A3B) and a denser, slower but stronger 27B model that the author recommends. The headline reaction on Hacker News, where this got a lot of attention, is that the 27B punches well above its size.

The catch is hardware. The author only half-jokes about his knees overheating and grabbing a thermal camera to photograph how hot his computer got running it. A capable local model is not free; it taxes your machine and your electricity bill. But it runs entirely on your own hardware, which means no API costs, no rate limits, and nothing leaving your computer.

That last part is why this matters beyond hobbyists. Everything in this week's briefings, from ChatGPT logs as courtroom evidence to governments gating model access, points the same direction: the data you send to a cloud model is not fully yours. A local model good enough for real work is the practical answer to that, and Qwen 3.6 is a sign that the open option is closing the gap fast.