Good morning. It's Friday, June 26th.

**Story one.** OpenAI has unveiled its first custom silicon — a chip called Jalapeño, built in partnership with Broadcom. It's purpose-designed for large language model inference in data centers. Not a general-purpose GPU that's been tweaked for AI — Jalapeño is a ground-up blank-slate architecture based on OpenAI's knowledge of how LLMs actually run. The company emphasized significantly lower inference costs compared to GPU clusters, especially for real-time coding workloads. This puts OpenAI in the same arena as Google, Amazon, Meta, and Anthropic — all racing to break free from NVIDIA's pricing power. The chip is already in production with TSMC, and OpenAI says it will begin deploying Jalapeño at scale before the end of the year. The AI lab that rented the most GPUs is now building the silicon to replace them.

**Story two.** SpaceX is building its own energy infrastructure in South Texas. The company announced plans to construct Starpipe — an eight-mile natural gas pipeline connecting to its Starbase launch complex in Boca Chica. The pipeline will terminate at a liquefaction facility that converts natural gas into liquid methane, which is what Starship's Raptor engines run on. The Texas Railroad Commission filing shows the project would enter service by January 2027. This is significant: right now Starship launches rely on trucked-in liquid methane, which is a bottleneck for cadence. A dedicated pipeline means SpaceX can fuel Starships on-site, rapidly. It's the difference between an experimental test site and an operational launch facility. The company is literally building its own gas station — because when you plan for millions of tonnes to orbit per year, you don't wait for trucks.

**Story three.** OpenAI published research yesterday showing that non-developer adoption of its Codex coding agents has grown one hundred and thirty-seven-fold over the past ten months. Legal, finance, recruiting, and research teams at OpenAI have all made Codex their primary AI tool. By May, more than eighty percent of individual users had delegated at least one task that would take a human thirty minutes or longer. A quarter of users are now assigning eight-hour-plus jobs to agents. At OpenAI itself, ninety-eight percent of employees use Codex, and it accounts for more than eighty-five percent of output tokens across all departments. The 99th percentile users are running over sixty hours of agent work per day across multiple parallel agents. The headline isn't that developers love AI coding tools anymore. The headline is that lawyers and recruiters are using them to build software.

**Story four.** Rocket Lab pulled off a remarkable demonstration this week. On June 19th, the U.S. Space Force gave Rocket Lab a launch order for the Victus Haze mission — and the Electron rocket was in orbit just sixteen hours and forty-two minutes later. That's a new record, beating the previous mark of twenty-seven hours set by Firefly Aerospace in 2023 by more than ten hours. The satellite went operational in under thirty-eight hours, commissioning well ahead of a 72-hour deadline. Victus Haze is designed for close-proximity operations with another satellite already in orbit. The message: commercial launch providers can now deliver military spacecraft to orbit on less than a day's notice. Space on demand is no longer a concept — it's a standing capability.

That's all for today. Have a great Friday.