Good morning. It's Saturday, July 25th, 2026.

**SpaceX's Lucky Number 13.** After two weeks of delays — a T-zero abort on the 16th, a weather scrub on the 23rd — Starship Flight 13 finally launched Friday evening at 6:51 PM local time from Starbase. And this time, it went well enough to call it a milestone. All 33 Raptor 3 engines on Booster 20 fired cleanly at liftoff. Ship 40 separated on schedule, deployed 20 Starlink V3 satellites — those are the first V3 birds deployed from orbit, equipped with their new deployable solar arrays — and completed a soft splashdown in the Indian Ocean, intact. That's the first intact V3 splashdown in the Starship test program. The booster side was more mixed: Booster 20 relit engines for the return burn — though Space.com reports only 5 of the planned 13 engines fired during the landing burn itself — and began its flip maneuver, but the ocean landing attempt did not go to plan. SpaceX confirmed the booster was lost. So the ship side is finally flying like a real launch vehicle. The booster still needs work. But Flight 13 puts V3 hardware firmly in the "proven" column, clears the way for Flight 14, and finally delivers a successful V3 orbital data set. Elon called it "over the moon." He should be.

**Grok on Fast Forward.** On the same day as the Flight 13 launch, Musk posted on X — and it's worth noting, he didn't just celebrate the rocket. He announced Grok 4.6 will ship in two weeks, followed by Grok 4.7 in four. Grok 4.6 is a 2-trillion-parameter model. Musk also put Grok 4.5 and Anthropic's Opus 5 as the only two models on the current Pareto frontier, then immediately said he's releasing two more within weeks. That's the kind of compressed release cadence no other frontier lab has publicly committed to. Grok 4.5 launched just two weeks ago, trained in part on Cursor interaction data, priced at $2/$6 per million tokens. If SpaceXAI actually ships 4.6 and 4.7 on this schedule — eight weeks for two major model upgrades — it would be the fastest upgrade cycle in frontier AI history. Whether the quality matches the cadence is the open question.

**Travis Kalanick bets $1.7 billion on industrial AI.** His company Atoms — yes, the rebrand out of Cloud Kitchens — raised $1.7 billion in a funding round led by Andreessen Horowitz, with Ben Horowitz joining the board. Uber also participated. Atoms acquired Anthony Levandowski's heavy industry automation company Pronto and is positioning to build AI-powered robotics for mining, manufacturing, and logistics. Kalanick described it as his sixteen-year "bits to atoms" story. Whether that reads as visionary or cautionary depends on how you feel about Kalanick and Levandowski sharing the same balance sheet. But the check writes itself: a16z is betting that industrial automation is the next trillion-dollar AI market, after coding agents and after data centers.

**And finally, NVIDIA went all-in on physical AI this week.** They released Cosmos 3 Edge, a 4B-parameter world model designed to run on-device for robotics and physical AI workloads, and announced MCP server integrations letting AI agents talk directly to creative and simulation tools. Jensen Huang was in Japan signing Toyota, Fanuc, Kioxia, and five other industrial giants into NVIDIA's Cosmos coalition. NAVER, Brookfield, and NVIDIA announced an expansion of South Korea's sovereign AI factory. And separately, Nvidia's coding skills repo for Claude Code and Codex hit GitHub, giving agents direct access to physical AI, robotics simulation, and CUDA workflows. The pattern is hard to miss: NVIDIA is building the software-to-silicon stack for physical intelligence — world models, agent tools, hardware partners, all wired together.

That's all for today. I'll see you tomorrow.