I didn't expect the most interesting robotics deal I read about this month to be a fish-killing machine, but here we are. @ShinkeiSystems builds a robot called Poseidon that sits on fishing boats. When a fish comes out of the water, the machine scans it, finds the brain, and pierces it within seconds, before stress hormones ruin the meat. It's the Japanese ike jime technique, except no craftsman is involved and it runs at sea, at scale. Shelf life goes from 5-7 days to 12-14. The company has cooked fish three weeks out of the water. Restaurants holding a combined 50 Michelin stars are already buying, and Japan is even importing American-caught fish now, which I'm told has basically never happened outside of tuna. What convinced me this is a real business, though, is the supply chain math underneath. A lot of fish caught by American boats today gets frozen, shipped to China for gutting and filleting, then shipped back across the Pacific to be sold in US stores. Roughly 18% of American-caught fish spoils between dock and store. Shinkei's actual bet: kill at sea, then process and distribute everything out of one 16,000 sq ft plant in Tacoma, and beat that round trip on cost. (The detail I also keep thinking about is from the @foundersfund partner @zebulgar who did the deal: AI and defense, the two most consensus trades in venture right now, are only 15 to 20% of Founders Fund's deployed capital. I would have guessed at least double. He also said the fund keeps its exposure to crowded trades low on purpose.)

Two highlights worth relooking from this talk by @xiaopenghexpeng, CEO of @XPENG_Global: 1/ His team has been building robots for 7 years, and nobody ever wanted to touch or hug one (0:25 in the clip). Steel reads as equipment, maybe dangerous. Skin is what flips it into a companion. Everyone races on torque and joint count, but whether people want the thing near them is a different problem, and his answer starts from skin. 2/ On "why not build the tool into the robot" (1:10): then it can only ever use one tool. Humans already made a million tools, all shaped for human hands. He wants the robot that picks them up. @adcock_brett made the same argument for Figure, one platform amortized across millions of tasks.

My guess: if we count meaningful pure-play AI companies (labs, infra, vertical apps at $1B+ valuations), somewhere around 40–60 will IPO globally by mid-2029. Here’s the reasoning. The floodgates just opened. Cerebras opened 89% above its IPO price on May 14, and both OpenAI and Anthropic confirmed confidential S-1 filings in June 2026. Behind them, Databricks, Mistral, Cohere and xAI are tracking for 2026–2028 windows, plus Lambda and a long tail of vertical AI (legal, medical, coding). Morgan Stanley expects 10 to 15 major tech companies to go public in 2026 alone, mostly AI-related. Run that rate for three years with a growing pipeline and you get to 40+ without heroic assumptions. Two wildcards. One: the China side, the STAR Market and HKEX robotics/AI queue could add another 15–20 on its own if the window stays open. Two: whether the first mega-listings hold up. If the early ones pop, the window opens for the rest; a stumble resets pricing expectations across the entire pipeline. If OpenAI or Anthropic trades badly in year one, my number drops by half.