Notes From a Scattered Week

Some weeks arrange themselves around a thesis. This one didn't. What follows is the pile as it accumulated, in roughly the order the tabs refused to close.

Money first, because the money got strange. The reported arrangement in which Nvidia would guarantee $250 billion of OpenAI's Ohio data center debt is the kind of structure that makes you reread the sentence twice, and the interesting part is what the guarantee reportedly excludes rather than what it covers. Sitting next to it, a piece on Shein's Hong Kong IPO prospectus makes the opposite argument in miniature: the headline number is the accounting artifact, and the number nobody's quoting is the one that would actually price the thing.

Then there's the question of who benefits when compute gets scarce. The case that Japan, South Korea and Taiwan have become AI export economies is the most cheerful set of macro numbers in the region in decades and also, read carefully, one of the more uncomfortable ones — because a single export category carrying that much of the growth is a concentration risk wearing a growth costume. Meanwhile Washington has found a new thing to worry about, and the word is distillation, which arrived in policy circles roughly eighteen months after it arrived in engineering ones.

Defense capital keeps moving in the same direction. The primes have now participated in a record $4.1 billion of venture rounds, concentrated in drones, interceptors, and the software that ties them together — categories where cost per unit and time to production matter more than exquisite capability. A different flavor of the same trend shows up in the argument that Palantir is winning a share fight rather than ending one, which gets the mechanism right and, usefully, refuses the more dramatic version of the claim.

On the water, the counting continues to disagree with the commentary. Bab al-Mandab transits fell to eleven vessels after strikes on Jizan and Yanbu, with Hormuz holding below ten across the same weekend. If you want to do that counting yourself rather than wait for someone else's summary, the practical guide to analyzing AIS, satellite imagery and port activity is the place to start, and it is refreshingly honest about where the data lies to you.

Two on tradecraft to finish. The argument that cheap visual manipulation doesn't need to persuade anyone to work is the sharpest thing I read all week — the capability ceiling barely moved, the floor collapsed, and the effect is corrosion rather than deception. And a contrarian defense of the worst source in the business: conferences are a terrible OSINT channel and still worth attending, for reasons that have nothing to do with the programming.

Ten tabs down. The pile regenerates by Tuesday.