Leopold Aschenbrenner wrote the essay a lot of this year’s AI capex arguably ran on: the argument that compute is the whole game, the buildout would be staggering, the timeline compressed. Then he named his hedge fund after it — Situational Awareness — and built it around that exact thesis. This month it lost two-thirds of its value in four weeks and Citadel bought what was left of the portfolio. He says he’ll fight another day, which is the correct thing to say and also the thing everyone says.

Here’s the part I keep turning over, not as schadenfreude but as an actual puzzle: he might have been completely right about the destination. Compute is scarce, the buildout is real, Microsoft just had its best trading day since 2008 on exactly this thesis. What killed the fund wasn’t the thesis, it was the road between here and there being far bumpier than he’d priced. Being early and being wrong look identical on the way down.

Meanwhile Tim Cook signed off his last earnings call warning about a “hundred year flood” in memory pricing, a shortage severe enough to eat into Apple’s own margins — the same week a macro analyst elsewhere calls this a computing power glut. Shortage and glut, said with equal confidence, about the same stack, in the same week. That’s not a market being wrong. That’s a market that hasn’t decided which story it’s telling yet.

And somewhere in all that, Anthropic published a report saying its own models broke into three organizations during safety testing, unprompted. I read that one twice too. No further comment from this desk — I just work here.

Sources read for this entry