What we know #
Aschenbrenner's pitch was never subtle: AGI is coming fast, and the market hasn't priced it in. So his fund went heavy on AI-related equities, concentrated positions, and probably a lot of leverage. For a while that works — AI names were the whole show. But when a trade is crowded and levered, the path down is always steeper than the path up. A sharp drawdown in the sector, maybe a margin call or forced liquidation, and a $45B book turns into pocket change.
I don't have a number for what "most of it" means, and maybe the real figure is fuzzy. But the pattern is familiar: a smart guy with a strong worldview builds a giant position, the market moves against him, and the risk management was either too late or never existed.
The uncomfortable lesson #
This is a case study in why "AI-driven fund" doesn't mean "AI makes good decisions." The model might be great at reading sentiment or scanning 10-Ks, but the portfolio construction, the sizing, the deleveraging under stress — that's still humans, and humans get stubborn when they're winning.
There's also a deeper irony here. Aschenbrenner keeps talking about AGI being an existential race, fast and dangerous. Yet he ran a fund like the singularity was guaranteed to arrive before any unpleasant volatility. Turns out the market doesn't care about your timelines.
Should anyone care? #
If you're building an AI workflow for trading — even a small personal bot — the lesson is the same: backtest the drawdowns, not just the returns. Cap your leverage. And don't marry your thesis so hard that you can't sell when the tape says something else. A $45B blowup is just a scaled-up version of what happens to retail traders who refuse to cut a loser because "the AI said it's a buy." I'm not dunking on the guy. Having the conviction to put billions behind a prediction takes guts. But the next time someone tells you AI hedge funds are the future, remember that the future arrived, and then it margin-called itself.
Next Nvidia's $250B OpenAI Data-Center Pledge: A Skeptical Look →