Leopold Aschenbrenner’s Situational Awareness - the AI-focused fund built off the essay that named it - sold the bulk of its public equity holdings to Ken Griffin’s Citadel after July’s momentum rout and pressure from prime brokers. Millennium and Jane Street were also in the conversation. The private book, including a large Anthropic stake, was not part of that sale.

That’s not a story about artificial intelligence dying. It’s a story about process failing under leverage.

The rise

Aschenbrenner left OpenAI in 2024, published Situational Awareness, and turned a thesis about compute, power, and near-term AGI into a hedge fund.

Seed capital was small by industry standards… backers included names that open doors in both Silicon Valley and trading.

Returns through June 2026 were reported in the stratosphere.

One letter cited 439% net for the first half of the year.

Assets were reported in the tens of billions at the peak, with some accounts putting the top far higher still before July turned.

The market treated him as an oracle. Positions in AI infrastructure, memory and neo-energy became a crowded map. When one young fund with a clear narrative and heavy size moves, others watch what’s going on…

But…

Understanding a technology is not the same as surviving a pressurised margin cycle.

What a margin call actually is

You borrow against positions with assets to increase size.

Brokers lend so you can run more exposure than your equity alone would allow. When prices fall, the collateral shrinks. Brokers demand more cash or securities. If you cannot post it, they force sales.

Forced sales hit the same names the rest of the market is long. Prices fall further. More margin is required. The loop is mechanical and it does not care how correct the long term essay was, and this is what we have seen recently in Korea too, which we noted in the first edition of Fink Signal.

AI hardware and infrastructure names collapsed in July, and short positions in software that moved the wrong way (again, Tim’s view from the last Fink Signal was long software, short semis, which worked exceedingly well this week) caused pressure from both sides of the book.

Leverage figures in the press have been high some saying 4x, others saying as high as 16x (wtf?). Exact terms are private, the movement in stock prices, well, not so much.

Citadel did not believe in AI in a romantic sense by buying the book.

Multi-strat balance sheets absorb inventory when someone else cannot.

The brokers get paid. The crowded tape finds a bid as we saw yesterday.

The forced seller loses the right to manage that risk on their own timetable.

How this hits other funds

One large, levered, thematic book does not need to be Archegos to matter.

When a concentrated AI long is sold in size, dealers and other funds holding the same names mark down.

Risk systems cut exposure and volatility-targeting strategies reduce gross exposure when realised vol rises - they sell into weakness by design, not by any form of opinion.

CTA and risk-parity flows can lean the same way. Correlation across AI infrastructure names drifts toward one. Liquid winners get sold because they still have a bid.

So the damage is not only Aschenbrenner’s PnL.

It is the second-order flow… margin, vol control, and crowding reinforcing each other. That is the same mechanism we discussed on correlation spikes in a quick email on Wednesday (check your inbox) - forced sellers, not a clean capitulation signal.

Other managers who ran a softer version of the same book feel the air go out even if they never met his prime brokers which enhances said selling.

The arrogance problem

A recurring pattern in tech-adjacent capital is that domain fluency is mistaken for invulnerability.

I posted a video on this yesterday…

If you can explain transformers, power loads, and memory bottlenecks, it is easy to conclude the only risk is being too early.

Position size, gross exposure, prime-broker terms, and what happens when the narrative is still intact but the path is down 30% in a month get treated as details for traders.

Aschenbrenner had no long public record as a risk manager.

The fund was small in headcount relative to the capital it attracted. That is a structural failing. Markets do not grade essays under stress. They grade whether you can still choose what you want to do.

Whether you’re not forced by a party like Citadel to sell them your listed equities at a discount.

The arrogance is not unique to AI. It shows up whenever a new competence feels like a shield. Crypto had its version. Dot-com had its version. The outcome is almost always the same… the thesis can be directionally right and the practical execution still breaks.

And this is what makes investing and trading SO DAMN HARD.

The irony

AI is not dead. Training runs continue. Capex plans will be revised, not erased. Anthropic and the private layer of the stack are a different liquidity regime than Nebius or Sandisk on a down market.

The irony is funnier.

The person most associated with forcing the industry to take the timeline seriously may have ended the public-markets chapter of his own fund career by running that timeline with borrowed money. The technology did not need him to be levered the way he was. The fund did.

The fund has also been forced to sell its Anthropic stake which is fine and will have probably covered a tonne of the listed equities losses. But it is also not the same job as running a high-profile, levered, public AI book that the whole market used as a sentiment gauge.

Thesis alive. Vehicle constrained. Those are different sentences.

Volatility control, in one paragraph

Vol-targeting funds scale exposure to hit a volatility budget. When markets get louder, they shrink. In a thematic dump, they do not wait for your view on AGI. They sell because realised vol said so. Layer that on prime-broker margin and concentrated longs, and you get air pockets that look like conviction changing when it is often just rules meeting leverage.

What actually matters for a process

Before the screen goes a deep, painful red, ask yourself these questions…

Aschenbrenner’s public unwind is a clean case study in those questions.

Not because AI was a fraud. Because path and leverage decide who still gets to hold the thesis.

We are lucky that we are long only.

Lots of these losses were caused as well by the short side going against him.

If you’re running a long short book, you need to concern yourself with negative correlation too… that is, software went to a correlation of -1 to AI/memory etc.

I.e while AI/memory were falling, software was rallying, amplifying losses!

What this scenario should show you is that investing and trading is not just what stock you think is super cool.

Sure, that’s one side of it.

But what makes investing and trading so damn difficult is when market mechanics get involved and you have to consider things like gamma, margin, forced unwinds, counterparty risk and all that jazz that no one really talks about.

Because these are the hidden risks.

The hidden risks that no one really talks about until it’s too late.

Which is why in the Academy we have a whole section called ‘Markets Trade Against Positions.’

Because we know exactly what happens every single time leverage starts to get too obese.