When a hedge isn't a hedge: the Situational Awareness collapse

I spent 25 years inside bank Treasuries, where the phrase that ended careers was never “we were wrong.” It was “we thought that position was hedged.”

That distinction is the whole story of what happened to Situational Awareness, the AI-focused hedge fund run by former OpenAI researcher Leopold Aschenbrenner. Press reports put the fund at roughly $45 billion at the start of July 2026, up several hundred percent since its 2024 launch, and down about 67% by the end of that single month — after which it reportedly sold its entire public equity book in one block trade to another manager and continued on as a private investment vehicle. It is not a fraud story and, as far as the public record goes, not a misconduct story. It is a structure story, and the structure is one that shows up in products a retail investor can buy in a single click.

The trade was one idea wearing two costumes

The fund’s thesis was that artificial intelligence would arrive faster than markets expected, and that the companies supplying the physical layer — chips, memory, data centres, power, cloud capacity — were undervalued relative to that arrival. Reports describe it expressing that view in two directions at once: long the infrastructure suppliers, and short a set of software businesses it expected AI to disrupt.

On a risk report, that looks like a hedged book. Longs on one side, shorts on the other, exposures partly cancelling. But look at what the short leg was actually betting on. It was betting that AI would arrive fast and reshape software. Which is the same sentence as the long leg.

A short position is only a hedge if it responds differently to the event that hurts your longs. If it responds to the same event the same way, it isn’t insurance. It’s a second helping.

In July, reporting says the AI-infrastructure names fell hard while several of the shorted software businesses rallied sharply. Both legs lost, on the same days, for the same reason. The costume came off.

I wrote a while back that ten funds which all fall together are one big bet wearing a costume. This is the institutional version of exactly that, and it is worth being precise that the underlying thesis may still turn out to be correct. AI infrastructure may well be undervalued today. The fund simply did not survive long enough to find out. Being right eventually and being solvent throughout are two different achievements.

What the leverage actually multiplied

Reports put the fund’s public book at roughly four times leverage, financed across three prime brokers.

Here is the part that is easy to get wrong, and it is not “4x losses became 8x.” The arithmetic is subtler than that. Take a stylised book — we don’t know the fund’s real split, so treat these as round numbers illustrating the shape:

Start with $1 of capital and four times gross leverage, so $4 of positions. Put $2.50 into longs and $1.50 into shorts.

• Net exposure: $2.50 − $1.50 = $1.00. One times capital. On paper, a modest, market-neutral-ish posture.
• Gross exposure: $2.50 + $1.50 = $4.00. Four times capital.

Which of those two numbers your risk actually lives on is decided entirely by correlation. If the legs genuinely disagree with each other, losses behave like the $1.00 net. If both legs go wrong together, losses behave like the $4.00 gross.

Run it: longs fall 30%, and the shorted names rise 30%. The long side loses $0.75. The short side loses $0.45. Total: $1.20 of loss against $1.00 of capital.

Nobody set out to risk everything. The leverage multiple was four, not a hundred. The damage came from the gap between the exposure the book appeared to carry and the exposure it actually carried once the correlation assumption failed. Correlation is the switch that decides which number leverage gets applied to — and that switch is thrown by the market, not by the manager.

The third blow: you don’t choose when to sell

The stylised arithmetic above produces a loss worse than 100% of capital, yet the reported figure was around 67%. That gap is not an error. It is the margin call.

Leverage is borrowed money, and lenders have their own risk limits. When the collateral falls far enough, the lender doesn’t ask about your thesis or your time horizon — it demands cash or it sells. Which means the positions get liquidated near the bottom, on the lender’s schedule, converting a paper drawdown into a permanent, realised loss. The margin call is what both caused the wipeout and, in the grim accounting, capped it.

This is the part that most changes how I think about leverage. Unlevered, a bad year is survivable — the investor decides whether to hold on. Levered, that decision is transferred to a counterparty whose only question is whether the collateral covers the loan today.

Which is worth stating more precisely than “leverage is dangerous,” because that version is lazy and the accurate version is more useful. Institutions use leverage durably all the time. What separates them from this episode is not the multiple — it is whether the de-risking decision was pre-committed. Serious leveraged books run hard, mechanical limits: position sizes set against how quickly a holding could actually be sold, caps on how much risk any one theme may carry, and drawdown triggers that cut exposure automatically, well before a lender would. The purpose of all of it is to make the manager reduce first, on their own schedule, rather than have the counterparty reduce for them at the bottom.

Two caveats stop that from being a recipe. A trigger is only as good as the risk estimate that sized it — a book believed to run one times capital while actually running four has its limits calibrated to an assumption wrong by a factor of four, so they fire far too late to help. The correlation error comes first, and it is what renders the discipline downstream ineffective. And no discipline makes a margin call impossible, because a lender can reprice its collateral requirements unilaterally, which is exactly what stress tends to produce: the call can arrive without the price moving another tick.

This is professional machinery, and worth naming as such. Running leverage with that apparatus is the work of institutional desks and of investors with equivalent experience and infrastructure. It is a different activity from holding a levered position in a brokerage account, which is where the next section goes.

What this transfers to an ETF portfolio

Three things, none of which require a $45 billion book to matter.

First, count relationships, not holdings. The reason a portfolio can hold a dozen funds and still be one bet is that diversification comes from holdings that respond differently to the same event — not from the number of line items. Measuring that properly means looking at how holdings actually co-move, across different market conditions and time horizons, rather than assuming a fund label implies independence. That measurement is the core of what the analysis engine behind this site does, and it is open source specifically so the correlation work can be inspected rather than taken on faith.

Second, understand what a daily-reset leveraged ETF is structurally doing. These products carry the same two mechanics in a retail wrapper. The leverage multiplies moves, and — critically — the daily reset makes the fund a forced trader: to maintain its stated multiple it must sell exposure after declines and add after rallies, every session, regardless of what its holders think. That is the margin-call dynamic built into the product’s plumbing, arriving on a schedule rather than as a phone call. Over any period longer than a day, in a market that moves around, the path taken subtracts from the outcome — which is why such a fund can lag its stated multiple of an index even when the index finishes flat. This is disclosed in the products’ own documentation; it is a design feature, not a surprise. And the institutional answer above does not transfer here: a stop does not address this, because the erosion is not a drawdown event that can be exited. It arrives continuously, a fraction at a time through each daily rebalance, in markets that rise as well as fall.

Third, treat “hedge” as a claim requiring evidence. A position described as protection deserves the question: protection against what specific event, and what is the evidence it behaves differently when that event occurs? A holding added for balance that quietly moves with everything else is not ballast. It is more of the same bet, purchased at the price of thinking otherwise.

None of this is a comment on AI, on the fund’s thesis, or on what any market does next. It is the oldest lesson on a Treasury desk, delivered at unusual scale: risk you have measured is risk you can hold. Risk you have assumed away is the one that decides how the story ends.

Want a quick read on how your own approach handles correlation and concentration? The free ETF Portfolio IQ Score takes a few minutes and asks for no dollar figures: etfwealthiq.com/iq-score.