Citadel bought the AI fire sale. The seller had run out of time.

Watch it play outCitadel took the stocks; the private book stayed

Before the sale, the fund held listed stocks and a separate book of private investments.

Citadel took most, and perhaps all, of the public stocks, depending on the report.

The private book stayed behind, including the fund's stake in Anthropic.

So the shorthand is wrong: Citadel bought the stocks, not the firm.

Replay

• "Margin call" is unconfirmed. Lender pressure and forced deleveraging are well documented, but Reuters reported that it could not determine whether the fund received a formal margin call before the deal.

The clear part: the fund had borrowed against a volatile, concentrated portfolio. When the portfolio fell, it lost the ability to wait for its thesis to recover.

From AI manifesto to a $45 billion gross portfolio

Aschenbrenner, a former OpenAI researcher, was known in AI circles before he ran money. In June 2024 he published Situational Awareness: The Decade Ahead, arguing that fast gains in computing power and algorithmic efficiency could produce artificial general intelligence within years.

The investment version went beyond "buy AI software." Building more capable systems would need an enormous physical buildout: accelerators, memory, data centers, power equipment, cooling and the networks connecting it all.

The fund was built around that thesis. Early backers reportedly included Stripe co-founders Patrick and John Collison, Nat Friedman and Daniel Gross. By March 2026, its US filing showed large positions in data-center power and compute, including Bloom Energy, CoreWeave, Core Scientific, IREN and Applied Digital. It also reported put positions (bets that a stock will fall) on semiconductor companies and a semiconductor ETF.

What the 13F did and didn't show

A 13F is the quarterly holdings report large US managers file with the SEC. The fund's March 31 Form 13F carried a headline value of $13.68 billion. That was not the fund's net asset value:

The filing also left out private companies, most foreign-listed securities, borrowing, cash and short stock positions. Three months later, those gaps mattered most.

The Wall Street Journal reported that the gross portfolio peaked at about $45 billion. Reported assets under management had been above $20 billion. With borrowing and derivatives, a fund can control $45 billion of securities without having $45 billion of investor money.

Leverage turned a strong idea into spectacular returns. It also tied the fund to the path the market took toward the future Aschenbrenner expected.

Why leverage changes the deadline

Take a fund with $10 billion of investor equity that borrows another $20 billion. It now controls a $30 billion gross portfolio.

• Up 10 percent: the portfolio gains $3 billion. Before financing costs and fees, investors are up 30 percent.
• Down 10 percent: it loses $3 billion, or 30 percent of the equity.
• Down 20 percent: $6 billion is gone, leaving far less collateral under the loan.

Leverage multiplies every move in the equity

Pick how much the fund borrows and how the portfolio moves

Borrowing on $10 billion of equity

None$10 billion$20 billion

Portfolio move

−20%−10%+10%

Change in investor equity−20%−10%+10%−40%−20%+20%−60%−30%+30%

−60%+30%

$10 billion portfolio$20 billion portfolio$30 billion portfolio

The lender doesn't share the manager's ten-year conviction. It cares whether the collateral covers the loan today. If asset values fall far enough, it can demand more cash, force positions to be cut or change the terms of the financing.

So "margin call" works as a description even though the contract is private. The mechanism was a collateral squeeze: the fund needed cash just as the assets it could sell to raise it were falling.

A collateral squeeze hands the exit date to the lender

1. AI stocks fall (A concentrated, volatile book) to Collateral shrinks (Less cover under the loan)
2. Collateral shrinks (Less cover under the loan) to Lender acts (More cash, cuts or new terms)
3. Lender acts (More cash, cuts or new terms) to Forced selling (Into falling prices)

Member research: A collateral squeeze hurts the borrower and pays whoever has cash. Issue 04 shows how Citadel was paid to provide liquidity once Situational Awareness could no longer finance its path.

The 67% month that followed a 439% half-year

A client letter cited by multiple reports said the fund returned 439 percent in the first half of 2026 and 1,551 percent since inception, net of fees.

Make a guess

The fund gained 439% in the first half of 2026, then lost about 67% in July. Where did that leave its return for the year?

Tap your guess

Down for the yearYour guess0 to 50%Your guess50 to 100%Your guessOver 100%Your guess

+01234567890123456789%Year to date after July

Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was +80%.

Still up about 80%, per a later investor letter. A 439% gain turns $1 into $5.39; losing 67% of $5.39 leaves about $1.78.

Then came July. A later investor letter reviewed by The Wall Street Journal said the fund lost roughly 67 percent in the month while staying up about 80 percent for the year.

The math checks out. A 439 percent gain turns $1 into $5.39. A 67 percent loss on $5.39 leaves about $1.78, still up roughly 78 percent on the year. A big loss after a big gain wipes out most of the year's dollars without turning the calendar-year return negative.

After a 439% gain, a 67% loss still leaves the year up about 78%

Start of 2026$1.00

First half, +439%+$4.39= $5.39

July, −67%−$3.61= $1.78

After July$1.78

The timeline:

• July 24: the firm told investors about its first-half performance and sought new capital.
• July 30: it had transferred most of the public-equity book.
• After the sale: a letter disclosed the scale of July's loss.

A strong year-to-date number did nothing for the immediate financing problem.

How the winning trade became a trap

A portfolio can grow more fragile as its track record grows more impressive. Five weaknesses built up at once:

1. One economic system. A power company, a GPU cloud, a memory maker and a bitcoin miner have different tickers but depend on the same AI spending cycle and the same willingness to finance it.
2. Volatile positions. AI infrastructure stocks had risen fast and could fall fast when expectations shifted.
3. Borrowing shortened the holding period. A solvent investor without debt can choose not to sell. A leveraged fund may lose that choice.
4. The best private asset was the least liquid. A stake in Anthropic could be worth billions yet take far longer to sell than listed stocks.
5. Crowding. When a well-known portfolio starts cutting large positions, other traders step back, hedge or demand a lower price.

The reversal came in June and July. AI infrastructure and semiconductor shares sold off, and a separate rally in software reportedly hurt another part of the book. The firm tried to raise capital and worked with its banks to cut exposure.

By July 30, the search for cash had become a portfolio sale.

From manifesto to fire sale in about two years

1. Jun 2024Manifesto publishedSituational Awareness: The Decade Ahead
2. Mar 31, 202613F snapshot$13.68 billion reported value
3. First halfUp 439%Net of fees, per a client letter
4. Jul 24Seeks new capitalInvestors told of first-half results
5. Jul 30Public book soldMost of it transferred to Citadel
6. Jul 31Anthropic sale reported$3.5 billion deal, then withdrawn

What Citadel actually bought

The reports differ at the edges:

All agree on the event. Citadel bought a very large block of public securities when Situational Awareness needed cash quickly.

Buying the block doesn't mean Citadel endorsed every position at its previous price. It could negotiate a block discount, hedge individual exposures, spread the risk across teams and finance the positions differently from the seller.

Later reporting put the discount at more than 10 percent, a figure from unnamed sources that we couldn't independently verify. Citadel's first layer of return didn't need AI stocks to recover. It started with the gap between the transfer price and the market price.

Same securities, opposite outcomes. Situational Awareness locked in losses because it had to sell at the wrong time. Citadel could profit because it had cash at exactly that time.

Same securities, opposite outcomes

Situational AwarenessThe seller

• Needed cash quickly
• Sold at the wrong time
• Locked in its losses

vs

CitadelThe buyer

• Had cash at exactly that time
• Bought at a discount of more than 10%
• Could hedge, spread risk and finance differently

The wedding-week contrast

As negotiations intensified, Aschenbrenner was preparing to marry Avital Balwit, chief of staff to Anthropic chief executive Dario Amodei. The Wall Street Journal reported that guests were arriving in Carmel, California, for a multiday celebration: a ceremony at a Tuscan-style villa, a forest-spa send-off and a pre-wedding colloquium.

Later accounts say the couple married that weekend, though we found no announcement from the couple themselves. The detail doesn't explain the losses. It shows how compressed the timeline was: one of the largest AI portfolios in public markets was being taken apart during one of the biggest weekends of its founder's life.

What remained after the sale

Situational Awareness did not disappear. It kept its private investments, most notably an Anthropic stake that reports valued at about $5 billion.

That stake wasn't untouchable either. The Journal's July 31 follow-up reported that the fund had agreed to sell about $3.5 billion of Anthropic shares to Sequoia Capital and Greenoaks, then pulled out of the deal the next morning.

The withdrawn deal covered about 70% of the Anthropic stake

Anthropic stake, as reported$5.0B

• 70%$3.5BAgreed sale to Sequoia and Greenoaks, withdrawn

Illiquidity didn't protect the private stake; it changed the process. The listed infrastructure stocks could be priced and moved as one block under pressure. Selling Anthropic shares needed a negotiated secondary deal, and that deal didn't close.

The fund's future turns on facts that are still private: remaining investor capital, what it owes its banks, redemption requests, the exact Citadel terms and what its private holdings ultimately fetch.

What investors can take from this

• Leverage turns a thesis into a deadline. In the example above, $10 billion of equity controls a $30 billion portfolio. A 20 percent fall costs $6 billion, which is 60 percent of the equity ($6 billion ÷ $10 billion). The lender acts on today's collateral, not on where AI will be in ten years.
• A 13F doesn't show a fund's real risk. The March filing's $13.68 billion headline included about $8.46 billion of shares underlying puts; the common stock was about $3.86 billion. It left out borrowing, cash, short positions and private companies. Copying the holdings list copies none of the leverage behind it.
• Different tickers can be one bet. Power, GPU cloud, memory and bitcoin-mining stocks all leaned on the same AI spending cycle, so they could fall together.
• Big percentage gains hide dollar risk. The fund was still up about 80 percent for the year after losing about 67 percent in July. The year's return stayed positive while most of the year's dollars disappeared.
• Cash is a position too. The buyer with cash on hand got a discount of more than 10 percent from the seller without it.

Questions readers ask

Did Situational Awareness get a margin call?

Nobody has confirmed a formal one. Reuters said it could not determine whether the fund received a margin call before the deal. Lender pressure and forced deleveraging are well documented, and economically the effect was the same: a collateral squeeze.

How can a fund lose 67 percent in a month and still be up for the year?

Because the loss hit a much bigger base. A 439 percent gain turns $1 into $5.39, and a 67 percent loss on $5.39 leaves about $1.78. That is still up roughly 78 percent on the year, close to the reported 80 percent.

Why would Citadel buy stocks a fund was forced to sell?

It bought at a reported discount of more than 10 percent, so its return started with the gap between the transfer price and the market price. It could also hedge individual positions, spread the risk across teams and finance the book differently from the seller.

What happened to the Anthropic stake?

The fund kept it; reports valued it at about $5 billion. On July 31 the Journal reported a deal to sell about $3.5 billion of those shares to Sequoia Capital and Greenoaks, which the fund withdrew from the next morning.

What the sale does and doesn't prove

Nothing here proves that AI investment will slow, that Anthropic is overvalued or that data-center demand will collapse. It shows that a correct long-term thesis can still end in a disastrous portfolio outcome when concentration and leverage take away control of the exit date.

Three separate questions:

• Is the AI infrastructure thesis directionally right?
• Were the securities bought at prices that offered an adequate return?
• Could the portfolio survive the volatility on the way to that return?

Situational Awareness may yet be right on the first. The forced sale of its public book shows the third can overrule the other two.

Citadel bought time from a seller who had run out of it.

The paid issue digs into the less obvious signal: why the 13F never showed the fund's true risk, how Citadel earned a return at the moment of purchase, and what to measure before copying a high-conviction portfolio.