Numbers first.
A hedge fund founded by a former OpenAI researcher reached more than $20 billion in assets under management within roughly two years of inception. At one point in 2025, it was up approximately 270 percent for the year. Then, in a single month โ July 2025 โ its net asset value collapsed by roughly 67 percent. Not over a quarter. Not over a year. Over a month. In the aftermath, the fund was forced to sell the bulk of its stock positions to Citadel, the most sophisticated risk-management machine in the world.
Stop and do the arithmetic before you touch the narrative.
A 67 percent drawdown in thirty days does not happen by accident. Equity markets do not perform that operation on unlevered, diversified books. A concentrated long-only portfolio would need its underlying names to fall 60 to 70 percent in a single month to produce that result. Did that happen in July 2025? No. The AI complex de-rated sharply. The selloff was real. But the plausible description of that month is a broad repricing of 15 to 25 percent on the hardest-hit names โ painful, yes. Ordinary by historical standards, also yes.
So there is only one way to reconcile the numbers. Leverage.
If the underlying basket fell 20 percent, a 67 percent NAV loss implies gross exposure of roughly 3.4 times. If the basket fell 15 percent, the implied leverage is closer to 4.5 times. Either way, the leverage is the story.
The thesis was not wrong. The architecture was. That is the entire lesson of this event in one sentence.
The ledger does not forgive emotion, only math. The math here reads like a textbook case of conviction multiplied by leverage, divided by zero risk management.
This article is a forensic autopsy. I am not here to bury Aschenbrenner or to mock the dead. I am here to audit the machinery that turned a credible technology narrative into a margin-call liquidation. I have spent eleven years watching narratives get priced, levered, and blown up โ in crypto, in DeFi, in algorithmic stablecoins, and now in the AI equity complex. Same pathology. Different ticker.
Context: The Man, The Essay, The Fund
Leopold Aschenbrenner is not a trader. He is a researcher. During his time at OpenAI, he worked inside the most consequential technology project of the decade. He wrote "Situational Awareness," a sweeping argument that artificial general intelligence is not decades away but years away, and that the world is structurally unprepared for its arrival. The essay made him a celebrity in the AI policy and investing worlds.
In 2024, he converted that intellectual capital into a hedge fund. Same name: Situational Awareness. The pitch, in essence: I am on the inside of the most important technology cycle in history. I can read the timeline. I will position your capital accordingly.
That pitch was rewarded. Two years. Over $20 billion. Portfolio managers spend thirty years chasing that number. A first-time money manager with no professional investment background reached it in two. That fact matters, and it should not be flattered. It tells you the capital was priced on narrative conviction, not on risk-adjusted track record. It tells you the product was a token of access to a worldview, not a vehicle of portfolio engineering. The market was paying a premium for the insider's map to AGI.
The year 2025 began as a victory lap. The AI complex ran hard through the first half. The fund's returns were stratospheric; peak marks reportedly approached 270 percent for the year. The AUM figure became a story in itself โ a proof that the smartest AI people would dominate the capital markets too.
Then July arrived. AI equities came under pressure from multiple directions: the weight of an enormous first-half run, escalating valuation scrutiny, and a street of short sellers that began circling the most crowded names. This is where the report's details get sharp: the fund's monthly loss, the margin calls, the sale of the majority of its stock positions to Citadel.
What do we actually know, and what is inference? Let me separate the evidence before I analyze it.
The Forensic Ledger: Confirmed, Inferred, Unknown
A disciplined audit starts with a chain of custody for facts. Here is mine.
Confirmed by multiple reporting channels and the fund's own communication: - The fund lost roughly 67 percent of its net asset value in July 2025. - It was forced to sell most of its stock positions to Citadel to meet margin obligations. - Management acknowledged the damage in a letter to investors: "We let you down this month." - The fund reportedly remained up roughly 80 percent for the year as of the end of July. - At its peak, the fund's AUM exceeded $20 billion. - The fund holds or held positions in private AI companies, including Anthropic. - Aschenbrenner, prior to founding the fund, had no professional investment experience.
Inferred, with high confidence, from the arithmetic: - The fund employed significant leverage โ likely three to five times gross exposure on a concentrated book. - The fund's positions were concentrated in the most liquid, highest-divergence AI names โ exactly the names a short seller can pressure and exit. - The fund lacked a professional risk infrastructure: no meaningful stop-loss regime, no stress-test matrix, no liquidity buffer that functioned.
Unknown, and frankly unavailable to the public: - The exact leverage ratio, notional exposure, and whether leverage was sourced from margin loans, swap contracts, or both. - The precise composition of the book: Nvidia-style semis, power and data-center infrastructure, megacap software, or a blend. - The fee structure and whether redemption gates or lock-ups exist. - The discount at which Citadel acquired the positions. - The governing covenants between the fund and its prime broker.
There is also a data inconsistency I want to flag, because it matters for how you read every headline about this event. The reported loss is about 67 percent. The reported year-to-date result after the loss is about plus 80 percent. The reported peak was about plus 270 percent. These three numbers do not fully reconcile with each other unless the peak occurred earlier in the year and the fund partially gave back gains before July, or the reporting uses different measurement windows. That discrepancy is a red flag about the quality of the real-time NAV disclosure โ and about how much of the fund's reported performance was verified versus approximated. Numbers do not lie, but narratives do. And the narrative around this fund was constructed from numbers that were never fully audited in public.
I will now walk through the anatomy of the failure in five parts.
Part One: Reverse-Engineering the Leverage
Let me be direct about the constraint set.
A fund does not fall 67 percent in a month in a market where its underlying universe falls 20 percent unless the fund is leveraged. The alternative โ that the entire AI universe fell 67 percent in July โ contradicts the observable fact that the year-to-date return remained positive at roughly 80 percent by month-end. The AI complex had a bad month, not an existential collapse.
So the loss was manufactured internally. It was manufactured by gross exposure.
There is a deeper point about the timing of the leverage. The fund did not need leverage to produce a spectacular 2025. A 270 percent year on an unlevered, concentrated, correctly-positioned AI book would have been one of the great hedge fund records of the decade. The leverage was not the strategy. The leverage was a competitive choice: to be the biggest AI fund fastest, to maximize AUM and fee revenue in a window of maximum narrative heat, to convert an intellectual reputation into an institutional balance sheet before the window closed.
That choice is the root error. It is the same error I have seen repeatedly in markets: the substitution of conviction for capital structure. When your edge is information โ an insider's reading of the AGI timeline โ you feel certain about the direction. You therefore believe you do not need a hedge, a stop, or a buffer. Certainty is the enemy of survival.
I modeled the Terra/LUNA algorithmic stablecoin in early 2022 using Monte Carlo simulations. My model produced a 68 percent probability of de-peg under high-volatility conditions. I flagged it to my supervisor. He ignored the report. When the de-peg came, the mechanism did not care that many brilliant people believed in the thesis. The mechanism cared about the math of the collateral. The same is true here: leverage is a mechanism, and mechanisms do not read essays.
There is also the short-seller dimension. The reporting mentions that short sellers amplified the selling. That should not be read as a conspiracy against Aschenbrenner. It should be read as market structure. When a fund is running a highly-concentrated, leveraged, high-beta book, it becomes a target. Short sellers do not attack diversified books with transparent risk limits; they attack opaque books with forced-selling potential. The leverage invited the attack. The concentration gave it power. The absence of a defined exit made the attack terminal.
Part Two: The Margin Call Machinery โ A Close Reading of the Forced Sale
Most people do not understand how a forced sale to another institution actually operates. Let me walk through it.
A levered fund pledges its securities to a prime broker or lender. The lender applies a haircut: a loan-to-value ratio that reflects the risk of the collateral. As the market falls, the collateral value falls and the loan-to-value ratio rises. When it breaches the covenant, the lender issues a margin call. The borrower must post additional cash or collateral, or the lender will liquidate positions to restore the ratio.
When the fall is violent and the book is concentrated, the fund faces a cascade. Selling to meet one call pushes prices lower, which triggers the next assessment, which requires more selling. In a vacuum of bids, the only buyers are the liquidity engines โ firms with the capital, the hedging infrastructure, and the risk appetite to take on a distressed book at a discount.
That is Citadel's role here. Citadel is not an AI bull. Citadel is a counterparty. When a forced seller hands a book to Citadel, Citadel is buying optionality: the right to hold a discounted position, to hedge it, to unwind it into any bounce, or to carry it if the fundamentals hold. The entry price is set at the moment of maximum duress. The discount is the price of the fire sale. And that discount is a real, realized loss to the selling fund's investors.
The "rescue," in other words, is the transfer of the recovery option from the seller to the buyer. If the AI complex rebounds in the fourth quarter of 2025, the profit from that rebound will accrue to Citadel, not to the LPs of Situational Awareness. That is not a rescue. That is a transaction.
Liquidity is a ghost; it vanishes when you blink. It was present in June, when everyone wanted in. It evaporated in July, the instant everyone wanted out. And the fund, because it had no liquidity buffer, no hedging layer, and no pre-negotiated facilities, sold into exactly the vacuum it had helped create.
I learned this lesson in the 2020 DeFi Summer. I deployed personal capital into a newly launched automated market maker and built a Python script to monitor gas fees and slippage in real time. When the protocol suffered a flash-loan attack โ the price oracle was manipulated โ my script executed a defined exit within 45 seconds. I recovered 92 percent of principal. The true believers, the ones who held conviction without a monitoring layer, lost materially. I did not out-predict the attack; I out-structured it. That is the difference that matters. A thesis without an exit mechanism is not an investment. It is a donation.
The same principle applies at institutional scale. A fund two years old with $20 billion under management should have a documented, tested, rehearsed liquidation protocol. It should know, in writing, exactly what it will sell, at what trigger, in what sequence, if the market breaks. The evidence suggests no such protocol existed. If it did, it either failed under stress or was never designed for a 20 percent drawdown on a levered book. Both are failures of the same category: institutional negligence dressed as intellectual confidence.
Part Three: The Bank Run Narration
One detail in the reporting deserves special attention. In his letter to investors, Aschenbrenner invoked the metaphor of a bank run.
Think about that choice. A manager with a levered, concentrated book, facing margin pressure, writes to his investors and uses the words "bank run." There is no universe in which that narrative calms redemption requests. There is no universe in which telling LPs that the fund is experiencing a bank run slows the exit of capital. If anything, it accelerates it. It tells every investor on the fence that others are leaving. It tells them the fund cannot satisfy everyone. It tells them the asset-liability mismatch is real.
The "bank run" framing was a self-fulfilling prophecy. The manager who intended to warn investors of the consequences of panic instead handed them a map of the exits.
Anchor pegs break before trust does. The peg here was trust: trust that a brilliant researcher's market judgment would outrun market mechanics. It broke. Not because the AI thesis is wrong, but because pegs buckle under structural pressure, and no amount of conviction calibrates them.
There is a compliance dimension too. In the funds I have audited, the first rule of crisis communication is: do not lie, do not hype, and do not invent metaphors that trauma your own investor base. The letter "We let you down this month" was honest. The "bank run" language undermined it. One sentence acknowledged the loss. The next sentence manufactured the panic. That is not a risk management failure; that is a communication failure layered on top of a risk failure.
Part Four: What Risk Architecture Was Missing
It is easy to mock the dead. It is harder to specify what should have existed. A competent risk framework for an AI-themed equity fund would have required, at minimum, five components.
First, concentration limits. No single name should have exceeded five to eight percent of NAV. A multi-hundred-billion-dollar technology complex offers thousands of candidate positions. A fund that cannot diversify across semis, power, cloud, software, and infrastructure has made a theological choice, not an investment choice. The evidence here points to a book that was effectively a single thesis with a single collateral pool.
Second, a stress-test regime. A mandatory matrix evaluating the book under 10 percent, 20 percent, and 35 percent drawdowns. The fund should be required to prove, in writing, that it can survive the worst case without forced selling. The July 2025 event would have been a test question, not a surprise. My own practice, developed in the 2022 Terra collapse and refined since, is that every portfolio must be survivable at the stress level that the position's leverage implies. The leverage implies a stress level. The fund's behavior implies it was never tested.
Third, a liquidity buffer. Cash or liquid treasuries equal to a defined percentage of margin requirements. This ensures the fund can meet a call without liquidating into a vacuum. A 5 percent liquidity buffer on a $20 billion fund is $1 billion. It is expensive. It is also the difference between a rough month and a funeral.
Fourth, a pre-specified hedging layer. Index puts or single-name collars that activate automatically in a crash, even when the manager remains in denial. Hedging is not a belief system. It is insurance. And insurance is purchased precisely because the insured direction might be wrong.
Fifth, alignment between asset liquidity and redemption terms. This is the Anthropic problem. If you hold private, illiquid AI shares with no transparent mark, you cannot run a levered public-equities book with daily margin obligations without a covenant that protects the fund from mismatched cash needs. A $20 billion fund holding private positions in Anthropic is structurally exposed: the public book can be forced to sell, while the private book stays frozen. That mismatch is a hidden death spiral. If the private stake is ever sold at a distressed valuation, it will confirm the contagion path from public forced selling to private markdowns.
Would such a framework have produced a 270 percent year? Possibly not. The leverage was doing a lot of the algebraic lifting. But it would have survived into August. And survival matters more than peak performance. A 100 percent gain followed by a 50 percent drawdown is a zero percent cumulative return. The manager who survives to fight again will always beat the manager who swings violently and requires rescue.
This is where my 2017 experience comes in. While my undergraduate peers bought ICO tokens on the strength of whitepapers, I spent three weeks auditing the Tezos smart contracts. I found a race condition in the delegation logic and published a warning. I sold my pre-mine allocation at mainnet for $4,200 while later adopters faced the consequences of buying narrative. The lesson was small in dollars and permanent in method: audit the code, not the promises. The code here โ the fund's risk control stack โ was the promise. There was no code to audit.
I audit the code, not the promises. The only code I could audit here was the margin math. And the margin math was the one thing that worked exactly as designed.
Part Five: The Return Trap โ Why "Still Up 80 Percent" Is a Distraction
The number you will hear in the fund's defense is: "But it is still up about 80 percent for the year."
Correct at the fund level. Meaningless at the investor level.
This is the time-weighted versus dollar-weighted return trap. The fund's NAV may be up 80 percent from January 1 to July 31. But the typical investor's experience is dollar-weighted, and flows chase returns. When did the capital flood in? At the peak. When the fund was up 270 percent. When the story was euphoric. When every LP in the AI orbit wanted exposure. That capital โ the highest vintage of assets โ took the full 67 percent hit.
So there are two distinct populations of investors. The early cohort, who are up enormously. And the peak cohort, who have lost two-thirds of their capital in one month. The fund-level average conceals the actual distribution of outcomes. And here is the deepest cut: the peak cohort's losses are essentially the early cohort's exit liquidity. Late money bought the thesis at the top. The market transferred their wealth to whoever sold early.
This pattern is not unique to this fund. It is the default pattern of performance-chasing capital. I have watched it in institutional flows for years. In 2024, I led a team that automated extraction of institutional flow metrics from our data infrastructure. We cut report generation time from four hours to 45 minutes and identified a $2.3 billion inflow trend before mainstream coverage picked it up. That edge โ reading flows ahead of narratives โ cuts both ways. Flows that enter a rocket are the flows that get burned at the peak. When you see a first-time manager with $20 billion in assets and no institutional risk pedigree, the flow pattern is not a prediction. It is a warning.
The correct question after July 2025 is not "is the fund up year-to-date?" The correct question is "what was the experience of the marginal dollar that subscribed at the worst time?". That dollar is the only honest measure of a fund's real relationship with its clients.
The Contrarian Angle: Three Comforting Myths and One Uncomfortable Truth
Myth one: this event discredits the AI trade.
It does not. The AI industry's long-run fundamentals โ model capability, compute demand, enterprise adoption, the actual trajectory of the technology โ do not change because one levered vehicle blew up. What changes is the marginal cost of capital. A blow-up like this tightens fundraising conditions for AI-themed funds. It raises scrutiny on sector leverage. It arms the short side with a scalpel. But it does not touch the technology's economic engine.
I watched the identical category error during the 2022 Terra/LUNA collapse. Mainstream observers concluded that the failure of an algorithmic stablecoin meant the failure of decentralized finance. It did not. It meant the failure of an unbacked mechanism. The difference matters, and it persists. The AI thesis will outlive this fund.
Myth two: Citadel is a rescuer.
Citadel is not a rescuer. Citadel is a counterparty. When a fund is forced to sell positions of this size in this time frame, the buyer holds all the negotiating power. The discount is a realized loss to the seller's investors. If the market rebounds, the rebound profit accrues to Citadel. The "rescue" is the purchase of the recovery option at a discount from a seller who had to sell. Nothing about that is charity.
Myth three: Aschenbrenner's inexperience was the problem, and experienced managers would not have blown up.
Partially true, but incomplete. Sophistication without discipline is its own disease. Experienced managers blow up the same way, because experience often breeds overconfidence in one's ability to time exits. The real structural problem is the incentive environment: fundraising rewards speed to size, and fee economics reward maintaining AUM over the risk-adjusted custody of capital. A first-time manager with an AI thesis and a $20 billion book is a mismatch of incentives. The market priced the narrative โ the insider's claim to the AGI timeline โ and underpriced the fragility of the narrative's vehicle.
The uncomfortable truth is the conflict-of-interest layer. A prominent AI commentator managing $20 billion in levered AI positions holds a public voice and a private book in the same asset. Every public statement he makes about the speed of AI progress moves the assets he owns. Whether or not he was deliberate, the incentive conflict is structural. In a market, sincerity does not excuse structure. When the narrator holds the book, the narrative and the book become a single position.
This is also an information-deflation story. Aschenbrenner's edge โ insider access to AI research and AGI timeline judgment โ was a depreciating asset from the day he left OpenAI. Information diffuses. The essay was published. The insights became public. The scarcity value converted to a commodity. What he did not have, and could not buy, was a mechanism to convert a decaying informational advantage into a durable capital advantage. The market recognized the narrative premium for a while. Then the market repriced it. The repricing was delivered through the margin call.
Scenario Analysis: What Happens Next
I do not make predictions. I structure probabilities.
Base case โ probability medium. The fund survives in reduced form. AUM shrinks through redemptions. The Anthropic stake remains intact at its last carry value. Aschenbrenner continues with a smaller, more conservative book. Impact: moderate brand damage, modest tightening of AI fundraising conditions.
Stress case โ probability medium. Redemptions continue and force the fund to sell the private stake at a discount to last valuation. That is the turning point. A discounted Anthropic transaction would mark the first major private-market AI valuation repricing triggered by a secondary forced seller. It would send a signal through the entire AI venture complex โ not that AI is over, but that private valuations built during a narrative mania are now subject to public-market discipline. Impact: high.
Recovery case โ probability low-to-medium. The AI complex rebounds sharply by the fourth quarter. Citadel's acquired positions appreciate. The fund's surviving book partially repairs. The manager commits to lower leverage and retains a meaningful asset base. Some credibility is salvaged. Impact: moderate positive.
Now the signals to track, in order.
Short term, one to four weeks: Does the fund issue additional correspondence? Do Citadel's quarterly disclosures reveal the acquired positions? Do block trades leak the discount? Do other AI-themed funds report similar drawdown pressure? If three or more levered AI funds are simultaneously in distress, this stops being a single-manager story and becomes a sector-structure story.
Medium term, three to six months: Does the fund disclose a new risk framework โ leverage reduction, a dedicated risk officer, position limits? Does Anthropic raise a financing round, and at what price? Does the fund syndicate its private stock to long-term strategic holders? The answer to each of these is a data point on the fund's viability.
Long term, six to twelve months: Does the AI equity complex recover and set new highs? Does this event become a standard case study in narrative leverage, permanently inserted into LP due-diligence checklists? Or does it vanish into the noise of the next cycle? The answer will tell you whether the market learned the structural lesson or merely updated its fear index by one data point.
The critical unknown remains the fund's internal controls. What were the actual leverage ratios? Who set them? Who had the authority to stop the bleeding? What did the risk committee โ if one existed โ know, and when did it know it? The letter to investors is one thing. The council of witnesses is another. Until the fund discloses its risk documentation, any claim about its path forward is inference, not evidence.
The Takeaway: Structure Survives the Storm
The lesson of this episode is not that artificial intelligence is overvalued. It is not that the AI trade is dead. It is not that researchers of exceptional intelligence cannot participate in markets.
The lesson is narrower and harsher. Cognitive edge is not a risk model. The ability to predict the direction of a technology cycle does not equal the ability to survive its volatilities. The market charges a premium for conviction โ and it collects that premium with interest, through the mechanism of forced liquidation.
Structure survives the storm; chaos drowns it. When you allocate to any high-conviction manager, in any asset class, ask one question before any thesis question: what happens when the book falls 25 percent? If the answer requires hope, walk away.
I have sat through the worst five minutes of three different markets. The most valuable sentence a manager can say is not "I was right." It is "I have a plan for being wrong."
The AI thesis will outlive this fund. The next brilliant insider who offers you 270 percent may even be right. Ask him who his risk officer is. Offer him a seatbelt. If he laughs, the market will soon laugh back at his ledger.