On the morning of Thursday, July 30, 2026, before the US market opened, the entire public equity book of the fastest-growing hedge fund in history changed hands in a single block trade. Roughly $16 billion of longs and shorts moved from Leopold Aschenbrenner's Situational Awareness LP to Citadel at a reported discount of about 10% to market. The deal was proposed the previous morning and closed inside twenty-four hours.
Then something happened that complicates every tidy morality tale written about it since. The stocks went up. Nebius closed up 27% that day. IREN up 26%. Bloom Energy up 26%. SanDisk up 25%. CoreWeave up 22%. The equal-weight S&P 500 was flat.
Two weeks later, the names Aschenbrenner was forced out of have rebounded 30% to 90% off their July 29 lows, and the earnings they reported in the interim were, with one important exception, extraordinary. That exception, SK Hynix, is the part of this story that actually matters for anyone holding AI infrastructure equities today.
This is an attempt to separate three things that have been mashed together in the coverage: what happened to a man, what happened to a balance sheet, and what happened to a thesis. They are not the same event, and only one of them is genuinely unresolved.
The chain of events
| Date | Event |
|---|---|
| Apr 2024 | Aschenbrenner fired by OpenAI. He says the cause was a security memo he took to the board; OpenAI's stated reason was a leak. |
| Jun 2024 | Publishes Situational Awareness: The Decade Ahead, 165 pages. |
| Jul 2024 | Launches Situational Awareness LP with roughly $225M seeded by Patrick and John Collison, Nat Friedman and Daniel Gross. |
| Feb 2025 | Buys into Anthropic at roughly a $60B valuation. |
| 2025 | Fund compounds hard; AUM goes from ~$255M (Q4 2024 13F) to $5.5B (Q4 2025 13F). |
| Mar 31, 2026 | Q1 13F shows $13.7B across 42 positions and $8.5B of it in put options. |
| May 28, 2026 | Anthropic closes a $65B Series H at a $965B post-money valuation. |
| Jun 8, 2026 | WSJ reports the fund manages "over $20 billion." |
| Jun 22, 2026 | Momentum peaks. The factor had returned ~57% in H1. This is where the drawdown starts. |
| Jun 30, 2026 | Fund reports +439% net for H1 2026. Q2 13F (filed Aug 14) shows a $20.2B book with the put hedge gone. |
| Jul 1, 2026 | NAV reported at ~$45B. |
| Jul 10, 2026 | SK Hynix lists ADRs on Nasdaq, raising $26.5B the largest foreign debut in US history. It closes up 12.8%. This is the top. |
| Jul 24, 2026 | Investor letter reports the 439%, calls the drawdown one of the most attractive opportunities since early 2025, and invites fresh capital for August 1. |
| Jul 27-30, 2026 | Morgan Stanley's momentum index falls 17.4% in four sessions, the worst four-day stretch in the history of the series. Its tech and media components fall 36%. |
| Jul 28-29, 2026 | Over $1 trillion comes off global chip stocks. SK Hynix posts record earnings that miss consensus. The Fed holds with three dissents; the 30-year yield runs to 5.25%, the highest since 2007. |
| Jul 29, 2026 | Every US name in the book bottoms on the same day. Griffin and Aschenbrenner speak that morning; Citadel analyses the book overnight, beating out Jane Street and Millennium. |
| Jul 30, 2026 | The block trade. AUM falls to ~$10B. The cohort rips 22–28% the same session. |
| Aug 1-2, 2026 | Aschenbrenner marries Avital Balwit, chief of staff to Anthropic's CEO, in Carmel. |
| Aug 4, 2026 | The fund puts $400M into an undisclosed private company, on top of $100M the previous month. |
| Aug 5, 2026 | Citadel's July numbers land: Wellington +5.9%, its best month since 2022. |
| Aug 7, 2026 | Bloomberg reports investors are lining up. Aschenbrenner tells LPs the fund is not taking new money. |
| Aug 11-12, 2026 | CoreWeave and Nebius report. Nebius jumps 34% in a session. |
Who Aschenbrenner is
He was born in Berlin around 2001 to two physicians, enrolled at Columbia at fifteen, and graduated valedictorian of the Class of 2021 at nineteen with a degree in economics and mathematics-statistics. He went to the Global Priorities Institute at Oxford, then to Sam Bankman-Fried's FTX Future Fund in February 2022, resigning before the November collapse. In 2023 he joined OpenAI's Superalignment team under Jan Leike and Ilya Sutskever and co-authored "Weak-to-Strong Generalization."
In April 2024 OpenAI fired him. The company pointed to a leaked document. His account, given on Dwarkesh Patel's podcast that June, is that he had written an internal memo calling OpenAI's security "egregiously insufficient to protect against the theft of model weights or key algorithmic secrets from foreign actors," taken it to board members after a security incident, and been told by HR that the memo was "a major reason" for his dismissal. He also says company lawyers questioned him on whether the Superalignment team was "loyal to the company."
Two months later he published the essay. A month after that he was running money. He was twenty-two.
That sequence is the whole psychological shape of what followed: a man who had twice watched an institution he was inside of come apart, who concluded that the institutions were not going to price the thing correctly, and who decided to price it himself with borrowed money.
The thesis
Situational Awareness: The Decade Ahead is 165 pages in nine parts, and it makes one structural argument with a financial corollary.
The argument. Capability gains come from compounding "orders of magnitude" of effective compute, consisting of raw compute scaling, algorithmic efficiency, and what he calls "unhobbling" (turning a chatbot into an agent). He puts the combined rate at roughly 0.5 OOM per year and concludes that "AGI by 2027 is strikingly plausible." From there, automated AI research compresses a decade of algorithmic progress into under a year, the intelligence explosion. The remaining sections are consequences: lock down the labs against Chinese exfiltration, solve superalignment, and expect that "by 27/28 we'll get some form of government AGI project," which he calls The Project.
The corollary. If you believe the capability curve, you must believe the physical inputs. The essay's most-quoted line is not about intelligence at all: "By 2030, a trillion-dollar cluster, 100 GW, more than 20% of all U.S. electricity generation. And that's just the training cluster." Many trillions, he wrote, "will go into GPU, datacenter, and power buildout before the end of the decade."
Translate that into positions and you get the fund almost mechanically:
- Long the bottleneck. Whatever is scarce on the path to 100 GW captures the rent. In practice: high-bandwidth memory and NAND (SK Hynix, Micron, SanDisk), neoclouds renting GPU capacity (CoreWeave, Nebius, IREN, Core Scientific), and power and thermal (Bloom Energy, Applied Digital, T1 Energy). Fortune quoted the thesis crisply: "The most valuable assets in the AI era may not be algorithms, but electricity and computing power."
- Short the disrupted. Software businesses whose moat is labour that models can replace.
- Own the compounder privately. About a fifth of the fund sat in Anthropic, plus MatX and Fluidstack.
This is a picks-and-shovels trade, not a model-maker trade, and on its own terms it worked spectacularly. It also contains a landmine that nobody, including its author, priced: bottleneck rents are the most violently mean-reverting cash flows in the economy. SanDisk printed an 84.6% gross margin in the June quarter. That is not a moat. That is a shortage. Shortages summon capacity, and capacity arrives on a fab schedule that has nothing to do with your redemption terms.
The two weeks
The popular version is that a stock market crash killed the fund. The market did not crash. CNBC's own headline was "Why Situational Awareness hedge fund imploded, even in a tame stock market." What crashed was a factor.
June 22, the top. Momentum, up roughly 57% in the first half, peaks and begins to bleed. The fund's book was, structurally, long momentum and short anti-momentum. Once the factor turned, both legs lost money on the same days and the short book stopped being a hedge and started being a second long.
July 10, the misread catalyst. A lot of commentary blames an SK Hynix event around this date. The event was real but it was not bad news: SK Hynix priced 177.9 million ADSs (selling their stock on the US market as a foreign company) at $149 and listed on Nasdaq, raising $26.5 billion in the largest-ever US debut by a foreign company, with demand more than seven times the book. It closed the day up 12.8% and ran to about $195 within days.
The problem is the mechanism, not the headline. That listing injected $26.5 billion of fresh supply into the single most crowded long in the market, at the exact moment the memory narrative was turning. Within three weeks SK Hynix traded below $137. Micron fell 33% from its peak, SanDisk 50%, Western Digital 42%. The reason was not AI demand. It was HBM capacity expansion plus CXMT's IPO funding another wave of Chinese domestic memory supply, supply-side news in a supply-constrained trade.
July 24, the letter. Six days before the liquidation, Aschenbrenner reported +439% net for the half, described the drawdown as creating some of the most attractive opportunities since early 2025, and invited investors to add capital from August 1. Everything you need to know about the mismatch between his conviction and his financing is in that document.
July 27-30, the four-day crash. Morgan Stanley's sector-neutral momentum index fell 17.4% in four sessions, the worst in the history of the series; its tech and media sleeve fell 36% against a prior record of about 20%. BTIG's Jonathan Krinsky: "we just witnessed the largest/fastest momentum crash in modern history." Over $1 trillion came off global chip stocks. On July 29 the Fed held with three dissents and the 30-year ran to 5.25%, the highest since 2007, a direct hit to every long-duration, cash-burning name in the portfolio. That same day SK Hynix reported record revenue and operating profit and still missed consensus.
At roughly 4x leverage, expressed through total return swaps across four prime brokers, a 25% drawdown is an equity wipe. Goldman, JPMorgan, BofA and Citi called for collateral daily.
He never missed a margin call. The banks lost nothing. Their dynamic margining had automatically delevered him on the way up. This is why comparisons to Archegos are lazy. Bill Hwang cost the banks over $20 billion and went to prison for fraud. Aschenbrenner ran out of equity, told his counterparties so in advance, and sold. BofA's Brian Moynihan told CNBC on August 5 that the bank "would have been fine" without Citadel, while also calling the episode a warning shot: "Valuations get out, leverage in the system gets there."
July 30, the trade. Griffin and Aschenbrenner spoke on the Wednesday morning. Citadel ran the book overnight with Pablo Salame, Gerald Beeson and Perry Vais, and bid against Jane Street and Millennium, Jane Street being, awkwardly, an investor in the fund. The whole thing, longs and shorts, went across in one pre-market block at a discount of about 10%.
The structure is the point. As SpotGamma put it: "The positions didn't hit the tape as supply; they changed hands." A piecemeal liquidation into the open market would have gone through the AI complex like a chainsaw. Instead the overhang was extinguished in a single print, the cohort immediately repriced upward, and Citadel's Wellington fund went from roughly flat on the month to +5.9% (its best month since 2022), with Tactical Trading at +11% and Equities at +14%. Neither Citadel nor Griffin has commented. An investor familiar with the deal told Reuters: "There are only a few firms that could take down this much risk."
Some interesting disclosure
The Q2 2026 13F, covering June 30 and filed on August 14, 2026, after everything above, reframes the story.
At March 31, the book looked defensive. Of $13.7 billion in reportable positions, $8.5 billion was in puts, including index puts and puts on Nvidia, Oracle, Broadcom, AMD, TSMC, ASML, and $584 million of puts on Micron. Long common stock was only $3.9 billion.
At June 30, that hedge is gone. The book is $20.2 billion across 26 positions and the entire put sleeve has collapsed to a single $5.2 million position in Infosys. Micron has flipped from a $584 million put to a $5.6 billion long, the second-largest position in the fund. SanDisk is the largest at $5.7 billion. Nebius, which appears in no Q1 filing at all, is a $1.2 billion long. Bloom Energy is $1.9 billion.
So the sequence is not "conviction bet goes wrong." It is: a hedged, 42-position book was converted over one quarter into a 26-position, essentially unhedged, memory-concentrated long book, and then run at 4x into the worst momentum crash on record. The single largest thesis reversal (Micron from short to long) was into the exact sector that broke.
This also corrects a claim repeated widely in late July: that Nebius, SanDisk, Micron and CoreWeave were the fund's top disclosed Q1 holdings. They were not. Nebius was absent, Micron was a put, and the largest disclosed Q1 positions were downside hedges.
Where the stocks actually are
Every US name in the book bottomed on the same session, July 29, 2026, which is itself the tell that this was a financing event, not a fundamental one.
| Pre-crash high (Jun) | Jul 29 low | Aug 14 close | Off the low | 2026 YTD | |
|---|---|---|---|---|---|
| SanDisk (SNDK) | $2,335 | $1,016 | $1,641 | +61% | +591% |
| Nebius (NBIS) | $287 | $148 | $277.68 | +87% | +232% |
| CoreWeave (CRWV) | $125 | $60.82 | $105.26 | +73% | +47% |
| Micron (MU) | $1,214 | $739 | $971.66 | +32% | +240% |
| Bloom Energy (BE) | $346 | $164 | $229.94 | +40% | +165% |
| Applied Digital (APLD) | $48 | $23.22 | $31.20 | +34% | +27% |
| SK Hynix (SKHY) | Listed Jul 10 | $126.79 | $166.33 | +31% | +12% vs offer |
And the earnings that landed on either side of the liquidation:
- SanDisk (Aug 5, FQ4): revenue $8.97B, up 372% year over year and 51% sequentially; GAAP EPS $43.97; gross margin 84.6%; guided FQ1 to $10.3–10.8B.
- Nebius (Aug 12, Q2): revenue $582.3M, up 454%; still lossmaking at $(190.4)M; stock +34% on the day.
- CoreWeave (Aug 11, Q2): revenue $2.58B, up 112%; remaining performance obligations of roughly $104B; net loss $(626)M; FY guide $12.4-13.2B.
- Bloom Energy (Jul 28, Q2): revenue $1.07B, up 166%; GAAP EPS $0.62; raised FY26 guidance to $3.9-4.2B.
- Applied Digital (Jul 27, FQ4): revenue $258.7M, up 407%; adjusted EBITDA $42.4M; still net lossmaking.
- SK Hynix (Jul 29, Q2): revenue ₩79.3T, up 257%; operating profit ₩60.5T at a 76% operating margin, and still short of consensus.
Note what that last line does to the narrative. The one company that disappointed did so while earning a 76% operating margin.
What this does to the thesis
It did not test it. The essay is a claim about 2027 through 2030. The fund was a 4x-levered vehicle with quarterly liquidity. Nothing about July resolved whether compute and power are the binding constraint on AI, it resolved whether a levered book can survive a 17% four-day factor drawdown. It cannot. As Aswath Damodaran wrote in his August 10 post-mortem, leverage "can shrink your time horizon by forcing you to liquidate your mispriced investments before the market corrects." His summary of the fee structure: that 2-and-20 is "an abomination that encourages reckless risk taking", is a separate argument, but it points the same way.
Except in one respect, where it did. The essay's own mechanism predicts this kind of volatility and Aschenbrenner appears not to have drawn the inference. If the story is that physical inputs are the constraint, then the entire equity thesis is a bet on a shortage, and shortages are self-liquidating. SanDisk's 84.6% gross margin and SK Hynix's 76% operating margin are the sound of a bottleneck, and they are also the loudest possible signal to every fab operator on earth to add capacity. HBM expansion and CXMT's domestic Chinese supply are that capacity arriving. The thesis was right about where the rent would accrue in 2025 and 2026. It says almost nothing about who keeps it in 2028.
And the reflexivity problem is real. The fund owned 2-3% of names like Nebius and Bloom Energy. A levered buyer that large in a small float is not just observing the price, it is setting it. Some portion of the 2026 melt-up in these names was Aschenbrenner's own bid. That does not invalidate the thesis, but it means the price action that appeared to confirm it was partly endogenous.
The bounce is not proof either. The cleanest counter-argument comes from the Archegos precedent: the media and streaming names dumped in that 2021 liquidation underperformed for years afterward even as the broader market rose. A rebound off a forced-selling low tells you the seller is gone. It does not tell you the buyer was right.
What it means for public AI equity fundamentals
Strip out the personality and four things are left standing.
- Demand-side evidence is now contractual, not narrative. CoreWeave's $104 billion of remaining performance obligations is a signed backlog, not a TAM slide. Nebius at +454% and Applied Digital at +407% are revenue, not bookings. Whatever else July proved, it did not put a dent in the contracted demand curve. This is the single biggest difference between 2026 and 2000.
- The risk has migrated from demand to supply and financing. The bear case is no longer "nobody wants this." It is margin compression as memory capacity lands, and a cost of capital that has moved against every one of these balance sheets. The 30-year at 5.25% is not a footnote for companies that fund GPUs and substations with debt and are still net lossmaking, Nebius, CoreWeave and Applied Digital all are. Watch memory ASPs and credit spreads, not AI headlines.
- Positioning is now a fundamental. These are small floats, heavily optioned, held by momentum-sensitive capital. The July crash happened with the index calm. That will happen again, and the trigger will again be flows rather than news. Anyone underwriting these names on fundamentals alone needs to size for a 40% drawdown that has nothing to do with the business.
- The differentiation trade has begun. Through 2025 the whole complex traded as one. It no longer does: SanDisk is up 591% YTD and Applied Digital is up 27%, and CoreWeave, despite the strongest backlog disclosure in the group, sits below its June level. Contracted revenue, gross margin durability and balance-sheet capacity are starting to separate these names from one another. That separation is the healthiest thing to come out of July.
Where it leaves him
Situational Awareness is not winding down. Assets are around $10 billion, mostly private: the Anthropic stake bought at roughly a $60 billion valuation in February 2025 and now marked against a $965 billion Series H, plus MatX and Fluidstack. Anthropic could go public as soon as October. Redemptions have reportedly been light, which the LP base explains: family offices and founders, not institutional allocators with committees.
The fund never stopped investing. It sold 11.5 million Core Scientific shares on August 3 for about $226 million, and on August 4 put $400 million into an undisclosed private company, following $100 million the month before. Bloomberg reported on August 7 that investors were queuing to add capital and that Aschenbrenner had told them the fund is not taking new money for now.
His own words to LPs at the end of July: "We took the steps that were necessary to fight another day. But our fund must always be structured such that we can take a loss and fight another day."
That is the correct lesson, and it is a lesson about structure, not about AI. The thesis of Situational Awareness has not been falsified and it has not been vindicated. What was falsified is the idea that a ten-year view can be financed with overnight money. The stocks recovered in eleven days. The fund could not wait eleven days. Both of those facts are about leverage, and neither is about whether the machines are getting smarter.
Disclaimer. Informational only. Not investment advice.
