The Situational Awareness Implosion, Explained With Data
A concentrated AI fund went from +439% to selling its entire book to Citadel in six trading days. The July 30 tape, where stocks the fund never owned bounced as hard as the ones it dumped, is some of the cleanest public data on forced selling you will ever see.
In six trading days at the end of July 2026, one of the best-performing hedge funds in the world went from reporting a +439% first half to selling its entire public book to Citadel: longs and shorts, in one negotiated block. The fund was Leopold Aschenbrenner's Situational Awareness. The story has been told as personality drama, as an AI-bubble morality play, and as a Wall Street cautionary tale.
This post tells it with data, because the data version is more useful. Buried in the July 30 tape is one of the cleanest natural experiments you will ever get on a question every investor eventually faces: when a stock drops 40%, how do you know whether something changed about the business, or just about who was holding it?
The setup
Situational Awareness launched with roughly $225M, with backers including Patrick Collison, Nat Friedman, and Daniel Gross. The book was a concentrated expression of one thesis: long AI "picks and shovels" (CoreWeave, IREN, Core Scientific, Bloom Energy, SanDisk, Nebius, SK Hynix, Micron, Applied Digital, and power names) against software shorts and put hedges on SMH and Nvidia.
Three numbers describe the risk profile better than any narrative:
- Top five positions: more than 76% of the disclosed long book
- Gross leverage: roughly 4x
- H1 2026 net return: +439%
The six trading days
| Date | What happened |
|---|---|
| Jul 17 | Meta announces a commercial compute/cloud offering. The GPU-cloud complex de-rates; CoreWeave is hit worst |
| Jul 24 | Investor letter: +439% H1, "best buying opportunity in over a year," invitation for new capital |
| Jul 25–29 | Core longs fall 27–54% on the month. The Kospi drawdown takes SK Hynix down roughly a third. Hedges fail in both directions: the puts underperform the longs, and the software shorts rally. Goldman Sachs, JPMorgan, and Bank of America issue margin calls |
| Jul 30, pre-open | The entire public book, longs and shorts, goes to Citadel in a single negotiated block, reported between $16B and $20B, at a discount |
| Jul 30, intraday | The dumped names V-bounce: Nebius +27%, IREN +26.5%, Bloom Energy +25.6%, SanDisk +24.6%, CoreWeave +22%, against SPY +1.8% |
| Jul 31 | LP letter: "We let you down this month." July return roughly −67%, still around +80% YTD. Reported assets fall from as much as $45B to roughly $10B |
| Aug 4 | First rebuild move: $400M into an undisclosed Sequoia-backed private company |
Why it broke: three mechanical failures
Leverage turned a drawdown into an ending. At 4x gross, a 27–54% monthly decline in the core longs is not a bad month. It is a solvency event. The fund's own July 31 letter conceded it "came closer to permanent capital impairment than is acceptable to us." Concentration removed the escape hatches. With more than three-quarters of the long book in five names, there was nothing uncorrelated to sell to meet margin calls. Every dollar of forced selling went into the same handful of tickers, which is exactly why those tickers kept falling, which triggered more margin calls. The doom loop was structural. The hedges were correlated with the longs. This is the subtle one. On paper, the fund was hedged: index and Nvidia puts against the longs, software shorts against the AI exposure. In the actual July selloff, the specific long book fell far more than SMH or Nvidia (so the puts paid out less than the longs lost), while software, the short leg, rallied. Both hedges lost money at the moment they were needed. A hedge that is positively correlated with your book under stress is not a hedge; it is more exposure wearing a costume.July 30: the tape that explains everything
Here is the part worth committing to memory. On July 30, with the forced seller finally out, the dumped names bounced violently:
| Ticker | Jul 30 move | In the SA book? |
|---|---|---|
| Cipher Mining | +28% | Never owned it |
| Nebius | +27% | Yes, dumped |
| IREN | +26.5% | Yes, dumped |
| Bloom Energy | +25.6% | Yes, dumped |
| SanDisk | +24.6% | Yes, dumped |
| CoreWeave | +22% | Yes, dumped |
| SPY | +1.8% | Benchmark |
If the July decline had been about fundamentals, with the market rationally repricing the AI capex cycle name by name, then the removal of one fund's supply should have mattered only for the names that fund actually owned. Instead, the entire category had been repriced by the presence of a forced seller, and the entire category snapped back when the selling stopped. That includes names connected to the story only by sector membership.
That is what a flow-driven move looks like: the selling pressure sets the price, and fundamentally unrelated names trade as one block. SpotGamma's write-up of the day is the best single source on this mechanic.
Thesis versus expression
The uncomfortable epilogue for anyone hoping this was a story about AI fundamentals cracking: the same week the fund was liquidating, Azure reported +43% growth and Nvidia's commentary remained supply-constrained. The post-mortem consensus across CNBC, Bloomberg, and most practitioner commentary is that the thesis did not break in July. The expression broke: 4x leverage, five names, correlated hedges.
Two more context points sharpen the lesson. First, this happened against record US margin debt of roughly $1.5 trillion, up 77% since April 2025. Situational Awareness was the most concentrated expression of a leveraged market, not an anomaly in an unlevered one. Second, the fund's surviving asset was its private book, anchored by an Anthropic stake worth around $5B (TechCrunch covered the stake), entered at a ~$60B valuation and marked against a May 2026 round at $965B. Public marks destroyed the fund while private marks saved it. Whether that says something profound about where AI value lives, or just something about how rarely private positions get marked, is one of the genuinely open questions of this episode.
A note on the numbers
Outlets disagree on some figures: peak assets are reported as both $20B and $45B (likely the difference between NAV and gross exposure), and the Citadel block as both $16B and $20B. I have used ranges where the sources diverge. CNBC and Bloomberg are the primary reporting; much of the rest is aggregation.
What this means for how you read the next selloff
The July 30 tape is a rare gift: a same-day, side-by-side comparison of forced-flow pricing and fundamental pricing across a whole sector. Most selloffs never give you that clean a signature. Which raises the practical question: when the next 40% drawdown arrives without a convenient natural experiment attached, what data would you check to tell a positioning event from an information event?
That question has a reasonably systematic answer, and it is the subject of the next post in this series. The short version: price can tell you that something happened, but only fundamentals-versus-valuation data can tell you what. A stock that has fallen 40% while its earnings trajectory is unchanged is in a very different situation from one that fell 40% because estimates fell 40%, and comparing a multiple to its own history is how you see the difference. That comparison (current multiple against the company's own valuation history, next to its fundamental trajectory) is exactly what we are building StockResearch to make self-serve, starting with the comparison and valuation tools live today.
Jake is the founder of StockResearch.app, where he writes data-first research on valuation dislocations. This article is for informational purposes only and does not constitute financial advice. Figures are drawn from the cited reporting and may be revised. Nothing here is a recommendation to buy or sell any security. Always do your own research before making investment decisions.