Situational Awareness: A 4x-Levered AI Book Meets the Tape We Measured

1What happened, per the reporting

The reported sequence, each claim attributed: Situational Awareness LP, co-founded by former OpenAI researcher Leopold Aschenbrenner, launched in 2024 with $225M and rode a concentrated, levered long book in AI infrastructure, power, datacentres and bitcoin miners to one of the fastest scalings in hedge fund history - up 439% in the first half of 2026 by Yahoo Finance's account, with peak assets estimated at up to $45B in early July 2026. Reported leverage ran as high as 4x.

Then the July rout - the one quantified in the next section - turned the leverage around. Margin calls forced the fund to sell the bulk of its public equity book to Citadel in a deal closing on or around July 30; Seeking Alpha sized the transferred portfolio at $16B, while earlier flashes circulated smaller figures. What remains, per Bloomberg and TechCrunch: roughly $10B of assets, anchored by a private Anthropic stake of about $5B - first bought at a reported ~$60B valuation in early 2025, marked near $965B by May 2026 - leaving what Bloomberg called essentially an Anthropic holding company with a hedge fund attached.

The arithmetic is brutal and coherent: up to $45B at the peak, a $16B public book handed to a single buyer under duress, $10B left standing - roughly half of it one private position. Where outlet figures still differ at the margin, we note it below rather than resolve it by preference. The core is consistent everywhere: extreme leverage, a July drawdown, a forced transfer of the public book to Citadel, and a surviving private stake.

2The tape it happened into, measured from our bars

Close-to-close, July 22 to July 29, from our own bar store: SPY -2.4%. Against that: Micron -23.0%, Marvell -22.6%, AMD -22.2%, Applied Materials -21.2%, Lam -21.0%, KLA -20.7%, Intel -20.2%, ASML -13.9%, TSMC -11.0%, NVIDIA -10.4%. Over 21 sessions: KLA -38.9%, Lam -38.6%, Intel -37.8%, Marvell -41.2%, Micron -35.5%, against SPY -1.6%.

Before the fund's name attached to the move, the cross-section already said two things. The derating was complex-wide - three semicap franchises within half a point of each other in five sessions is sector-level repricing, not idiosyncratic news. And it hit the SUPPLIERS far harder than the flagship, which is what forced selling of a broad infrastructure basket looks like: a liquidating book sells what it owns, not what the narrative says. The reported composition of the fund's book - AI infrastructure, power, datacentres, miners - maps onto our casualty list with uncomfortable precision.

3What our desks recorded at the transfer

On July 30 - the reported closing date - our flow desk recorded put blocks bought at 100% of ask, all expiring July 31, one day out: TSM $1.2M in 390-strike puts, AMD $2.4M in 490s, MU $1.9M in 830s. Our dark-pool desk recorded institutional size transacting through the derating: an AMD block of $246M, an AVGO block of $207.8M, an AMAT block of $60.2M.

We cannot attribute any specific print to the Situational Awareness transfer - blocks do not sign themselves, and a $16B portfolio moving to Citadel would mostly travel off-tape anyway. What the desks establish is narrower and still useful: institutional size was changing hands in exactly these names on exactly the reported date, and someone was still paying full ask for one-day downside protection after a 20% fall. Our July 29 scanner note on the MU strangle resolved the same way - $3.1M of event volatility bought with spot near $900, and MU closed July 29 at $739, through the put strike by $160.

4The macro squeeze that set the clock

Our macro briefs were pointing at this trade's financing before the name surfaced. July 30, 13:06 UTC: the 30-year at 5.21% and rising, against the brief's own 5.25% stress line - and AI infrastructure is the most long-duration capex trade in the market. Gold +2.46% with copper +2.89% against weak equities - a stagflation-shaped repricing that punishes expensive growth. USD/JPY at 162.92, strengthening from 163.88 - the carry-unwind transmission our earlier note mapped, in which levered books sell their most crowded liquid longs first.

A 4x-levered long book is mechanically short volatility and short rates whatever its equity thesis says. The instruments that were moving - long end, vol, yen - are the ones that call margin, and margin, not narrative, is what sets a forced seller's calendar.

5Thesis versus balance sheet

Here is the part worth keeping after the headlines fade. The fund's core claim - that AI infrastructure is structurally scarce - is the same claim our bottlenecks board tracks in the physical world, and NOTHING in that structure loosened this month: advanced packaging is still two suppliers, HBM three firms, the transformer queue still quotes in years. The thesis and the balance sheet failed on different clocks. At 4x, a one-fifth drawdown in the underlying is close to a total loss of equity in the levered position; the structure being right in 2028 is irrelevant if the margin call arrives in July 2026.

The surviving asset makes the same point from the other side: the Anthropic stake - private, unlevered, unmarginable - is the position that could not be taken from the fund by a fast tape, and it is what the vehicle reportedly reorganises around.

6Situational awareness, the discipline

The name deserves one respectful observation. Situational awareness, as a discipline, means knowing which of your inputs is an instrument and which is a story - and knowing what your own structure does to your time horizon. The fund's story may yet prove right; its structure did not get to wait. The instruments that mattered were not AI benchmarks but the 30-year, the yen, and the margin desk's arithmetic.

For our part: this note's first draft, written twelve hours earlier, measured the entire event from our desks without knowing the fund's name, and said so plainly rather than guessing at a catalyst. The name arrived; the measurements did not change. That is the product working as designed, and it is also the reason the instrument sections above survive contact with tomorrow's corrections to the reported figures.

Sources and method

Fund facts are press-reported and attributed inline, with conflicting figures printed as conflicts; nothing about the fund comes from this desk's own knowledge. Market measurements are ours: the bar store, flow and dark-pool desks, and the macro briefing, each line dated. The first draft of this note measured the event without knowing its name; the name changed the title, not the numbers.

• Press-reported core, consistent across outlets: 4x max leverage, margin calls in the July rout, bulk of public book to Citadel on/around 07-30, ~$10B remaining anchored by a ~$5B Anthropic stake
• Our bars, 5 sessions to 07-29: SPY -2.4% vs MU -23.0%, AMD -22.2%, semicap trio ~-21% each - the reported book composition maps onto the casualty list
• Flow desk 07-30: TSM/AMD/MU put blocks at 100% of ask, 07-31 expiry; dark pool: AMD $246M, AVGO $207.8M, AMAT $60.2M
• Macro brief 07-30: 30Y 5.21% vs the brief's own 5.25% line; gold/copper up against weak equities; USD/JPY 162.92 strengthening
• MU closed 07-29 at $739 vs the $900 put strike in our 07-29 strangle note - the outsized-move premise printed in five sessions

Underlying sources

• Bloomberg, 2026-07-30: Situational Awareness assets fall to $10B after losses
• CNBC, 2026-07-30: Aschenbrenner forced to unwind public stock positions, sources say
• TechCrunch, 2026-07-30: fund sold public portfolio, retains Anthropic shares
• Yahoo Finance / Seeking Alpha, 2026-07-30: $16B equity portfolio to Citadel; 439% H1 return
• TechTimes, 2026-07-30: 4x leverage ends the fund's 1,000% run
• Parallax bar store (closes 07-22 to 07-29); flow and dark-pool desks (07-30 sweeps); macro brief 2026-07-30 13:06 UTC; Parallax notes: MU volatility dislocation (07-29), Anatomy of a Carry Unwind (07-29)

Written by Parallax Research's research writer on 2026-07-30 and reviewed against the dossier it was given. Figures are as of that date and are not updated in place. Our scoring model's out-of-sample AUC is published in the whitepaper; nothing here is a forecast.