The fund, from the filings
Situational Awareness LP is a Delaware limited partnership formed in 2024 and based in San Francisco, serving as investment adviser to Situational Awareness Partners LP and its offshore feeder. Leopold Aschenbrenner, formerly of OpenAI’s Superalignment team, is Chief Investment Officer; Carl Shulman is Director of Research. Its Form ADV filed 27 April 2026 reports discretionary regulatory assets under management of $9.28bn as of 31 December 2025.
The stated objective is exposure to the development of artificial intelligence “with a significant focus on the impact of the development of artificial general intelligence.” The mandate is unusually explicit about what follows: the adviser “engages in hedging (through short sales and otherwise) and uses leverage through borrowing on margin and investing in swaps, options, futures, and other derivatives,” with — in the ADV’s own words — “no limits on the types of securities, positions, concentration, or leverage” beyond what the offering documents specify. Concentration and leverage were not a drift from the strategy. They were the strategy, disclosed in advance.
The Q1 2026 Form 13F, filed 15 May 2026 for the period ended 31 March, discloses 42 entries totalling $13.68bn. Its shape is the first surprise, because it is not the shape the coverage described:
| Sleeve | Value $m | % of filing | Lines |
|---|---|---|---|
| Puts | 8,459.1 | 61.85% | 11 |
| Equity long | 3,855.8 | 28.19% | 26 |
| Calls | 1,361.8 | 9.96% | 5 |
| Total | 13,676.7 | 100.00% | 42 |
A book that is 62% puts by disclosed value is not a long portfolio with an overlay. It is a two-sided book whose disclosed short side was more than twice its disclosed long side. Whatever else went wrong here, “they forgot to hedge” is not it.
The put book — 61.9% of the filing
| Underlying | Company | Notional $m | % filing | vs 29 Jun, at trough | vs 29 Jun, at 31 Jul | Shift (pts) |
|---|---|---|---|---|---|---|
| SMH | VanEck Semiconductor ETF | 2,042.7 | 14.94% | −20.2% | −14.5% | +5.7 |
| NVDA | Nvidia | 1,568.3 | 11.47% | −2.5% | +0.0% | +2.6 |
| ORCL | Oracle | 1,072.9 | 7.84% | −20.3% | −12.1% | +8.2 |
| AVGO | Broadcom | 1,006.2 | 7.36% | −0.6% | +4.5% | +5.1 |
| AMD | Adv. Micro Devices | 969.2 | 7.09% | −20.4% | −11.7% | +8.6 |
| MU | Micron — also held calls | 583.7 | 4.27% | −31.0% | −23.2% | +7.8 |
| TSM | TSMC — also held calls | 535.1 | 3.91% | −17.7% | −11.2% | +6.5 |
| ASML | ASML Holding | 494.1 | 3.61% | −17.7% | −13.5% | +4.2 |
| INTC | Intel | 159.1 | 1.16% | −37.8% | −31.5% | +6.3 |
| GLW | Corning | 21.0 | 0.15% | −51.5% | −45.9% | +5.6 |
| INFY | Infosys | 6.8 | 0.05% | +17.8% | +12.4% | −5.3 |
| Put book | 11 lines | 8,459.1 | 61.85% | Notional >15% below base: $5,878m → $764m |
Two things stand out. The best-performing hedge in the entire book was its second-smallest: the Corning put, on an underlying down 51.5% at the trough, carried $21m of notional — one thousandth of the SMH line. And the largest lines were pointed at the calmest names. Nvidia and Broadcom together are $2.57bn of puts, 18.8% of the whole filing, struck on underlyings that were down 2.5% and 0.6% respectively when the infrastructure longs were down 23% to 60%. By Friday, Nvidia was flat on the month and Broadcom up 4.5%.
The long book — 28.2% of the filing
| Position | Company | Value $m | % filing | Jul peak→trough | Trough→31 Jul |
|---|---|---|---|---|---|
| BE | Bloom Energy — power | 878.7 | 6.42% | −44.5% | +25.7% |
| SNDK | SanDisk — storage | 724.4 | 5.30% | −50.0% | +19.6% |
| CRWV | CoreWeave — GPU cloud | 556.1 | 4.07% | −32.4% | +18.0% |
| IREN | IREN — miner→HPC | 401.0 | 2.93% | −33.2% | +25.6% |
| CORZ | Core Scientific — miner→HPC | 389.1 | 2.84% | −24.5% | +14.3% |
| APLD | Applied Digital — datacentre | 320.0 | 2.34% | −34.6% | +18.0% |
| RIOT | Riot — miner→HPC | 142.2 | 1.04% | −23.9% | +10.6% |
| CLSK | CleanSpark — miner→HPC | 104.5 | 0.76% | −23.1% | +14.6% |
| SEI | Solaris Energy Infrastructure | 62.5 | 0.46% | −40.9% | +18.4% |
| TE | T1 Energy | 43.9 | 0.32% | −59.5% | +12.1% |
| BITF | Bitfarms / Keel Infrastructure | 38.8 | 0.28% | −40.9% | +24.8% |
| BTDR | Bitdeer Technologies | 29.8 | 0.22% | −42.4% | +18.2% |
| PSIX | Power Solutions International | 26.3 | 0.19% | −31.4% | +8.5% |
| WYFI | WhiteFiber | 20.9 | 0.15% | n/a | n/a |
| AMD | Adv. Micro Devices — also held puts | 20.2 | 0.15% | −23.0% | +10.8% |
| BW | Babcock & Wilcox | 19.9 | 0.15% | −37.7% | +17.4% |
| SHAZ | SharonAI — see 13G below | 18.1 | 0.13% | −49.9% | +20.4% |
| PUMP | ProPetro Holding | 13.1 | 0.10% | −25.7% | +8.7% |
| SMH | VanEck Semis ETF — also held puts | 10.3 | 0.08% | −18.7% | +7.2% |
| INTC | Intel — also held puts | 8.9 | 0.07% | −35.5% | +10.2% |
| TSM | TSMC — also held puts and calls | 7.6 | 0.06% | −17.1% | +7.9% |
| HIVE | HIVE Digital | 6.4 | 0.05% | n/a | n/a |
| ASML | ASML — also held puts | 6.1 | 0.04% | −15.9% | +5.1% |
| MU | Micron — also held puts and calls | 5.9 | 0.04% | −28.4% | +11.4% |
| GLW | Corning — also held puts | 0.7 | 0.01% | −43.8% | +11.4% |
| NVDA | Nvidia — also held puts | 0.5 | 0.00% | −10.6% | +2.6% |
| Equity | 26 lines | 3,855.8 | 28.19% | Every priceable line bottomed 29 July |
| Position | Underlying | Notional $m | % filing | Jul peak→trough | Trough→31 Jul |
|---|---|---|---|---|---|
| MU calls | Micron — memory | 422.3 | 3.09% | −28.4% | +11.4% |
| SNDK calls | SanDisk — storage | 388.8 | 2.84% | −50.0% | +19.6% |
| TSM calls | TSMC — foundry | 354.8 | 2.59% | −17.1% | +7.9% |
| CRWV calls | CoreWeave — GPU cloud | 140.6 | 1.03% | −32.4% | +18.0% |
| BE calls | Bloom Energy — power | 55.3 | 0.40% | −44.5% | +25.7% |
| Calls | 5 lines | 1,361.8 | 9.96% |
The positions that are not in the 13F
Two of the fund’s most-discussed holdings barely appear in the 13F, or do not appear at all, and reach the public record through Schedule 13G instead. This matters for the concentration argument, which is best made from ownership percentages rather than dollar values.
| Filed | Form | Company | Shares | Ownership | Jul peak→trough | Trough→31 Jul |
|---|---|---|---|---|---|---|
| 2026-06-29 | 13G | SharonAI Holdings | 5,404,540 | 19.90% | −49.9% | +20.4% |
| 2026-05-27 | 13G | Nebius Group | 12,410,060 | 5.60% | −35.3% | +28.5% |
| 2025-10-14 | 13D/A | Core Scientific | 28,756,478 | 9.40% | −24.5% | +14.3% |
The SharonAI line is the sharpest illustration of the structural problem. Owning 19.90% of a company is not a position you reduce quietly; it is a position you exit by finding one buyer for the whole thing. The same is true, less extremely, of Core Scientific at 9.40% and Nebius at 5.60%. Stakes of that relative size in names of that liquidity have exactly one exit under duress, and on 30 July the fund used it.
The month, day by day
1 Jul — Fund assets peak at a reported $45bn, up from roughly $225m at the late-2024 launch. SanDisk closes at $2,032, its highest close of the month.
10 Jul — SK Hynix’s US-listed shares (SKHY) begin trading, days from the sector top. The line is down 22.8% from its debut within thirteen sessions.
10–24 Jul — The AI-infrastructure complex breaks. Core longs fall 30–60% from their highs. Global long-short funds head for their worst monthly drawdown on record; Asia-focused funds average −18.6% through 28 July.
24 Jul — FT reports the H1 letter: +439% net through June. Aschenbrenner calls the sell-off “some of the most attractive opportunities since early 2025” and invites fresh capital by 1 August. It does not arrive.
27 Jul — Citadel Securities publishes Fed Views: The Case for July, breaking with consensus to call for a surprise 25bp hike. Swap pricing moves toward one-in-three; open interest in August fed funds futures sets a record.
28–29 Jul — Prime brokers including Goldman Sachs, JPMorgan and Bank of America press the fund to raise capital or cut positions. Both sides of the book are marketed. Sequoia and Greenoaks are approached over a slice of the Anthropic stake; talks fail.
29 Jul — 2:00pm ET: the FOMC holds at 3.50–3.75% on a 9–3 vote, with three regional presidents dissenting for a hike — the most one-directional dissents since 2016. The Dow closes down more than 840 points. Every priceable position in the filing bottoms this session.
30 Jul — Pre-open, the bulk of the public book crosses to Citadel in a single block. CNBC’s David Faber breaks it; the WSJ and Reuters name the buyer. Reuters reports Citadel took the portion financed by broker leverage. On the close the cohort reprices sharply higher — IREN +30.5%, Nebius +27.1%, Bloom +26.5%, SanDisk +24.6%. Adobe, reported as a fund short, falls 5.9%.
31 Jul — Reuters reports the fund will hold roughly $10bn after the deal, chiefly the private Anthropic stake, which was not sold. The WSJ reports the portfolio down about 67% in July. The complex gives back a little into the weekend: SanDisk −4.0%, Micron −5.9%, Core Scientific −5.0%.
The Fed on the same afternoon
Any account that attributes 29 July purely to one fund’s margin calls is incomplete. That afternoon was also the most genuinely uncertain FOMC in years. Chairman Kevin Warsh had removed forward guidance from the statement, and the market went into the meeting split: CME FedWatch put the odds of a surprise hike near 38%, and open interest in the August fed funds contract hit a record 967,136 as traders scrambled to hedge the tail.
The Fed held at 3.50–3.75% — but on a 9–3 vote, the most one-directional dissents since September 2016. Warsh’s characterisation was “I asked for a good family fight and I got one.” Markets read the dissents rather than the decision: the Dow fell more than 840 points, the 30-year yield rose over nine basis points to 5.19%, and rate-sensitive equities took the brunt.
That matters here because the fund’s long book was, almost by construction, the most rate-sensitive cohort in the equity market: pre-profit or thinly-profitable, capital-hungry infrastructure names whose valuations are long-duration bets on cash flows years out, several funded by convertible issuance. A hawkish surprise hits that cohort hardest. So 29 July has two sufficient explanations operating at once — a forced seller liquidating into the close, and a macro repricing that would have hit these names regardless.
The arithmetic that did it
The early read on this unwind, ours included, was that the hedges “failed.” The filings show that framing was wrong. At the 29 July trough, ten of the eleven put underlyings were lower than a month earlier, several heavily — Corning −51.5%, Intel −37.8%, Micron −31.0%, AMD −20.4%, Oracle −20.3%. Roughly $5.88bn of put notional, 69% of the put book, sat on underlyings down more than fifteen percent. A book that is 62% puts by disclosed value, with 69% of that notional in the money against the base, is not a portfolio that forgot to hedge.
The two sessions after the block took most of it away. Every put underlying except Infosys rallied into Friday: SMH narrowed from −20.2% to −14.5% against the same base, Oracle from −20.3% to −12.1%, Intel from −37.8% to −31.5%, AMD from −20.4% to −11.7%. Notional sitting more than fifteen percent below the base collapsed from $5.88bn to $764m in two days. Whatever the hedge book was worth on Wednesday afternoon, it was worth a fraction of that by Friday’s close.
What survives of the original point is narrower and more specific. The two largest single lines in the whole filing after the SMH ETF — Nvidia at $1.57bn and Broadcom at $1.01bn, together 18.8% of everything disclosed — were struck on the two large-cap semiconductors that barely moved. At the trough Nvidia was down 2.5% and Broadcom 0.6%, against infrastructure longs down 23% to 60%. Nearly a fifth of the filing was pointed at the two names least capable of paying for what was happening elsewhere in the portfolio.
But the decisive asymmetry was beta, not direction. Measured from their July peaks the sixteen infrastructure longs fell 23–60%; measured against the same 29 June base the put underlyings fell 0.6–52%, half of them clustered between 17% and 21%, with a median decline of 20.2%. A book long high-beta infrastructure and short low-beta large-caps loses on a sector-wide drawdown even when both legs move the “right” way, because the legs do not move by the same magnitude. Add leverage — Reuters reports Citadel took on the portion of the portfolio financed by broker margin — and a 30% decline in the long book compounds well past the equity cushion. The hedge book cushioned; it could not plausibly have covered that.
Concentration then converted a loss into a forced sale. Top-ten concentration across the filing was 72.66%, and the fund’s largest ownership stakes were in companies far smaller than the positions taken in them: 19.90% of SharonAI, 9.40% of Core Scientific, 5.60% of Nebius. Stakes of that relative size cannot be reduced quietly in names of that liquidity. Once the pressure came, a single block was the only exit that existed — and it is why July’s declines ran far beyond anything the fundamentals reported: for three weeks the price action was consistent with the market anticipating a persistent forced seller in exactly these names.
What the tape settled — and what it didn’t
The block structure explains the shape of the recovery. A piecemeal liquidation of a fifth of a float is days of price-insensitive supply; a single transfer to one deep-pocketed buyer moves the position without ever hitting the tape. That is why 30 July was a violent repricing upward: the cohort had carried a liquidation discount for three weeks, and the discount expired the moment the seller did. Two sessions later every priceable line in the equity book is higher — Nebius +28.5% off the trough, Bloom Energy +25.7%, IREN +25.6%, Bitfarms +24.8%, SharonAI +20.4%, SanDisk +19.6%.
The clearest single confirmation is on the short side, and it comes with a caveat that has to come first. Adobe is not in the 13F and cannot be: direct equity shorts are not reportable on Form 13F. That the fund was short it comes from press reporting, not from a filing, and we flag it as such. Taking that reporting at face value, Adobe’s highest close of July was $263.43 — set on 29 July, the exact session in which every long in the filing bottomed. It fell 5.9% the following day as the block cleared and is −4.9% from that peak. A short that stops squeezing on the day the forced coverer is removed, having topped on the day the forced seller capitulated, is about as direct a piece of evidence for the mechanism as a tape can offer.
The bounce was also not confined to the long book. Oracle rallied 8.2 points against its June base in two sessions and Broadcom turned positive on the month — names the fund was short via puts, not long. A recovery spanning both sleeves simultaneously is consistent with July’s move having been about positioning and liquidity across the whole complex rather than anything specific to these particular lines.
Four things remain unsettled. The fundamental question — memory pricing, HBM allocation, the capex cycle — was never adjudicated by this episode and will be decided in guidance; the thesis may yet prove right in an unlevered owner’s hands. The short side now sits with Citadel, and how those positions are worked out will move the software tape. The macro overhang has not cleared: three dissents for a hike, Jackson Hole on 27–29 August and a September meeting with live odds mean the rate-sensitivity that helped break these names in July is still there. And one fund’s exit does not un-crowd a theme — global long-short funds just booked their worst month on record. The forced seller is done. Whether the repricing is done is a separate question, and only the first was answered this week.
It is worth separating the two risks that ran together here, because only one of them proved fatal. Thesis risk is the possibility that AI build-out slows, that memory pricing rolls over, that the neoclouds never earn their cost of capital. Nothing in July adjudicated that question — and the recovery once the seller cleared suggests the market did not think it had been adjudicated either. Structure risk is the possibility that you are correct and still cannot survive the path: broker-financed leverage, positions running to a fifth of a company, the largest hedges struck on names with a fraction of your longs’ beta, and prime brokers with the contractual right to decide when the argument ends. An unlevered fund holding this identical book had a brutal month and would now be sitting on a large but survivable drawdown. The levered version did not get to have a month. The thesis may yet be right; the capital structure was not built to find out.
Sources. Situational Awareness LP Form ADV filed 27 Apr 2026 and Q1 2026 Form 13F (period 31 Mar 2026, filed 15 May 2026), with line items and 13D/G filings as compiled by Fintel and cross-checked against Quiver Quantitative and Nasdaq institutional holdings; totals independently recomputed. SharpeIQ price history for all trough, close and horizon moves. Federal Reserve FOMC statement, 29 Jul 2026. Citadel Securities Global Macro Strategy, Fed Views: The Case for July, 27 Jul 2026. Reuters (Citadel purchase, prime brokers, ~$10bn residual, Anthropic stake retained, Asia long-short −18.6%); CNBC (D. Faber, 30–31 Jul; momentum crash, BTIG and Morgan Stanley data); Wall Street Journal (buyer identified; investor letter, −67% July); Financial Times (H1 letter, +439% net); Yahoo Finance / Blockspace (13F composition). Figures attributed to reporting are not independently verified by SharpeIQ.
For informational and educational purposes only. Not investment advice, and not a recommendation to buy or sell any security. Position, leverage and transaction details are drawn from regulatory filings and public reporting and may be revised or superseded. Options positions in Form 13F are reported at notional value of the underlying and do not indicate realised profit or loss. Price data is close-to-close from SharpeIQ market data and may contain errors or omissions. Past performance is not indicative of future results.