
Situational Awareness LP
Leopold Aschenbrenner (founder, ex-OpenAI Superalignment)
Pure AGI-thesis fund built around Aschenbrenner's 2024 'Situational Awareness' essay — concentrated long bets on the AGI compute/power buildout (neoclouds, power, memory) plus a large private Anthropic stake, originally hedged with puts on the headline semiconductor names it saw as overvalued. After the Jul 2026 forced deleveraging the shorts and the broker financing are reported closed, leaving an unlevered public book alongside the retained private stakes.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
- Self-reported — Headline returns are manager- or firm-reported and unaudited.
- Levered — Leverage is material to outcomes, including on the downside.
Top holdings
Largest disclosed public long (Q1 2026); ~5.6% ownership stake; neocloud GPU compute
Private position, not in the 13F; entered Feb 2025 at a ~$60B valuation
Data-center power thesis; stock ran sharply after the stake was reported
Neocloud / AI compute long
SHORT — ~$1.6B notional put position (Q1 2026) — a hedge against the headline GPU name, alongside puts on Broadcom, Oracle, AMD, Micron, ASML, Intel
SHORT — ~$2B notional put position (Q1 2026) — basket hedge on the semis complex
Recent moves
On or about 30 Jul 2026 the fund sold the bulk of its broker-financed public equity book to Ken Griffin's Citadel (the ~$71B hedge fund, NOT Citadel Securities) in a single block, after prime-broker margin calls on a book reported levered up to ~4x; Goldman Sachs, JPMorgan and Bank of America are named as facilitating brokers and Millennium reportedly bid and lost. Reported by Reuters, Bloomberg and CNBC on unnamed sources — no filing or on-record statement corroborates it, and no source states the price, size or discount. The private book (Anthropic, MatX, Fluidstack) was not part of the sale. Per an investor letter reported 31 Jul 2026, all shorts were closed and portfolio financing removed, leaving a 'fully-paid-for public book'. The fund was RAISING, not returning, capital into the event — a 24 Jul letter invited commitments from 1 Aug 2026, which Bloomberg reports fell short.
Loading disclosed positions…
Our take
The most iconic internet-native fund of this AI cycle — born directly from a viral essay and a Collison/Gross/Friedman backer list. The returns are real but fragile: concentration in one private mark (Anthropic) plus levered neocloud/power longs means a single re-rate down the AI capex curve could reverse the story as fast as it built it. — Update (Jul 2026): that fragility materialised. The fund was forced out of its levered public book by prime-broker margin calls and sold it to Citadel in a block. Note what did NOT happen: the fund was not liquidated and Citadel did not buy it — its SEC adviser registration was ACTIVE with no Form ADV-W when checked 1 Aug 2026.
July 2026 tested the structure rather than the thesis, and the structure failed first.
Read the distinction carefully, because the shorthand circulating is wrong on both halves: the FUND was not liquidated and Citadel did not buy it. Registration was ACTIVE with no Form ADV-W when checked 1 Aug 2026, and every account describes a purchase of POSITIONS — not the LP, the management company or a GP stake. What did change is the vehicle's shape: leverage and shorts gone, a reported ~$10B left, private book (Anthropic, MatX, Fluidstack) retained. The instructive part is that the fund's own financing, not its idea, is what fired. A concentrated levered expression of a long-horizon thesis can be forced out of its position by a margin clerk before the thesis is ever adjudicated — and a manager who is right about 2030 can still be liquidated in July.
Thesis
A pure expression of one idea: AGI arrives via a physical buildout, so own the bottleneck (power, neoclouds, memory) and short the crowded way of expressing the same trade. Launched Sept 2024 by Leopold Aschenbrenner off his own viral essay. In July 2026 the expression broke before the idea did: prime-broker margin calls on a book reported levered ~4x forced the sale of the bulk of the public portfolio to Citadel, leaving a reported ~$10B, no leverage and no shorts. Reported +439% net through Jun 2026, ~-67% in July, ~+80% for the year.
Concentrated long/short AGI-infrastructure book. Long: neoclouds and power suppliers — Nebius (12.41M shares / 5.6% of the company, ~$2.6B, via a passive Schedule 13G filed late May 2026, separate from the Q1 13F), Bloom Energy, CoreWeave, IREN, Core Scientific — plus a private Anthropic stake at ~20% of the fund. Short: ~$8.5–8.7B of put notional on chips: SMH $2.04B, Nvidia $1.57B, Oracle $1.07B, Broadcom $1.01B, AMD $969M (Q1 2026 13F).
- 2026-07Accounts diverge on scope. The 30 Jul 2026 letter says a portion of the public portfolio was traded and the fund now runs a fully-paid-for public book, long stock and long paid-up options; CNBC reports all of it went.
Assessment
- Long the physical bottleneck, short the consensus semi expression was a non-obvious frame. We had called it 'a spread trade, not levered AI beta' — July 2026 falsified that half.
- Went early and large on the power/neocloud leg (Bloom, Nebius, CoreWeave) before 'picks and shovels' became the consensus AI trade — entry timing was ahead of the crowd.
- Reported terms match the illiquidity — a multi-year lockup against a ~20% private mark — rather than promising liquidity the book could not support.
- The thesis is written down and public in the founder's essay, so the book can be checked against a stated position rather than reverse-engineered from filings.
- Return attribution is thin. A ~20% Anthropic stake entered Feb 2025 at a reported valuation just above $60B against a ~$965B latest mark is roughly 15x — one call, not a demonstrated repeatable process.
- Reported 2/20 would accrue incentive fees on an unrealized private mark the manager neither sets nor can exit at scale — though we could not verify the fee terms.
- Reflexivity: the essay is both the thesis and the marketing. The manager is the most prominent public evangelist for the narrative that revalues his largest holding.
- The hedge question is now settled and it went the wrong way: in July 2026, the month the AI-capex complex actually fell, the put book did not hold the fund up — it lost ~67% and was forced out by margin calls rather than choosing its exit.
- Scale outran the structure. AUM went from a reported $225M to press-cited >$20B in under two years, and the gains were earned on a far smaller base; by Jul 2026 the leverage that amplified that growth is what removed the position.
- Leverage was never in the disclosed book. A ~4x-levered public portfolio is not visible in a 13F, so the risk that actually ended the position could not be read from the filings we critique here.
- 2026-07The semi hedge did pay: SMH -17.6% against a $2.04B put. What cost the book was Nvidia +0.3% ($1.57B put), Broadcom +3.1% ($1.01B put) and the Adobe short, up about 36% from late June to late July 2026.
- 2026-07The ~20% is a portfolio weight, not ownership of Anthropic: the retained stake is reported near $5B on 30 Jul 2026 against the $965B May mark, roughly 0.5% of the company, and about half of a ~$10B fund post-sale.
Record
Two regimes inside one year, and the second is unaudited. Reported +439% net from 1 Jan to 30 Jun 2026, then ~-67% in July, leaving roughly +80% for the year — the July figures being the manager's own current, unaudited estimate relayed via press, not a marked or administrator-verified number. Attribution matters more than the level: the drawdown traces to FORCED deleveraging on a broker-financed book, not to a re-rating the manager chose to sit through, so it measures financing structure rather than selection. And no single AUM figure describes this fund — $9.28B ADV RAUM (11 Jun 2026), $13.7B Q1 13F-reportable, >$20B press (Jun), ~$45B cited peak, ~$10B after the sale are different measures that no source reconciles.
Risks & fit
- Financing risk already fired: reported ~4x leverage drew prime-broker margin calls that forced the Jul 2026 block sale (Reuters). Leverage and shorts are now reported removed.
- Anthropic mark risk: a down round, a flat secondary, or an extended private hold reprices the largest retained position with no exit available.
- Concentration: after the sale the book is reported at ~$10B weighted toward private stakes whose marks are negotiated, not cleared.
- Key-person and short operating history: a manager in his mid-20s, founded 2024, whose first real drawdown ended in a forced sale.
- Capital-raise risk: the fund was soliciting commitments from 1 Aug 2026 into a 67% monthly loss; Bloomberg reports the appeal fell short.
Partly adjudicated, and against the hedge. Our prior test was whether the put book could hold the fund flat in a quarter when AI-capex actually fell; July 2026 was that quarter and the answer was no — the fund fell ~67% and was forced out by margin calls rather than choosing its exit. What remains open is the idea: if the retained private book and an unlevered public book compound from here, the buildout call was right and only its financing was wrong. The cleanest remaining tests are primary — whether the Q2 2026 13F-HR (due 14 Aug 2026) is filed, and whether any ADV amendment or ADV-W follows.
Analytically instructive for anyone learning to read a 13F correctly — this filing is a textbook case of option notional inflating a headline and a private leg hiding the real driver. It is also not accessible: reported terms include a $25M minimum, Qualified Purchaser status and a multi-year lockup (unverified). Its value to a reader here is as a case study in return attribution, private-mark accounting, and capacity decay.
Secondary commentary reports 2/20, a $25M minimum and a ~two-year lockup; unverified against an offering document or Form ADV. The structural question is the base, not the level: incentive fees accrue on an unrealized private mark.