In this article
← All articles
Investment allocation

Aschenbrenner's Q3 2026 13F: nine AI names and an incomplete risk picture


Executive summary

Situational Awareness's 30 September 2026 disclosure still centres on AI infrastructure, but its $1.58bn reported value cannot tell you what the fund earned or how much investor capital was at risk. The 13F filed on 16 November reports calls at the value of the underlying shares. Those calls make up 66.2% of the total; their purchase prices and terms are missing.

The clearest performance finding is hypothetical. If held unchanged from 30 September to 16 November, the five disclosed share positions would have lost 61.4%, against 2.5% for MSCI World net. The shares underlying the calls fell 44.8%; that is not the calls' return. The filing cannot show whether the manager kept these positions during the fall.

Copio's assessment: the disclosed public positions share substantial AI risk, while the evidence does not establish that the option structure improves long-term compounding. A fully paid call limits the loss on that instrument to its premium. That does not establish the fund's total borrowing, protection or diversification, because cash, shorts and private assets are absent from the disclosure.

For someone investing through a neobroker, the useful check comes before copying any ticker: add up the holdings that depend on the same spending cycle, including those inside ETFs. Then test the joint loss against money you may need to withdraw. Our first study of Situational Awareness covers the June positions and the reported July unwind; this follow-up examines what the September filing adds.

What is in the fund?

Form 13F is a regulatory snapshot of certain holdings at quarter-end, rather than a complete fund account. The adviser's Form 13F-HR for 30 September 2026 lists 23 lines in nine issuers, with a reported value of $1,577,454,028. Five lines are shares worth $532.7m. Eighteen are calls on four underlyings, reported at $1,044.8m, and there are no puts. The 30 June filing had 26 lines in 24 issuers, worth $20.24bn, of which calls were 0.3%.

Situational Awareness 13F at 30 September 2026, reported value of $1,577m: SharonAI shares $440.7m (27.9%), Bloom Energy calls $389.2m (24.7%), Nebius calls $332.6m (21.1%), Sandisk calls $182.7m (11.6%), Roundhill Memory ETF calls $140.3m (8.9%), CleanSpark shares $38.9m, IREN shares $21.6m, Bitdeer shares $20.3m and Fermi shares $11.2m.

30 September 2026: SharonAI shares $440.7m (27.9%); calls on Bloom Energy $389.2m, Nebius $332.6m, Sandisk $182.7m and the Roundhill Memory ETF $140.3m; CleanSpark, IREN, Bitdeer and Fermi shares $92.0m together. Source: 13F-HR information table, 30 September 2026; every line for 30 June and 30 September (CSV).

The three largest issuers make up 73.7% of reported value. Memory exposure through the Sandisk and memory-ETF calls adds up to $323.0m, or 20.5%. These percentages describe the filing's mixture of shares and option underlying values. They are not weights in the fund's net assets and cannot measure how much investor capital each position could lose.

What the call figures mean

Every call line equals its number of underlying shares times the 30 September closing price: Bloom Energy at $276.98, Nebius at $235.88, Sandisk at $1,739.89 and the memory ETF at $60.36. The SEC's 13F guidance (question 44) says most column entries for an option "refer to the underlying security, rather than to the option itself". So $1,044.8m is the value of the shares the calls refer to, not the premium paid or the calls' worth.

A call gives its buyer the right to buy shares at a set price before or at expiry, depending on its terms. The premium is the price paid for the option; a fully paid long call can lose that entire amount. Its percentage gains and losses can be much larger than the share's, even without borrowing to buy the option.

The filing omits the strike price, expiry and premium. It also omits delta, which measures how sensitive the option's value is to the share price. Without those inputs, $1,044.8m of underlying value cannot be translated into capital invested, a maximum portfolio loss or an effective share exposure.

What changed since 30 June

Only two issuers are new: Fermi, with 2,671,898 shares, and the Roundhill Memory ETF, held only through calls. SharonAI, CleanSpark, IREN and Bitdeer were all in the June book. CNBC's 16 November report presents the five share stakes as positions taken as the fund got back on its feet; for four of them, the filings show resized June holdings and cannot say whether they were kept throughout or sold and bought back.

SharonAI is the clearest case. The June 13F lists 5,396,127 shares, and a Schedule 13D for 27 August reports the same 5,396,127 plus pre-funded warrants for 2,674,823 more, 21.1% of the class. A Schedule 13D/A filed on 17 September reports that the fund exercised those warrants on 15 September, and the September 13F lists the sum, 8,070,950 shares.

CNBC wrote on 16 November that the summer's near-collapse forced Aschenbrenner "to sell all of its public equity positions" to Citadel. The August 13D reports no other common-share transactions in its preceding 60-day window apart from a June warrant exercise, and the 13D/A reports none since the August filing, qualifying that account for SharonAI.

The 13D says the fund bought the stake from the company under a purchase agreement of 17 June 2026, with resale depending on a registration statement the company agreed to file. It states: "The amount of funds used in purchasing the Issuer's Class A Ordinary Common Stock was $523,882,863.18." The 30 September value of $440.7m was 15.9% below that cost; at the 16 November close the same shares were worth $162.5m, 69.0% below it.

Seventeen June issuers are absent from the September table, including Micron, CoreWeave, Core Scientific and TSMC. Nebius, Sandisk and Bloom Energy appear as calls rather than shares: June's 4,464,260 Nebius shares compare with calls on 1,410,000 in September, 2,495,344 Sandisk shares with calls on 105,000, and 6,272,808 Bloom shares with calls on 1,405,000. The IREN, CleanSpark and Bitdeer share counts fell by 94.4%, 75.4% and 43.8%.

What the filing cannot show

The SEC's guidance excludes short positions (question 41), options the manager writes (question 43) and shares listed only outside the US (question 40). So the filing cannot say whether the calls are halves of spreads, or whether the fund still runs short positions like the software shorts covered in our first article. It shows no cash, borrowing, private stakes or trade dates. CNBC reported on 16 November that assets fell from a $45bn peak to about $10bn, "consisting mostly of private stakes"; no filing confirms either figure.

One inconsistency is unexplained. A Schedule 13G/A filed the same day reports 8,010,000 Nebius shares, 3.4% of the class, beneficially owned at 30 September. The 13F lists calls on 1,410,000 Nebius shares and no Nebius share line. Neither filing explains the gap, so we use the 13F lines and do not treat the 8,010,000 shares as a confirmed holding.

Why does Aschenbrenner use this structure?

No filing explains the September book. The SharonAI 13D repeats the standard purpose that the shares were bought because the stock "was undervalued and represented an attractive investment opportunity". The wider rationale is Aschenbrenner's June 2024 essay, Racing to the Trillion-Dollar Cluster, which says "many trillions of dollars will go into GPU, datacenter, and power buildout before the end of the decade."

The September names map onto that essay's bottlenecks. It calls power "the single biggest constraint on the supply-side"; CNBC describes Bloom Energy as a "data center power play". It says HBM memory and advanced packaging "are already key bottlenecks for the current AI GPU scaleup", and Sandisk and the Roundhill Memory ETF are memory holdings. CNBC calls SharonAI and Nebius neoclouds, and CleanSpark, IREN, Bitdeer and Fermi "AI-centric plays".

The explanation for the calls is second-hand. Hedge Fund Alpha, summarising Aschenbrenner's investor letter on 31 July 2026, wrote that the public book is "now run on a fully-paid-for, unlevered basis". Fully paid options can avoid an instrument-level borrowing obligation while retaining substantial sensitivity to share prices. The September 13F cannot verify the fund's overall financing arrangement or why the manager chose these particular calls.

Historical performance versus MSCI World

No defensible fund-versus-MSCI World comparison is possible for the third quarter of 2026. The sources reviewed provide no complete fund value or return series, and none of the filings contains one. The only fund returns we found are two interim, unaudited figures for July reported by Hedge Fund Alpha, examined in our first article. Without a return series, we cannot calculate the fund's compound annual growth rate (CAGR), or its Sharpe and Sortino ratios measuring return relative to risk, for a period covering the third quarter.

The 92.2% fall in reported 13F value is not a performance figure. It can reflect position changes, withdrawals or transfers, price movements and the switch between shares and options reported at underlying value. For scale only, MSCI World net in US dollars rose 1.9% from 30 June to 30 September, while SharonAI's share price fell 35.5%. Those price changes do not establish the return on a continuously held fund position.

What can be measured is narrower: what the positions disclosed for 30 September did afterwards, at fixed share counts. The chart shows the five share lines, the shares underlying the calls, and MSCI World net on the same dates.

Hypothetical buy-and-hold of the positions disclosed for 30 September 2026. Five share lines: -47.9% at 16 October, -55.2% at 30 October, -62.7% at 5 November and -61.4% at 16 November. Shares underlying the calls: -24.5%, -29.6%, -44.9% and -44.8%. MSCI World net: +0.1%, -0.8%, -2.8% and -2.5%.

Hypothetical, 30 September to 16 November 2026: five share lines -61.4% ($532.7m to $205.5m); shares underlying the calls -44.8% ($1,044.8m to $576.4m); MSCI World net -2.5%. Source: 13F-HR for 30 September 2026, Yahoo Finance daily closes and MSCI World net USD; daily path (CSV), price inputs (CSV).

Every underlying fell: SharonAI 63.1%, IREN 59.9%, Bitdeer 57.1%, Nebius 54.6%, Sandisk 53.5%, CleanSpark 51.1%, Fermi 41.1%, Bloom Energy 36.0% and the memory ETF 34.9%. Our closes reproduce CNBC's October figures: SharonAI -58.5%, IREN -41.4% and Nebius -34.7%.

The middle bars are the underlying shares, not the calls. An option's return also depends on its strike, remaining life and market pricing. The underlying shares' losses cannot establish how much the calls lost; an option can expire worthless even if its underlying company remains in business.

Sharpe and Sortino ratios over 33 trading days of a hypothetical basket would describe one sell-off, not the fund, so we do not report them. The filing arrived 47 days after the date it describes. A reader who saw it on 16 November was looking at positions whose shares had already lost between 34.9% and 63.1% since that date.

Copio's reading

Is it optimal for geometric growth?

The evidence does not establish a growth advantage for the disclosed structure. We assess expected log wealth, meaning growth that compounds through gains and losses, over several years without forced sales. The relevant alternatives are the same investment thesis at a smaller size, held through shares, or combined with paid protection. Each would need to meet the same funding and withdrawal requirements.

First, does the expected return justify the concentration? The risk is visible: the shares behind the September positions had 105% annualised daily volatility in the third quarter, calculated retrospectively at fixed share counts. The earlier upside was substantial too: our first article found that the six largest June longs rose 287% in the first half at fixed counts. Neither historical result supplies a reliable expected return for the September options.

Second, do the calls improve the whole portfolio after their cost? A fully paid long call caps that instrument's loss at its premium, but can lose all the money committed to it. Whether that trade-off supports faster compounding depends on the premium, position size, expiry and the assets held alongside it. The 13F does not provide enough information to compare it with holding the shares or reducing the exposure.

Third, what reaches investors after fees and withdrawals? The 13D mentions fees and allocations "based on assets under management and realized and unrealized gains", without public fee terms. Investor cash flows and the complete funding structure are also missing. The needed evidence is a whole-fund return series after costs, alongside feasible alternatives with the same funding constraints. A headline gain, a lower loss on one instrument or a larger number of tickers cannot replace that comparison.

Is it crash aware and protected?

The disclosed positions remain exposed to a shared AI downturn; whole-fund crash protection is unknown. The September filing lists no puts, against $8.46bn of underlying put value in the 31 March filing. Puts give their buyer a right to sell at a specified price, subject to their terms. Their absence from this table does not exclude other hedges, including shorts that a 13F omits.

Liquidity also matters for surviving a fall. The disclosed SharonAI position is 8,070,950 shares, equivalent to 5.6 days of the stock's entire average daily volume in September and 3.9 days in October. That is a warning about exit capacity, not an estimate of how long a sale would actually take or its price impact. The share price fell 63.1% from 30 September to 16 November.

The hypothetical share lines bottomed at -62.7% on 5 November. Recovering from the 16 November level would take a 159% rise, and 81% for the shares underlying the calls; our recovery maths article explains why rebounds are earned on a smaller base. Our study of put hedging budgets shows what protection that pays in such falls costs in ordinary years.

Is it diversified without sacrificing growth?

The disclosed public names share substantial economic risk; the filing cannot establish diversification across the entire fund. A correlation measures how closely two return series move together, with 1 meaning perfect positive co-movement. In the third quarter, the nine underlying shares had correlations between 0.40 and 0.88, averaging 0.64. From 1 October to 16 November the range was 0.36 to 0.96 and the average 0.74. Excluding Fermi, the other eight correlated 0.75 to 0.96 with each other.

Correlation matrix of daily price returns, 1 July to 30 September 2026, for SharonAI, Bloom Energy, Nebius, Sandisk, the Roundhill Memory ETF, CleanSpark, IREN, Bitdeer and Fermi. Pairs range from 0.40 to 0.88, average 0.64. Sandisk with the memory ETF 0.88; CleanSpark with Bitdeer 0.82; SharonAI with the memory ETF 0.80; Fermi is the least correlated, 0.40 to 0.56.

1 July-30 September 2026, 64 days: pairs 0.40 to 0.88, average 0.64; Sandisk with the memory ETF 0.88, CleanSpark with Bitdeer 0.82; Fermi 0.40 to 0.56 with the others. Source: Yahoo Finance daily closes; correlation matrices for both windows (CSV). These are observed correlations of the underlying shares from short samples of 64 and 33 daily returns. They do not establish future relationships or justify relying on the second decimal.

Fermi, the least correlated name in this group, still fell 41.1% after September. Imperfect correlation does not promise an offsetting gain in a crash. Equally, two holdings can diversify one another while both fall: the relevant question is whether their combination improves whole-portfolio growth after costs, rather than whether each chart looks different.

CNBC reported on 16 November that the roughly $10bn left after the July sale consisted mostly of private stakes, and Hedge Fund Alpha's 31 July report names a multibillion-dollar Anthropic stake among them. Those reports raise a further question about shared AI exposure outside the public positions, but cannot establish its size or behaviour. Our falling-together study shows the same rise in correlations across asset classes in bad months.

Copio does not grade this portfolio in its checker, because nine single names, call options and undisclosed private stakes cannot be represented by an ETF replica.

What a household can take from it

Start with the economic exposure behind the names in your account. A chip fund, a power supplier and a data-centre company can depend on the same investment budgets. Count the same companies inside broad ETFs as well as direct shares before deciding that another ticker improves diversification.

  • Measure the joint loss. Apply a stated fall to the related holdings together and compare the portfolio impact with accessible savings and dated commitments.
  • Understand the funding. Record any borrowing, collateral demands and option expiry dates, including whether losses could require additional cash.
  • Compare the growth case. Ask what expected return supports the concentration and whether a smaller exposure could meet the same objective after costs.
  • Price any protection. Compare its cost in ordinary markets and payoff in stressed markets across the whole portfolio, without assuming that a milder fall means a higher CAGR.

Our leveraged ETF study explains how exposure and the sequence of returns interact. The practical objective is to keep long-term capital compounding within the household's real funding needs. Neither this filing nor a retrospective sizing formula supplies the right exposure for an individual reader.

Our analysis of Dan Loeb's Third Point reviews another fund's Q3 2026 13F, which added Meta, Nvidia and Microsoft, and tests its record against MSCI World since 2008.


Sources and methodology

Checked on 17 November 2026. Filings come from SEC EDGAR for Situational Awareness LP, CIK 2045724: the 13F-HR for 30 September 2026 (accession 0000935836-26-000610, filed 16 November), the 13F-HR for 30 June 2026 (0000935836-26-000418), the SharonAI Schedule 13D (0000935836-26-000468, event date 27 August 2026) and 13D/A (0000935836-26-000500, filed 17 September) and the Nebius Schedule 13G/A (0000935836-26-000604).

The Form ADV on file is still the 11 June 2026 amendment, unchanged since our first article. Essay quotations come from Racing to the Trillion-Dollar Cluster: the chapter summary, the Power section and the Chips section. The SEC's Form 13F FAQ, questions 40, 41, 43 and 44, defines what the filing includes.

Press figures are attributed to outlet and date and are not confirmed by any filing: CNBC, 16 November 2026, Hedge Fund Alpha, 31 July 2026 and an Investing.com post of 16 November 2026 reporting a 92% fall to $1.58bn, which matches our calculation of 92.2%.

Shares and calls are summed from the information table's value column; the total equals the cover page's $1,577,454,028. For all nine issuers, value divided by shares equals the 30 September close. The hypothetical holds the 13F share counts fixed from 30 September and uses unadjusted Yahoo Finance daily closes; dividends are ignored. MSCI World is the net total return USD index (990100), daily levels from MSCI.

Volatility is the annualised standard deviation of 64 daily returns of the whole book's underlying shares at fixed counts, 1 July-30 September 2026. Correlations are Pearson coefficients of daily price returns. SharonAI volume is the Yahoo Finance daily average for each month.

The 13F values describe positions at quarter-ends, not returns or investor capital, and options appear at underlying value. The fund's NAV, fees, borrowing, shorts, option terms and private valuations are not public. Every calculated figure can be reproduced from the CSVs linked above with the formulas stated here. This article is educational, not personalised investment advice.

FAQ

What did Situational Awareness hold at 30 September 2026?

Its 13F filed on 16 November 2026 lists 23 lines in nine issuers, with a reported value of $1.58bn. SharonAI shares are the largest line at $440.7m. Calls on Bloom Energy, Nebius, Sandisk and the Roundhill Memory ETF make up 66.2% of the value, and small share stakes in CleanSpark, IREN, Bitdeer and Fermi make up the rest.

Did Aschenbrenner's fund lose 92% in the third quarter?

The filing does not show that. The reported 13F value fell 92.2%, from $20.24bn at 30 June to $1.58bn at 30 September 2026, but that figure covers only long US-listed positions at quarter-end. It moves with sales, transfers, withdrawals and the switch from shares to option notional, and it leaves out cash, shorts and private stakes.

Why are the call options valued so high in the 13F?

Because a 13F reports options at the value of the underlying shares. Each call line in the September filing equals the number of underlying shares times the 30 September closing price, for example 1,410,000 Nebius shares at $235.88. Neither the premium paid nor the calls' market value is disclosed.

How did the disclosed positions do after 30 September 2026?

Badly, if they were held. At fixed share counts, the five share lines fell 61.4% from 30 September to 16 November 2026 and the shares underlying the calls fell 44.8%, while MSCI World net fell 2.5%. This is a hypothetical buy-and-hold of the disclosed positions, not the fund's result: the 13F does not show what was traded in October.

Can an investor copy Situational Awareness from this 13F?

Not usefully. The filing arrived 47 days after the quarter-end it describes, and over those days the nine underlying shares fell between 34.9% and 63.1%. It also omits option strikes, expiries and premiums, any shorts and the private stakes that CNBC says make up most of the fund.

Optional cookies help us measure visits and advertising. Choose each purpose, or decline both. You can change your choice using Cookie settings in the footer. Cookie policy.