The essay said chips were underpriced. The fund took puts on $7.4bn of them in a single quarter.

In April 2024 OpenAI dismissed a researcher from its safety team over an alleged leak. Two months later Leopold Aschenbrenner published Situational Awareness: The Decade Ahead, an essay arguing that artificial intelligence was about to consume more electricity and capital than almost anyone had planned for. Then he did the thing essayists almost never do: he raised a fund, named it after the essay, and bought the argument.

Bought it literally. In the portfolio it reported for 31 March 2026, the fund’s six largest holdings were Bloom Energy, SanDisk, CoreWeave, IREN, Core Scientific and Applied Digital — a fuel-cell maker, a memory maker, and four companies whose business is putting computers in buildings and buying the power to run them. Those six were 84.8% of the shares it owned, by value.

The one sentence in the essay about buying anything

Fifty thousand words, and one sentence is investment advice: a note in brackets at the end of the chapter titled “Racing to the Trillion-Dollar Cluster.”

“What all of this means for NVDA/TSM/etc. I leave as an exercise for the reader. Hint: Those with situational awareness bought much lower than you, but it’s still not even close to fully priced in.”

Chips are cheap, and the people who understand this already own them. That is the essay’s entire view on what to own.

What the fund held

American investment managers above a size threshold file a form called a 13F once a quarter, listing what they held on its last day. It reports shares and options held long, never short sales or borrowing. For an option it reports the value of the underlying shares, not what the option cost, what it is worth or what anyone stands to lose.

Here is Nvidia, across every filing the fund has made:

No Nvidia long line appears in any of the seven except one: 2,855 shares, about half a million dollars, in the same filing as the $1.57bn put.

Nor was it only Nvidia. Counting AMD, ASML, Broadcom, Intel, Micron, Nvidia, TSMC and the VanEck semiconductor ETF, the underlying value of the fund’s semiconductor puts ran:

Read the two bold rows twice. At the end of December the fund held no semiconductor puts at all. Ninety days later it held puts referencing $7.36bn of chip stock, and its largest line was a $2.04bn put on a semiconductor ETF.

The ninety days

That is the finding, and it is bigger than the reversal, because two other things happened in the same window.

First: the long column stops moving, and that is not what it looks like. $3,914m to $3,856m is a market value, not a measure of buying; it can sit still while everything under it changes, and here it did. Eight positions in the December filing are absent from the March one. Ten appear that were not in December. Of the sixteen names in both, nine carry more shares and three fewer — Bloom Energy, the largest holding, down 35.6% of its shares, while CleanSpark rises 648%, Bitfarms 188%, Bitdeer 92% and Riot 87%. Two quarter-end photographs cannot say when that happened. They do say that thirty lines changed, and share counts are the one field on this form that counts rather than values.

Second: the protection inverted. At the end of December a $3.9bn book of memory, power and cloud equities had one put behind it — $8.9m of Infosys, an Indian IT services company in none of the businesses the rest of the book was in. Ninety days later the put column was 219% of the long column.

The ratio divides shares referenced by puts by the market value of shares owned, so it is not hedge coverage, net exposure or capital at risk. It shows one thing: at $9m the puts were too small to matter to a $3.9bn book, and at $8.46bn too large to have been sized against it.

What happened next

From 31 March, the quarter-end date of that filing, to 30 July:

Every one rose, as did the semiconductor ETF, the largest put line. Three names in the put book fell: Oracle, referencing $1.07bn, down 13%, and Corning and Infosys, the two smallest lines, down 1% and 7%. None of the three appears in the essay.

In July the fund lost 67% of its value, it told partners in a letter the WSJ reported, and it sold the bulk of its public stock portfolio to Citadel. It said the year was still up about 80%. Those are the manager’s own unaudited figures; the price moves above are not, and neither set explains the other.

Before July the press barely covered the fund. TEXXR holds 5 articles matching “Situational Awareness” in the two years from June 2024, and 23 from 1 July to 27 October 2026.

Where this reading is weakest

A put may be an outright bet or a hedge against something the form does not show, and the filing cannot tell you which. The fund may have been short in ways no 13F reports. A quarter-end photograph filed weeks later is a poor guide to anyone’s conviction; this one was filed on 18 May, seven weeks after the date it describes.

There is also a reading with no contradiction. The essay argues about what will be built, not about who keeps the margin: compute demand can be enormous while the companies selling it are badly priced. On that reading the put book asks the second question, which the essay never did.

What the record does establish is narrow: the essay said chip stocks were underpriced, and for one quarter the fund named after it held puts referencing $7.36bn of them.

One caution: the March filing’s gross value, $13.68bn, is not the size of the fund. $8.46bn of it is put notional, shares the options reference, not money anyone put up; only the $3.86bn of long equity is the market value of shares owned.

What the June filing showed

This piece set itself a test: if the June filing showed the puts gone and the long book rebuilt in the same names, its reading was the weaker one. The filing of 14 August shows exactly that.

Ten of the eleven put lines are gone; an Infosys put referencing $5.2m remains. Long equity rose from $3.86bn to $20.17bn, and 56% of it sat in SanDisk ($5.67bn) and Micron ($5.57bn). In March Micron was mostly a put, referencing $583.7m. TSMC went from a $535.1m put to $1.27bn of shares.

By its own test, this piece lost. The reversal was one quarter’s position, closed by 30 June. The filing cannot say when the puts came off, or why.

Held unchanged, the June book’s priced long lines fell 30.4% from 30 June to 30 July at our own closes; SanDisk fell 43.7%, Micron 24.2%. The fund told partners it lost 67% that month. Borrowing, which no 13F reports, would account for much of the gap; the fund sold positions to meet margin calls, Bloomberg’s source said. In the same weeks memory stocks moved as one.

Every filing behind this piece is laid out, priced forward against our own sessions, on the Situational Awareness fund page.

Sources

Position figures are summed from the position lines of seven 13F-HR filings by CIK 0002045724, with the raw SEC documents archived alongside; every period reconciles to the filing’s own stated total. Share-count comparisons are matched on CUSIP between the 2025-12-31 and 2026-03-31 filings.

The March 2026 filing is accession 0002045724-26-000008, period 2026-03-31, filed 2026-05-18. The June 2026 filing is accession 0000935836-26-000418, period 2026-06-30, filed 2026-08-14. Until 26 October our reader of these filings parsed it as empty, so this page went two months without it; the figures above come from the corrected reading, which reconciles to the filing’s own stated total of $20.24bn.

“Semiconductor” above means AMD, ASML, Broadcom, Intel, Micron, Nvidia, TSMC and the VanEck semiconductor ETF; CoreWeave, Oracle, Infosys and Corning are excluded. Prices are as-traded closes from our point-in-time archive, 2026-03-31 to 2026-07-30, unadjusted for splits. The archive has no 31 March close for Infosys or the ETF, so their moves run from the price implied by the March filing to 29 July, the nearest close it holds.

The essay is quoted from situational-awareness.ai, chapter IIIa, “Racing to the Trillion-Dollar Cluster,” retrieved 2026-08-01. Performance figures are unaudited, from the manager’s July 2026 letter to partners as the Wall Street Journal reported it on 31 July (TEXXR). The sale to Citadel: Wall Street Journal, 31 July (TEXXR). Margin calls: Bloomberg, 30 July (TEXXR).

Coverage count: TEXXR article search for “Situational Awareness”, run 2026-10-27: 5 articles dated 1 June 2024 to 30 June 2026, 23 dated 1 July to 27 October 2026. A text match, so a story logged twice counts twice.

Small long stubs sit beside eight of the eleven put lines in the March filing, between 0.03% and 5.6% of the put notional; they are a habit of the book rather than a position.

A 13F reports long positions only, is filed up to 45 days after the quarter it describes, reports the value of the underlying rather than a premium or strike for option lines, and never reports leverage — so gross 13F value is not assets under management and is not treated as such here. Not investment advice.