Leopold's Fund Fell From $45B To $10B. The 9 Names He Reported Are Down 35-63%.

Mathematical Money | November 17, 2026


Leopold Aschenbrenner's fund sent its latest quarterly report to the SEC on Monday. The reported total was $1.58 billion. The one before it was $20.24 billion. That's a fall of 92%.


The nine companies in it are all lower since the quarter ended on 30 September. They've fallen between 35% and 63%. The S&P 500 is down 3%.


If he had simply held everything in that report from 30 September until Monday's close, I estimate the listed positions would be worth about 70% less. I'll show how I got there further down.


First, the story. It's one of the most unusual in recent finance. A researcher is fired from OpenAI, writes an essay, starts a fund named after it, and runs it to around $45 billion. Then it nearly comes apart in four weeks. I'll go through all of it: who he is, what he argued, what the fund owned, what happened in July, what he's done since, and what the new filing really shows.


Who is he?


He grew up in Germany and went to school in Berlin. He finished Columbia in 2021 as valedictorian, at 19, with a degree in economics and statistics. He also co-founded the university's effective altruism group, and he did research with Oxford's Global Priorities Institute.


In February 2022 he joined the FTX Future Fund, the philanthropic arm of Sam Bankman-Fried's exchange. He resigned on 11 November 2022. That was the day FTX filed for bankruptcy.


In 2023 he joined OpenAI's Superalignment team, led by Ilya Sutskever and Jan Leike. It worked on keeping very capable AI systems under control. He co-wrote a research paper there called "Weak-to-Strong Generalization".


After a hack of OpenAI's internal messaging in April 2023, he wrote a memo to the board. It argued that the company's security was too weak against foreign spies.


OpenAI fired him in April 2024 for allegedly leaking information. He says it was a brainstorming document he shared with three outside researchers for feedback. OpenAI says the firing had nothing to do with the memo. He says he was told the memo was a major reason. I can't settle that, and neither can anyone outside the company.


About a month later the Superalignment team was dissolved.


What did he argue?


In June 2024 he published a 165-page essay called "Situational Awareness: The Decade Ahead". It's a long argument, and the fund trades on it.


The core claim is that AI systems as capable as a human researcher could plausibly arrive by 2027. After that, AI could start doing AI research itself and compress years of progress into months.


The practical part is about money and power. The essay says the largest computer clusters would go from $10 billion to $100 billion to a trillion dollars. According to summaries of it, it puts power needs at 1 gigawatt in 2026, 10 gigawatts by 2028 and 100 gigawatts by 2030.


It also argues the labs aren't secure enough against state spies. And it expects a US government effort, which it calls "The Project", to take over around 2027 or 2028.


You can disagree with all of it. But read it as an investor and it's a shopping list: chips, memory, power equipment, data centres. That's exactly what the fund bought.


How did the fund start, and how did it do?


The fund launched in July 2024 under the same name. CNBC reports about $225 million of starting money, from Patrick and John Collison of Stripe, Nat Friedman and Daniel Gross. Jane Street is also named as an investor.


The idea was a hedge fund version of the essay. Own the companies that benefit from AI buildout. Short the software companies AI might hurt. Keep private stakes in AI companies, with a very large one in Anthropic.


It worked, and then some. The Wall Street Journal reported gains of more than 1,000% since launch. The Financial Times reported a 439% net return for the first half of 2026, from an investor letter dated 24 July.


The fund also led private rounds. It led an $830 million round in the data-centre company Fluidstack in January at a $7.5 billion valuation, according to Bloomberg via Hedgeweek. Other reports date that round differently. It also reportedly co-led a $500 million round in the chip start-up MatX. Bloomberg values the Anthropic stake at about $5 billion.


The usual hedge fund fee applies, 2% a year plus 20% of gains, according to Damodaran.


CNBC puts the fund at about $45 billion at its peak. Other reports give lower numbers, and I can't find that figure in any filing.


What did it own?


The clearest window is the filing for 30 June. It listed $20.2 billion across 26 positions.


The five biggest were these.


SanDisk, $5.7 billion. Micron, $5.6 billion. Bloom Energy, $1.9 billion. Taiwan Semiconductor, $1.3 billion. Nebius, $1.2 billion.


That's memory chips, power for data centres, and AI cloud companies.


There was also a block of bitcoin miners that had turned into AI hosts: Core Scientific, Riot, IREN, Applied Digital, CleanSpark and Bitdeer. Plus CoreWeave at $745 million.


A filing like this shows listed longs and options. It doesn't show shorts or private stakes. So the public face of the fund was only part of the picture.


What happened in July?


On 24 July the fund sent its investor letter with the 439% figure. According to Inc, it ended with a postscript saying this seemed like a good time to add funds.


Within days the picture changed. The AI infrastructure names fell hard. Reports say the software shorts went the other way, with Adobe named as one. And leverage, reported at as much as 400%, magnified both.


CNBC didn't confirm that leverage figure, and the fund didn't respond to its request for comment. If it's right, the fund held about four dollars of stock for every dollar of its own money.


The last normal trading day was 29 July, according to Damodaran. On 30 July the fund sold most of its listed portfolio to Ken Griffin's Citadel, according to the Wall Street Journal and CNBC. CNBC says assets went from about $45 billion to about $10 billion.


Then came the letter. It put the unaudited loss for July at 67% after fees. For the year so far it showed a gain of about 80%.


Those two numbers fit together. A 439% gain is 5.4 times your money. After a 67% loss that's about 1.8 times. So he gave back about 82% of the year's gains in one month, and still ended the year ahead.


Press reports quote the letter as saying, "We let you down this month." They also quote it saying the fund "came closer to permanent capital impairment than is acceptable to us."


The letter also said the fund wasn't shut down. It wasn't liquidated either, and it wasn't turned into a private-only fund.


The estimates of the loss don't agree. Damodaran, an NYU professor who wrote about it, put the loss at more than 43% of capital over four weeks, with the listed holdings down almost 67%. The letter's figure is 67% for the month. They may be measuring different things.


Here's the part I checked myself. I took the nine companies in the latest filing and pulled their closing prices from 30 June to 29 July.


Eight of the nine fell more than 35%. The S&P 500 fell 2.3%.


So July wasn't a market crash. It was one theme falling all at once.


What has he done since?


He kept going. Reports say the fund put another $400 million into a private chip start-up in early August, on top of $100 million already held. It's reportedly called Source Foundry. The fund declined to comment.


According to the letter, the public side now runs without borrowed money. That fits what came next. Between 2 and 10 September, CNBC's David Faber reported that the fund bought call options on AMD, Bloom Energy, CoreWeave, SK Hynix, SanDisk and a memory-chip fund. A call option gives you big exposure for a fixed upfront cost.


Jim Cramer wrote, "HE's back, he's bigger than ever." By the next Monday all six names were lower, some by as much as 8%, according to Protos.


In late October, Bloomberg reported the fund was asking its existing clients to co-invest in a new Fluidstack round at a $40 billion valuation, about double the earlier one. A Fluidstack co-founder disputed the report. The fund had also reportedly considered selling its private stakes in Anthropic, MatX and Fluidstack during the crunch. It's not known whether it sold any.


CNBC now puts the fund at about $10 billion, mostly private stakes.


The timeline in one place


Here are the dates in order, so you can see how fast the second half moved.


- 2021. Graduates from Columbia as valedictorian, at 19.

- February 2022. Joins the FTX Future Fund. Resigns on 11 November, the day FTX files for bankruptcy.

- 2023. Joins OpenAI's Superalignment team.

- April 2024. OpenAI fires him over an alleged leak. He disputes the reason.

- June 2024. Publishes the 165-page essay.

- July 2024. The fund launches.

- 24 July 2026. The investor letter reports a 439% gain for the first half.

- 29 and 30 July 2026. The last normal trading day, then the sale of the listed book to Citadel.

- Early August 2026. Reportedly puts $400 million into Source Foundry.

- 2 to 10 September 2026. Buys call options on six AI names.

- Late October 2026. Reported to be asking clients to co-invest in Fluidstack.

- 16 November 2026. Files the new report.


What does the new filing really show?


I read the filing itself on SEC EDGAR, not just the news.


It lists 23 lines, worth $1,577,454,028 in total. Two thirds of that, $1.04 billion, is call options. A filing values an option at the shares it controls, not at what the buyer paid. It doesn't give strikes or expiry dates. So you can't read his real risk from it.


The plain shares are $533 million. SharonAI is most of that, at $441 million.


The rest is small. CleanSpark is $39 million. IREN is $22 million. Bitdeer is $20 million. Fermi is $11 million.


The calls are on four names: Nebius, Bloom Energy, SanDisk and the Roundhill Memory ETF.


One correction to how this is being read. Some coverage treats SharonAI as a new stake. It isn't. The June filing already had it at $457 million, with 5.4 million shares. He added about 50% more shares since.


Three things to check on any fund's filing before you trust a headline about it:


1. Options are shown at the shares they control, not at what was paid.

2. It's a snapshot of the last day of the quarter, published weeks later.

3. Shorts and private stakes aren't in it at all.


I checked every line against Tiger's closing prices on 30 September. Each value is the number of shares times that day's close, to the cent. SharonAI is $54.60 a share. Nebius is $235.88. So this is a quarter-end snapshot and nothing else.


What would it be worth now if he'd held?


This is an estimate, and I'll show the working. CNBC says it isn't known whether he sold anything in October.


The plain shares are easy. The five lines were worth $533 million on 30 September. At Monday's close the same shares are worth $205 million. That's down 61%.


By name, SharonAI went from $441 million to $163 million. CleanSpark went from $39 million to $19 million. IREN went from $22 million to $9 million. Bitdeer went from $20 million to $9 million. Fermi went from $11 million to $7 million.


The nine names since 30 September, from Tiger closes:


SharonAI is down 63%. IREN is down 60%. Bitdeer is down 57%. Nebius is down 55%. SanDisk is down 54%. CleanSpark is down 51%. Fermi is down 41%. Bloom Energy is down 36%. The memory ETF is down 35%.


The calls are harder. The filing counts them at $1.04 billion, the value of the shares they control. On Monday's prices those shares are worth $576 million, down 45%.


But a call falls much faster than the stock under it, and the filing doesn't say what he paid. So I priced them with a standard options model. I used each stock's own recent volatility, a 4% interest rate, and a strike at or 10% above the 30 September price. I tried expiries in January, March and June 2027.


On those assumptions the calls were worth roughly $215 million to $385 million on 30 September. On Monday they're worth roughly $7 million to $72 million. That's a fall of 80% to 97%.


Add the shares and the calls together. They were worth roughly $750 million to $920 million on 30 September. They'd be worth roughly $210 million to $280 million now. That's a fall of about 70% in every case I tried.


Two things to keep in mind.


First, the real money in the options was much smaller than $1.04 billion. On my assumptions it was about a fifth to a bit over a third of it. The headline number in the filing overstates what was at risk.


Second, this is a small part of the fund. Say the fund is about $10 billion. The loss on these positions is roughly $0.5 billion to $0.65 billion. That's about 5% to 6% of it.


The bigger question for him is the private side, and that depends on companies like Anthropic.


If he sold early, or hedged, or the calls had strikes far below the price, the picture is better than this. The model can't know.


What should you take from it?


I don't think the lesson is that he was careless. The thesis was early and large, and for two years it paid.


The lesson is about how fast a good idea can move against you when many positions are really one. You can check that on your own book.


1. List your top holdings, or any fund's.

2. Pick your worst four-week stretch of the last year. Pull each holding's closing price at the start and end of it.

3. Count how many fell together, and compare with the index. If most fell far more than the index, you hold one bet.

4. Take your leverage. Divide 100% by it. That's the fall that takes all your own money. At four times it's 25%, at two times it's 50%.


For his nine names, July gives you the answer: eight of nine fell more than 35% while the market barely moved.


What's next?


The big item is Anthropic. It's the stake that survived July and is now a large part of what the fund holds. Its listing is expected on 19 November, as I wrote on 12 November.


Then the next filing, for the quarter ending 31 December, is due in February. It will show whether he held the calls or the shares through the October fall.


Nothing here is advice to buy or sell anything. It's a case study from public filings and public prices.


If you had read his essay in 2024, how would you have used it? I'd like to hear whether you'd have bought the thesis, the companies, or neither.


Sources: Situational Awareness LP's quarterly reports to the SEC (13F-HR) for 30 June 2026, filed 14 August, and for 30 September 2026, filed 16 November, on SEC EDGAR, CIK 0002045724. CNBC, 31 July and 16 November 2026. Bloomberg, 3 August 2026, and Bloomberg via Hedgeweek, 29 October 2026. Aswath Damodaran, "The Situational Awareness Fund Blow-up". Inc, Hedge Fund Alpha and other press reports of the 24 July investor letter. TechCrunch, 30 July 2026. 24/7 Wall St on Yahoo Finance, 17 September 2026, which relays CNBC's 11 September report and Protos on 14 September. Biographical details from CNBC and Wikipedia. The essay is at situational-awareness.ai. Price moves and every per-share figure are Tiger closes. The July window is 30 June to 29 July 2026. The 35-63% falls run from 30 September to 16 November 2026. The estimate of the positions' value uses the Black-Scholes model. It takes each stock's own 60-day realised volatility and a 4% rate. Strikes are set at the 30 September price and at 10% above it. Expiries are January, March and June 2027. The filing gives no strikes, expiries or prices paid, so this is an illustration, not his actual result. The $45 billion peak and the "as much as 400%" leverage are reported by CNBC and others and are not confirmed by any filing I could find.


$NBIS$ $SNDK$

#AI #hedgefunds #13F #semiconductors #investing


Stop guessing. Start calculating.


Live to fight another day. 🤙

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Report

Comment(2)

  • Top
  • Latest
  • fishhhh
    ·11-17 16:13
    Big narrative plus huge size is how these things turn into stampedes fast. Live to fight another day fr
    Reply
    Report
  • bubblyo
    ·11-17 16:13
    92% drawdown implies way more than a plain 4x story. Leverage math still looks off
    Reply
    Report