Situational Awareness spent two years as the most talked-about new fund on Wall Street, and it took four weeks in July 2026 to turn it into the most talked-about near-disaster. The Securities and Exchange Commission has now sent subpoenas to several of the largest banks in the United States asking what they knew, when they knew it, and what they said to the fund while its positions were falling apart.
The New York Times reported the subpoenas first. Reuters put names to the banks, and CNBC carried the detail on 25 August 2026: Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America have all been asked for records relating to Situational Awareness — its trades, its borrowing, and its conversations with the lenders that financed it.
Nobody has been accused of anything. That point is worth making early and making plainly, because the story is unusually easy to tell badly. Regulators routinely look at funds that lose a lot of money very quickly, and a great many of those reviews end without a fine, a charge or even a public finding. If you follow how the money behind this boom is being raised and spent, our AI models, tools and releases hub covers the product side of the same story.
What follows is the full picture: what the regulator has asked for, how the collapse actually worked mechanically, what the fund owned, why Citadel was the buyer, why the headline numbers in different outlets do not match, and what any of it should change for a business that is buying artificial intelligence rather than trading it.
Table of contents
- What The SEC Wants From Situational Awareness And Its Banks
- How Situational Awareness Fell From $45bn To $10bn
- Inside The Book Situational Awareness Actually Owned
- The Leverage Question At The Centre Of The Situational Awareness Probe
- Why Citadel Bought The Situational Awareness Portfolio
- Situational Awareness, Archegos And The Prime Broker Problem
- Leopold Aschenbrenner And The Essay That Named The Fund
- Why The Situational Awareness Numbers Do Not All Agree
- What The Situational Awareness Probe Means For Everyone Else
- What Happens Next In The Situational Awareness Investigation
- Frequently Asked Questions About Situational Awareness
- References
What The SEC Wants From Situational Awareness And Its Banks
The subpoenas are the news. Everything else is context for them.
Which banks received them
Four names have been reported: Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America. All four acted as prime brokers to Situational Awareness, meaning they held its positions, financed them, and set the margin terms that governed how much the fund could borrow against them.
What the regulator has asked for
Three categories, according to the reporting. The sequence and timing of the fund’s trades. The leverage the banks extended and how they monitored it. And the communications between the banks and the fund’s management during the period when the positions were under stress.
The preservation notice matters as much as the subpoena
Alongside the requests, the banks were told to preserve any information relating to Situational Awareness. A preservation instruction is not itself an allegation, but it is the step a regulator takes when it expects to want a full record later rather than a summary now.
The banks are saying nothing
Goldman Sachs, JPMorgan and Citigroup declined to comment. Bank of America was contacted and had not commented at the time of publication. That is the normal posture for an open inquiry and should not be read either way.
What the fund said
Situational Awareness issued a short statement that neither denied nor conceded anything. “It is to be expected that regulators would closely examine any funds that are high profile, produce significant returns, or have particularly dramatic drawdowns,” it said. “We are a highly-regulated business and will cooperate to the fullest extent with any regulatory request.”
| Bank | Reported role | Public comment |
|---|---|---|
| Goldman Sachs | Prime broker, margin lender | Declined to comment |
| JPMorgan Chase | Prime broker, margin lender | Declined to comment |
| Citigroup | Named lender to the fund | Declined to comment |
| Bank of America | Prime broker, margin lender | Contacted, no comment given |
How Situational Awareness Fell From $45bn To $10bn
The drawdown is the reason the subpoenas exist, so the mechanics deserve care.
The starting position
By the first days of July 2026 the fund was, on the most widely quoted figures, running about $45 billion. It had reported a net return of 439% through 30 June, a number that made Situational Awareness the best-performing large fund in the world on a one-year view.
The sell-off that started it
Through July, AI infrastructure and semiconductor stocks corrected hard. The Nasdaq 100 fell more than 10%. Korea’s Kospi index, where the fund held a large position in SK Hynix, fell roughly 33%. Individual holdings dropped far further than either index.
Why leverage turned a bad month into an emergency
Reported gross leverage ran as high as four times, which CNBC rendered as leverage “of up to 400%”. At that ratio the arithmetic is unforgiving: a 30% fall in the long book translates to roughly a 120% hit to the equity underneath it. Situational Awareness did not need to be wrong about artificial intelligence to be destroyed by it — it only needed to be early.
The margin calls
As the value of the collateral fell, the prime brokers issued margin calls. The fund reportedly tried the conventional escapes first: raising fresh capital, and selling assets directly to its own investors. Neither closed the gap fast enough.
The single block trade
Before the market opened on 30 July 2026, Situational Awareness sold its entire public equity book — longs and shorts together — in one block. Reuters put the size of that book at roughly $16 billion. By the end of that week the fund’s assets stood at around $10 billion.
What survived
The private holdings did. Chief among them is a stake in Anthropic, bought during that company’s $1 billion Series H round in May 2026. Private positions cannot be margin-called the way listed equities can, which is precisely why they were still there when the listed book was gone.
| Date | Event |
|---|---|
| June 2024 | Aschenbrenner publishes the essay the fund is named after |
| Late 2024 | Situational Awareness launches, reportedly with about $5m |
| 31 March 2026 | Last filing before the collapse: $3.86bn across 26 names |
| May 2026 | Stake taken in Anthropic’s $1bn Series H round |
| 30 June 2026 | Net return reported at 439% |
| July 2026 | Nasdaq 100 falls over 10%, Kospi falls about 33% |
| 30 July 2026 | Entire public book, about $16bn, sold to Citadel pre-market |
| 21 August 2026 | Griffin says Citadel has shed about 80% of that risk |
| 24 August 2026 | The SEC inquiry into Situational Awareness is reported |
| 25 August 2026 | Four banks named as having received subpoenas |
Inside The Book Situational Awareness Actually Owned
A fund is its positions. The disclosed ones tell you a great deal about how this ended.
The 13F picture
The last regulatory filing before the collapse, covering positions as of 31 March 2026, showed a disclosed long book of about $3.86 billion across 26 names. That figure is far smaller than the fund’s headline assets, which tells you most of the exposure sat in swaps, options, and holdings listed outside the United States.
Concentration was the strategy, not an accident
The top five disclosed positions accounted for more than 76% of that book. Bloom Energy alone was 22.8%. Nothing about this was hidden or unusual for the manager — concentration was the stated approach, and it is what produced the 439% before it produced the drawdown.
The theme underneath the names
Read the list and the thesis is obvious: electricity, memory, and data centre capacity. Bloom Energy for power. SanDisk and Micron for storage and memory. CoreWeave, IREN, Core Scientific and Applied Digital for the compute estate. It was a bet on the physical plumbing of cloud computing rather than on the model developers themselves.
The short side
Situational Awareness was not only long. It ran shorts against software companies it expected artificial intelligence to disrupt, including Adobe, and held put options with roughly $1 billion notional on the SMH semiconductor ETF and about $0.6 billion notional on Nvidia.
The overseas position that did the damage
SK Hynix, the Korean memory maker, was a large holding and is not captured in a US 13F. When the Kospi fell about a third, that position moved against the fund at a scale the domestic filings would never have shown an outside observer.
The Leverage Question At The Centre Of The Situational Awareness Probe
If the inquiry has a single organising theme, this is it.
Borrowing is not the allegation
Leveraged funds are legal, common and closely supervised. Every hedge fund of scale borrows against its positions. The question a regulator asks after an event like this is never “was there leverage” but “was the leverage understood, disclosed and controlled by the people who extended it”.
Four times is a lot for one theme
Four times gross exposure would be unremarkable in a diversified book. Applied to a portfolio where five names are more than three quarters of the disclosed longs, and where every one of those names responds to the same news, it stops being a financing choice and becomes a directional bet on a single idea.
Each bank saw only its own slice
This is the structural problem, and it is why the subpoenas went to lenders rather than to the fund alone. Each prime broker could see the exposure it had financed. None of them necessarily saw the total. The aggregate risk that Situational Awareness was carrying may have been visible to nobody, including in some readings the fund’s own investors.
The conflicting-roles angle
Reporting has also pointed at the multiple hats worn by the same institutions: lender, margin-caller, and facilitator of the eventual sale. Whether those roles were managed cleanly, and whether the fund’s investors were told what they needed to know, is a reasonable thing for a regulator to test.
Why the timing questions are sharp
The SEC has asked about the sequence of trades and the timing of the margin calls. In a forced liquidation, who moved first determines who absorbed the loss. That is exactly the kind of question that separates an unlucky trade from a supervisory failure.
| What the SEC asked for | What it is testing |
|---|---|
| Trade records and their sequence | Who sold first, and whether anyone traded ahead of the unwind |
| Leverage extended and monitored | Whether margin terms matched the concentration being financed |
| Communications with the fund | What the banks knew about total exposure, and when |
| Preservation of related records | That nothing is lost before the review is finished |
Why Citadel Bought The Situational Awareness Portfolio
The rescue is the most instructive part of the whole episode.
One buyer, one price, one morning
Ken Griffin’s Citadel took the bulk of the book in a single pre-market transaction. Buying everything at once removed the risk of a days-long public liquidation in which each sale pushed the next one lower.
The discount
The purchase was understood to have been done at a discount of around 10%. For a forced seller with no time and no alternative bidder of comparable size, that is the price of certainty.
Citadel has already reduced the risk
In an investor letter dated Friday 21 August 2026, Griffin said Citadel had since offloaded roughly 80% of the risk associated with the acquired portfolio. Citadel’s Wellington fund gained 5.9% in July.
The market reaction proved the positions were sound
The day the block traded, the transferred names surged: Nebius rose 27.1%, IREN 26.5%, Bloom Energy 25.6%, SanDisk 24.6%, CoreWeave 22.0%, Core Scientific 20.7%. Against that, the broad market barely moved. The equal-weight S&P 500 was flat and SPY rose 1.8%.
What that gap tells you
Those moves were not news about the companies. They were the price recovering once the seller was gone. Situational Awareness had not been wrong about the assets; it had run out of the ability to hold them. SK Hynix and CoreWeave have both rallied since.
Situational Awareness, Archegos And The Prime Broker Problem
Every comparison being drawn in the market points to the same 2021 precedent.
What Archegos was
Archegos Capital Management collapsed in March 2021 after concentrated, heavily financed positions moved against it. The banks on the other side of those trades lost more than $10 billion between them, with Credit Suisse and Nomura taking the heaviest damage.
The family resemblance
Concentrated bets. Borrowed money. Several banks, each with a partial view. An unwind compressed into days rather than months. Situational Awareness is not Archegos, but the shape of the failure is recognisable enough that the regulatory reflex was immediate.
The important difference
Archegos left banks with enormous losses. In this case the banks were made whole, the positions were absorbed by a single buyer at a modest discount, and the assets recovered almost at once. The system worked, which is a genuinely different outcome.
What ESMA concluded last time
Europe’s markets regulator found that total return swaps arranged through synthetic prime brokerage were what allowed Archegos to build leverage while keeping its positions invisible to regulators. The same instrument family explains the gap between the fund’s small disclosed book and its very large real exposure.
Why regulators care about the pattern
The lesson from 2021 was that disclosure regimes built around ownership do not capture exposure built through derivatives. That gap has not closed, and Situational Awareness has now demonstrated it a second time on a much larger notional scale.
| Factor | Archegos, 2021 | Situational Awareness, 2026 |
|---|---|---|
| Trigger | Single-name price falls | Sector-wide July correction |
| Structure | Total return swaps, family office | Swaps, options and overseas listings |
| Unwind | Disorderly, over several days | One block, before the open |
| Bank losses | More than $10bn | Reported as none |
| Fund outcome | Wound up | Survived, roughly $10bn remaining |
| Asset recovery | Slow | Same day, in double digits |
Leopold Aschenbrenner And The Essay That Named The Fund
The manager is inseparable from the story, and the fund’s name is not a coincidence.
From Columbia to OpenAI
Leopold Aschenbrenner graduated from Columbia University in 2021 as valedictorian, with a degree in economics and mathematics-statistics, having been born in 2001. He worked as a researcher on the FTX Future Fund, then joined OpenAI in 2023 on the Superalignment team led by Ilya Sutskever and Jan Leike, co-authoring the paper “Weak-to-Strong Generalization”.
The dismissal
He was fired in April 2024 over what OpenAI characterised as an information leak. Aschenbrenner disputed that account, saying the document in question was a benign brainstorming memo and that the real friction concerned a security warning he had raised internally.
The essay
In June 2024 he published “Situational Awareness: The Decade Ahead”, a long essay running to well over 150 pages that argued AI systems would be capable of conducting their own AI research by 2027. It was read far outside the technical community and it gave the fund both its name and its investment thesis.
The backers
The money came from names that carry weight in technology rather than in finance: Patrick and John Collison of Stripe, Nat Friedman, Daniel Gross, and the trading firm Jane Street. The fund launched in 2024, reportedly with about $5 million of initial capital.
The incentive problem
The valuation expert Aswath Damodaran, writing on his own newsletter, argued that the standard 2% management and 20% performance structure “encourages reckless risk taking”, particularly for a young manager with everything to prove and a short track record to prove it in.
The scale of the reversal
On Damodaran’s reading the fund’s public equity value fell by nearly two thirds in four weeks, with a loss of principal above 43%, after peak returns approaching 450%. The number that made the reputation and the number that broke it were produced by the same positions.
Why The Situational Awareness Numbers Do Not All Agree
Anyone reading widely on this will hit contradictions. They are worth explaining rather than papering over.
Three different peak figures
Reuters and CNBC put the peak at about $45 billion. Some analysts put peak assets nearer $24 billion. Other outlets have used $30 billion. All three appear in credible reporting published within a month of each other.
They are measuring different things
Assets under management, gross exposure and the value of the public book are three separate quantities, and at four times leverage they diverge enormously. A $10 billion equity base supporting $40 billion of positions can honestly be described with either number.
The disclosed book was never the whole book
The $3.86 billion in the March filing is not a fourth contradictory figure. It is what US disclosure rules require a manager to publish, and it excludes swaps, options, and shares listed in Seoul. Treating it as the fund’s size would be a serious misreading.
How to read the coverage
Prefer figures that name their basis. “Assets fell from about $45 billion to around $10 billion” is a claim about the fund. “The public book sold to Citadel was about $16 billion” is a claim about one transaction. Both can be true at once.
What is not in dispute
The direction and the speed. Whatever the starting number, Situational Awareness lost most of it inside a single month, was forced to liquidate, and is now the subject of a regulatory inquiry. No source disagrees on any of that.
What The Situational Awareness Probe Means For Everyone Else
Most readers are not hedge fund investors. The episode still carries three practical lessons.
Leverage, not artificial intelligence, was the failure point
This was not a verdict on whether the technology works. The fund’s picks recovered within hours of the seller leaving the market. What failed was the financing structure wrapped around a correct-looking view, and that distinction matters if you are budgeting for AI rather than trading it.
Your suppliers may be inside somebody’s leveraged bet
CoreWeave, Nebius, IREN and Core Scientific are not abstractions — they are compute providers that businesses actually buy from. When a single large holder is forced to sell, their share prices move violently for reasons unrelated to service quality. Check the balance sheet behind any critical supplier with the same seriousness you would apply to its cybersecurity posture, and build an exit into the contract. That is ordinary AI strategy discipline rather than market timing.
Concentration risk is not only a portfolio idea
The fund’s mistake — five names, one thesis, one set of catalysts — has a direct analogue in procurement. If your model provider, your compute provider and your tooling all depend on the same capital cycle, you do not have three suppliers. You have one, three times over.
Price volatility is not the same as service failure
None of the companies involved stopped operating. Their equity moved because of a forced seller, not because their capacity vanished. Do not renegotiate a working contract on the strength of a share price chart.
The useful question to ask a vendor
Not “are you profitable” but “what happens to your delivery commitments if your funding round is delayed by two quarters”. That single question surfaces more real dependency risk than any amount of market commentary, and it belongs in your digital transformation planning.
What Happens Next In The Situational Awareness Investigation
The honest answer is that nobody outside the SEC knows, but the range of outcomes is well understood.
Early stage means early stage
Subpoenas to third parties are an information-gathering step, not a charging decision. Many inquiries that reach this point are closed quietly with no public statement of any kind.
The likely timeline is long
Reviews of this type are measured in quarters and often in years. Records have to be produced, reviewed and followed up, and the banks involved will have their own counsel managing the pace.
What would signal escalation
A shift from subpoenas aimed at the banks to formal requests aimed at the fund and its principals, or a Wells notice, would be the first genuinely meaningful change. Neither has been reported.
The market’s own estimate
Prediction market pricing on Polymarket has put the odds of Aschenbrenner facing criminal charges before the end of 2026 at roughly 6%. That is a crowd guess rather than evidence, but it is a reasonable summary of how the market reads the risk.
The wider policy question
Whichever way this one inquiry lands, the systemic issue is unresolved: leverage is quietly underwriting a large part of the AI buildout, and no regulator currently sees the whole picture. Situational Awareness made that visible in a single month. The next fund to demonstrate it may not have a Citadel waiting on the other side.
Frequently Asked Questions About Situational Awareness
What is Situational Awareness?
It is an AI-focused hedge fund founded in 2024 by Leopold Aschenbrenner, a former OpenAI researcher, and named after his June 2024 essay on the trajectory of artificial intelligence.
Has Situational Awareness been accused of wrongdoing?
No. The SEC has subpoenaed banks for information. Neither the fund nor any bank has been accused of breaking any rule, and such inquiries frequently close without enforcement action.
Which banks were subpoenaed?
Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America, according to Reuters, citing a person familiar with the matter. The New York Times first reported that subpoenas had been issued.
How much did the fund lose?
Its assets fell from roughly $45 billion in early July 2026 to around $10 billion by the end of the month, after a forced sale of its public equity positions.
Did Situational Awareness shut down?
No. It sold its listed portfolio but retained its private holdings, including a stake in Anthropic acquired in that company’s May 2026 funding round, and continues to operate.
Who bought the portfolio?
Citadel, the multi-strategy firm founded by Ken Griffin, bought the bulk of the book at a reported discount of around 10% in a single pre-market block on 30 July 2026.
Was this an AI bubble bursting?
The evidence points the other way. The holdings rallied by more than 20% the same day the forced seller disappeared, which suggests a financing failure rather than a collapse in the underlying demand.
What is the connection to Archegos?
Both involved concentrated, derivative-heavy, heavily borrowed positions spread across several prime brokers, none of whom saw the total. The outcomes differed sharply: Archegos cost banks over $10 billion, while these positions were absorbed without reported bank losses.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.