A 24-year-old raised a fortune on a 165-page AI manifesto, returned more than 1,000% since inception, and then lost roughly 67% in July 2026. Yet the fund remained up about 80% for the year, survived the drawdown, and retained the private investment that most directly expressed its founder’s belief in frontier AI.
That combination is more instructive than the various tens-of-billions loss figures attached to the episode. A sound technology thesis can coexist with an unsound portfolio. The physical AI buildout may continue for years, while a manager financing that view with borrowed money can still be forced to sell during a one-month reversal.
Matthew Kanterman and I approached the episode from different sides. Matthew built thematic investment products and launched CHAT, which he describes as the first generative-AI ETF, in 2023; by his account on the show, it had returned about 56% since inception at the time of recording. I came from the technology side, where Aschenbrenner claimed his advantage. Matthew focused on concentration, financing, and what the market had already priced. I focused on whether deep understanding of the technology can still give an investor an edge.
Our shared conclusion is narrower than “Leopold was right.” His knowledge helped him identify the buildout early. His portfolio structure nearly prevented him from owning it long enough to benefit.
The résumé sold access to an unusual point of view
Leopold Aschenbrenner was born in Germany and raised in Berlin by two doctors. He finished high school at 15, enrolled at Columbia at the same age, and graduated as the college’s 2021 valedictorian in economics and mathematics-statistics. Columbia faculty described his academic record as unmatched in the economics department in 20 years.
He then worked at OpenAI from 2023 until April 2024 on the Superalignment team, which studied how humans might supervise systems more capable than themselves. He contributed to the team’s launch post and co-authored “Weak-to-Strong Generalization.” Ilya Sutskever and Jan Leike led the group; Aschenbrenner was a researcher, not its leader.
OpenAI fired him over an alleged leak. Aschenbrenner says he shared a preparedness document with three outside researchers for feedback and that a security memo he wrote was a major reason for his dismissal. OpenAI’s account differs, and the company has never publicly identified the confidential information at issue. The available reporting does not reconcile the two versions.
The dispute matters here because it shaped the pitch that followed. Aschenbrenner emerged as a former Frontier-Lab researcher willing to publish what he believed the rest of the market had not understood. His credibility came from proximity to the compute curve, not from an investing record.
One personal connection later became financially relevant: in August 2026 he married Avital Balwit, chief of staff to Anthropic CEO Dario Amodei. The fund’s Anthropic stake became its most important surviving asset.
A technology forecast became a portfolio
In June 2024, Aschenbrenner published Situational Awareness: The Decade Ahead. Its boldest claim was that “AGI by 2027 is strikingly plausible.” He compared the advance from GPT-2 to GPT-4 with a leap from a preschooler to a smart high-school student, then argued that continued gains in compute, algorithms, and “unhobbling” could produce another jump of similar scale.
The investable portion of the argument appeared in “Racing to the Trillion-Dollar Cluster.” On Dwarkesh Patel’s podcast, Aschenbrenner described the mechanism plainly:
“Unlike most things that have recently come out of Silicon Valley, AI is an industrial process. The next model doesn’t just require some code. It’s building a giant new cluster. It’s building giant new power plants. Pretty soon it’s going to involve building giant new fabs.”
In that model, software demand becomes a physical capital cycle. More capable systems require memory, power, land, cooling, networking, data centers, and semiconductor capacity. If the forecast is right, the companies controlling those bottlenecks collect a large share of the spending.
That view attracted Patrick and John Collison, Nat Friedman, and Daniel Gross as anchor investors. Sources disagree on the launch capital: CNBC reported a commitment near $225 million from the four, while The New York Times reported roughly $100 million raised by November 2024. Either figure was remarkable for a manager in his early twenties with little conventional investing experience.
The performance then supplied the missing credential. Reporting put the fund above 1,000% net since inception and up 439% through June 2026. Assets passed $20 billion. The broader investing operation reportedly reached roughly $45 billion at the start of July, although that was a partly levered figure and should not be confused with regulatory assets under management.
Twenty-six securities expressed one economic view
The last public snapshot before the crisis was the June 30, 2026 Form 13F, filed on August 14. It showed approximately $20.24 billion across 26 lines. The holdings occupied different parts of the AI supply chain:
Memory and storage: SanDisk and Micron.
On-site power: Bloom Energy.
Data-center landlords: Core Scientific and Applied Digital.
GPU-cloud operators: CoreWeave and Nebius.
Miners converting sites to high-performance computing: CleanSpark, Bitdeer, HIVE, Riot, and T1 Energy.
The list offered diversification by ticker, but little diversification by economic driver. Each position depended, in a different way, on sustained AI infrastructure spending and continued market enthusiasm for its beneficiaries. Twenty-six securities expressed one view.
The short book added another layer of risk. According to The Wall Street Journal’s reconstruction, the fund was long chips and infrastructure and short software companies expected to lose from AI. Traders call that a Texas hedge: both sides benefit from the same underlying outcome. If infrastructure stocks fall while software rallies, the longs and shorts lose together.
Public filings obscured that geometry. A 13F reports US-listed long positions and listed options, but it omits private holdings, foreign securities, most shorts, swaps, cash, and financing. Listed options appear at notional value rather than delta-adjusted exposure. A $2 billion put line therefore does not establish a $2 billion economic short, and the filing does not show whether the put is an outright bet, protection for a long position, or one leg of a spread.
Copying the filing produced a weaker version of the fund. A follower could buy the visible, crowded public equities but could not reproduce the private Anthropic position, the financing arrangements, or the investor lockups. The copycat inherited the theme without the full structure.
Four forces turned a drawdown into a forced sale
The July loss came from the interaction of concentration, leverage, correlation, and liquidity.
Concentration made the portfolio sensitive to a single factor. The companies differed, but their valuations all relied on confidence in rapid AI capital spending.
Leverage magnified the reversal. The Wall Street Journal reported that the fund borrowed roughly $3 for every $1 of its own capital. That implies about $4 of assets for each $1 of equity, or roughly 4x gross assets-to-equity. It is a reconstructed financing ratio from unnamed sources, not a disclosed measure of net exposure.
Correlation changed when the trade reversed. Infrastructure longs fell together while software shorts rose. Positions that appeared distinct in normal markets behaved like one large exposure during the selloff.
Liquidity determined the outcome. A position can have a quoted market price and still be impossible to sell at that price in the required size. When a fund owns a large fraction of a stock’s ordinary trading volume, liquidation pushes the price against the seller.
The feedback loop is mechanical. Falling prices reduce net asset value. Lower equity raises the leverage ratio. Lenders demand more collateral. Sales to meet those demands push prices lower, which produces another call.
That mechanism separates investment horizon from financing horizon. A manager may value an asset on expected cash flows through 2030, while a prime broker requires collateral tomorrow morning. The lender’s deadline wins.
July’s decline was severe across the book. SK Hynix fell roughly 50% from peak to trough, and reporting put Situational Awareness down about 67% for the month. The fund was still positive for 2026 because its earlier gains had been so large. That fact says little about each investor vintage's experience, for which public data are unavailable.
Citadel bought time at a discount
As pressure increased, Situational Awareness negotiated two potential transactions. It offered a portion of its Anthropic stake to existing investors at roughly a 20% discount, then withdrew the sale after reaching a deal on the public portfolio.
Citadel bought much of the leveraged public book at approximately a 10% discount to market. The transaction eliminated Situational’s leverage and gave the fund liquidity. Citadel later reported unwinding more than 80% of the inherited risk through more than 100 block trades with over $4 billion in market value. Its Wellington fund returned 5.94% in July.
Citadel’s advantage was its ability to hold and distribute risk. A forced seller effectively gives the buyer control over timing. The discount compensates the buyer for supplying liquidity when the seller has no comparable alternative.
The subsequent rebound supports a limited conclusion. It suggests that at least part of the transaction discount reflected forced selling rather than an immediate collapse in the companies’ prospects. It does not validate “AGI by 2027,” prove a decade of infrastructure demand, or establish that the same stocks were attractive after they recovered.
Matthew compared the transaction with Citadel’s purchase of Sowood Capital’s distressed portfolio in 2007. Both involved a respected manager, a levered book, rapidly deteriorating liquidity, and a buyer able to evaluate and absorb billions of dollars of risk on short notice.
The episode differs from Archegos in a central respect. Bill Hwang was convicted of fraud and market manipulation. As of the reporting used for this article, no regulator had publicly accused Aschenbrenner or Situational Awareness of fraud. This was a portfolio and financing failure, not a criminal case.
The private asset supplied the surviving equity
The fund’s private Anthropic stake never appeared on a 13F. Reporting says Situational invested in Anthropic’s February 2025 round at a $61.5 billion valuation. By June 2026, the position represented about 20% of the portfolio; after the crisis, Bloomberg and Reuters valued it at roughly $5 billion.
The public and private books carried different risks. Public equities were priced every day, pledged against financing, visible to copycats, and vulnerable to margin calls. Anthropic was illiquid, negotiated, and difficult to sell quickly. Those limits made it poor collateral during the crisis, but they also prevented a broker from liquidating it through the public market.
Illiquidity did not make Anthropic safe. Its reported value depended on a private mark, transfer restrictions, and a future transaction. It did, however, preserve a large equity stake after the public book was sold. The position the fund could not readily trade became the position it retained.
The filings cannot resolve the central exposure question
The reported 4x figure is useful only within its limits. Gross exposure adds the absolute value of longs and shorts; net exposure subtracts shorts from longs.
Consider a fund with $10 billion of equity:
A $40 billion long book with no shorts has 400% gross exposure and 400% net exposure. A 20% decline costs $8 billion before financing expenses, erasing 80% of equity.
A $40 billion long book paired with $30 billion short has 700% gross exposure but only 100% net exposure. That structure can still lose on both sides if the longs fall and the shorts rise.
Situational’s filings do not reveal its full short book, swaps, foreign holdings, cash, financing terms, or exposure at the moment of liquidation. We know the fund borrowed heavily. Public data cannot reconstruct its exact gross, net, or factor exposure.
Other uncertainties remain: the current asset mix, the precise Anthropic mark, investor lockups and redemptions, the returns by investor vintage, and whether the fund will resume substantial borrowing. Those gaps limit any confident judgment about its present risk.
The investment debate now turns on demand, price, and duration
The buildout thesis survived the July liquidation as a hypothesis. Its future value depends on whether AI revenue grows fast enough to support the committed capital.
Matthew framed the bearish case around the scale already reached. On the episode, he estimated that the frontier AI complex had grown to a $220 billion to $230 billion annual revenue run rate. If its principal addressable market is roughly $500 billion of enterprise software, the industry may have captured close to half of that opportunity unusually quickly. Capital cycles often overshoot demand, and the infrastructure suppliers would be exposed if revenue growth slowed while fixed investment continued.
My counterargument is that AI agents may address labor budgets far larger than enterprise-software spending. If AI systems perform work rather than merely replace software seats, the relevant market expands and the infrastructure cycle can continue much longer.
That disagreement produces a measurable test. Watch frontier-lab revenue growth against hyperscaler capital-expenditure guidance. Two consecutive quarters of materially slower revenue growth while capex commitments remain elevated would weaken the buildout case. Continued revenue growth against a broader labor market would support it.
Price remains separate from thesis. A rebound after a forced liquidation may confirm that the sale cleared below short-term market value. It also means a new buyer faces a different entry price. Company quality, industry growth, and stock attractiveness are three different judgments.
Matthew also raised a longer-term policy possibility: OpenAI, Anthropic, and NVIDIA may be becoming systemic institutions for the knowledge economy. If a large share of software production, research, and services depends on a few model providers and one dominant chip supplier, their failure would create consequences beyond their shareholders.
Systemic importance can create implicit support, but it also invites regulation, capital requirements, competition policy, and political control. The same concentration that may protect the largest firms can cap their economics. Investors should treat policy as part of the valuation, not a distant tail risk.
What the episode teaches investors
1. Leverage converts a debatable thesis into a fixed deadline. Concentration and correlation created the loss; borrowed money allowed lenders to determine when the positions had to be sold. An unlevered investor can wait through a drawdown. A levered investor may lose that choice.
2. A forced sale does not settle the underlying thesis. The rebound after Citadel’s purchase suggests liquidity pressure contributed to July prices. It validates neither Aschenbrenner’s AGI timetable nor the current valuations of AI infrastructure stocks. Investors still have to compare future demand with the price they already paid.
3. Situational awareness is useful only when the portfolio can preserve it. Technical understanding helped Aschenbrenner identify the physical AI buildout before much of the market. It did not protect him from crowding, correlation, or financing risk. For an individual investor, the practical advantage is not speed or access to leverage. It is the ability to understand a technology, size the position conservatively, and hold without a lender choosing the exit date.
The next evidence to watch is concrete: frontier-lab revenue growth, hyperscaler capital spending, the fund’s future use of borrowing, and the price paid for the infrastructure companies after their rebound. Those variables will show whether July marked a temporary financing failure or an early warning from an overbuilt capital cycle.
This article is for informational and educational purposes only and is not investment advice. Figures are as reported at the time of recording and should be verified against current data before acting.
