Leopold Aschenbrenner vs Buffett: Conviction vs Margin of Safety
- The two strategies differ less in what they believe than in what they assume about being wrong. Aschenbrenner’s fund sized its AI-infrastructure positions for the thesis being right and used leverage to amplify it. Buffett’s method, inherited from Benjamin Graham, sizes positions so that being wrong is survivable — and accepts lower returns as the price.
- Buffett has said outright that this costs him money. In the 2017 Berkshire letter: “Our aversion to leverage has dampened our returns over the years. But Charlie and I sleep well.” The margin of safety is not a claim to better forecasting. It is a decision to not need one.
- Situational Awareness reportedly returned 439% in the first half of 2026, then was forced to sell the bulk of its public equity portfolio to Citadel after AI infrastructure stocks fell and margin calls followed. Reported assets went from a peak near $45 billion to roughly $10 billion. These accounts come from unnamed sources; Aschenbrenner has not commented publicly.
- This is not a verdict on his thesis. A one-month drawdown in AI infrastructure equities says nothing about whether AGI arrives around 2027, and the fund retains private positions including an Anthropic stake reported at about $5 billion. What it tested was not the forecast but the structure built on top of it.
In the first half of 2026, Leopold Aschenbrenner’s fund reportedly returned 439%. By 30 July it was reported to have sold the bulk of its public equity portfolio to Citadel, after losses on AI infrastructure stocks triggered margin calls. Assets that had reportedly peaked near $45 billion were put at roughly $10 billion afterwards.
Across the same period, Berkshire Hathaway ended its first quarter holding $397.4 billion in cash and short-term Treasury bills, the largest liquid position in its history, roughly $339 billion of it in T-bills yielding around 3.7%.
It would be easy, and wrong, to read that as one investor being clever and the other being foolish. Both approaches are internally coherent. They differ on a single question, and it is not “will AI be big?” It is: what happens to you if you are wrong?
What did Situational Awareness actually bet on?
Aschenbrenner, a former OpenAI researcher, published the essay Situational Awareness in June 2024, arguing that continued scaling of compute and algorithmic progress made artificial general intelligence by around 2027 strikingly plausible. He launched a fund of the same name in September 2024 with roughly $225 million of seed capital from Patrick and John Collison, Nat Friedman and Daniel Gross, co-managed with Carl Shulman. Jane Street later invested — unusual for a firm that rarely backs outside managers.
The portfolio expressed a more specific claim than “AI will be big.” His argument was that the binding constraint on the build-out is not model quality but electricity, storage and data-centre capacity. Accordingly the fund held no meaningful position in the obvious names — no Nvidia, Microsoft, Google, Amazon or Meta. It owned the inputs instead: reported holdings have included Riot Platforms, Core Scientific, CleanSpark, Bloom Energy, AMD and Oracle, alongside memory makers SK Hynix and SanDisk and the neocloud provider Nebius. Private positions include a stake in Anthropic reported at around $5 billion, plus the chip designer MatX and the data-centre firm Fluidstack.
That is a real thesis, expressed precisely. It is also concentrated in a single macro bet, in the most cyclical part of it, and it was levered.
What is a margin of safety?
It is Benjamin Graham’s idea, set out in chapter 20 of The Intelligent Investor, and the reason it belongs in this comparison is that it is not a forecasting technique. It is the opposite of one.
The premise is that you cannot reliably predict what a business or a market will do. So rather than trying harder to be right, you buy at a price low enough that a substantial error in your own analysis still does not ruin you. The safety is structural: it lives in the gap between what you paid and what the thing is plausibly worth, and it works whether or not your reasoning was any good.
Buffett did not invent it and has never claimed to. He has called Graham’s book the best ever written on investing and said that chapters 8 and 20 have been the bedrock of his investing activities for more than sixty years. Its relevance here is that it is a discipline designed around the assumption of being wrong — which is exactly the assumption a levered, concentrated thesis position cannot afford to make.
Why does Buffett refuse leverage if it would raise returns?
Because he has decided the trade is bad, and — unusually for an investor describing his own method — he says openly that it costs him.
From the 2017 Berkshire shareholder letter:
Our aversion to leverage has dampened our returns over the years. But Charlie and I sleep well. Both of us believe it is insane to risk what you have and need in order to obtain what you don’t need.
That is not a claim that leverage does not work. It is an admission that it does, and a refusal anyway. The same letter sets out the reasoning with a table of the four occasions on which Berkshire’s own shares fell sharply — 59.1%, 37.1%, 48.9% and 50.7%. Buffett’s comment:
This table offers the strongest argument I can muster against ever using borrowed money to own stocks. There is simply no telling how far stocks can fall in a short period. Even if your borrowings are small and your positions aren’t immediately threatened by the plunging market, your mind may well become rattled by scary headlines and breathless commentary. And an unsettled mind will not make good decisions.
And the payoff, in the same letter:
When major declines occur, however, they offer extraordinary opportunities to those who are not handicapped by debt.
The $397.4 billion is that sentence in balance-sheet form. Cash yielding 3.7% is a poor asset in a rising market and Berkshire knows it. What it buys is the guarantee of never being a seller at someone else’s chosen moment.
What happened at the end of July 2026?
Reported, and worth reading with the sourcing in mind: these accounts come from unnamed sources cited by the Wall Street Journal, the Financial Times, CNBC and Bloomberg. Aschenbrenner has not commented publicly.
AI infrastructure equities fell sharply through July, as public investors grew concerned that enormous capital spending was not converting into near-term revenue. The names hit hardest were the ones the fund was concentrated in: SK Hynix, SanDisk, Bloom Energy and Nebius each fell more than 30% in a month. Leverage amplified the losses into margin calls, and the fund was reported to have sold the bulk of its public holdings — Citadel taking the other side.
The concern driving that sell-off is not abstract, and it is measurable in the filings of the companies doing the spending. In the June 2026 quarter Meta’s capital spending consumed 97.5% of its operating cash flow and free cash flow fell to $784 million; Microsoft, spending more in absolute terms, consumed 66.2% and kept $18.7 billion. We looked at that in detail in Microsoft vs Meta free cash flow . The market’s worry about capex not yet earning its keep is the same phenomenon that took Situational Awareness’s positions down — seen from the other end.
Is this a verdict on the AGI thesis?
No, and treating it as one would be the least interesting reading available.
A one-month drawdown in AI infrastructure equities is not evidence about whether AGI arrives around 2027. The fund’s private positions, including the Anthropic stake, were not part of the liquidation. Over its life it remains, on the reported figures, extraordinarily profitable — $225 million of seed capital in September 2024 does not reach the tens of billions by being wrong about direction.
What the month tested was not the forecast. It was the structure built on top of it: whether a portfolio could hold its thesis through a drawdown it did not predict. A margin call is not a referendum on your analysis. It is a statement that your position was larger than your ability to wait.
The two structures, side by side
| Situational Awareness | Berkshire Hathaway | |
|---|---|---|
| Core assumption | The thesis is right, and timing is tractable | The thesis may be wrong, and timing is not |
| Position sizing | Concentrated on one macro bet | Diversified, with permanent excess cash |
| Leverage | Used, and central to the outcome | Refused, explicitly and at a stated cost |
| What a 30% drop does | Forces selling | Creates buying |
| What it optimises for | Return if right | Survival if wrong |
| Reported result, H1 2026 | +439%, then a forced liquidation | $397.4bn in cash, earning ~3.7% |
Neither column is the correct one in the abstract. The first produced a return the second cannot reach; the second produces an outcome the first could not guarantee. What is not available is both.
What to watch
Whether the reported account is confirmed. Everything about the liquidation currently rests on unnamed sources with no comment from the fund. Position-level disclosure would come from a subsequent 13F, which reports with a lag and only covers certain US-listed holdings.
Whether the fund raises new capital. Reporting indicates it is seeking it. That, more than the drawdown itself, will show what investors concluded — about the thesis and about the structure separately.
Berkshire’s second quarter. The $397.4 billion figure is from 31 March 2026 and the next set of numbers is due within days. Whether the cash pile grew again under Greg Abel, who became chief executive on 1 January 2026 with Buffett remaining as chairman, is the readable signal about whether the stance is Buffett’s personally or Berkshire’s institutionally.
Whether the capex worry persists. The sell-off was driven by doubt that AI capital spending is converting into revenue. That question gets answered quarter by quarter in the cash flow statements of the companies doing the building, not in commentary.
Sources
Each was checked on 31 July 2026. Figures relating to Situational Awareness are reported figures resting on unnamed sources, and are described as such above.
| Source | What it supports here |
|---|---|
| Berkshire Hathaway 2017 shareholder letter (PDF) | All three Buffett quotations, verbatim: the aversion to leverage dampening returns, the table of four Berkshire declines and the argument against borrowed money, and the point about those not handicapped by debt |
| TechCrunch, 30 July 2026 | The 439% first-half return, the reported assets peaking near $45bn and falling to about $10bn, the Citadel purchase, the four worst-hit holdings and their falls, the capex-versus-revenue explanation, and the retained private positions including Anthropic, MatX and Fluidstack |
| Investing.com on Berkshire’s cash position | The $397.4bn total at the end of Q1 2026, the roughly $339bn in Treasury bills, and the approximate 3.7% yield |
| CNBC on the 2026 Berkshire annual meeting | Greg Abel as chief executive from 1 January 2026 with Buffett continuing as chairman |
| Kingswell on chapter 20 of The Intelligent Investor | The margin of safety as Graham’s concept, and Buffett’s description of chapters 8 and 20 as the bedrock of his investing for more than 60 years |
How we verified this
Buffett’s words are quoted from the primary source. The three passages used here were read from the 2017 Berkshire Hathaway shareholder letter itself, not from secondary coverage of it, and the wording is verbatim. One line commonly attributed to Buffett — that leverage “produces zeros” even for very smart people — is not in that letter, and rather than source it to a quotation site, it is not used at all.
The margin of safety is attributed where it belongs. It is Benjamin Graham’s concept, set out in chapter 20 of The Intelligent Investor. Buffett’s role is that of its most prominent adherent: he has called that book the best written on investing and described chapters 8 and 20 as the bedrock of his investing for more than 60 years. The article does not present the idea as his invention.
The Situational Awareness figures are reported ones, and are labelled as such throughout. The 439% first-half return comes from an investor letter reported by the Financial Times; the forced liquidation and the Citadel purchase were reported by the Wall Street Journal; the assets under management figures vary between outlets and across time, which is why a range is given rather than a single number. All of it rests on unnamed sources. Aschenbrenner has not commented publicly and had not responded to press approaches at the time of writing. Nothing here is drawn from a regulatory filing describing these events, because none has been identified.
Berkshire’s cash position is the $397.4 billion reported at the end of the first quarter of 2026. Its second-quarter figures are due within days of publication and will supersede it.
What this article does not do: declare a winner, or offer any view on what a reader should buy, sell or hold. It compares the risk structures of two publicly documented approaches. The outcome of one month is not treated as a judgement on either.