Leopold Aschenbrenner

leopold aschenbrenner was, until about a week ago, the closest thing the ai trade had to a folk hero. a former researcher, still in his mid-twenties, who wrote a document that half of silicon valley claimed to have read, then turned that document into a hedge fund that went up more than a thousand percent. and then, over the course of a few days at the end of july, most of it came apart. that's the arc. but the arc only makes sense if you understand the person, so let's start there.

he was born in germany, around 2001 or 2002, in berlin, to two parents who were both doctors. he went to the john f. kennedy school in berlin, the bilingual german-american school, and then he left for new york. he enrolled at columbia and graduated in 2021 as valedictorian, at nineteen, with a degree in economics and mathematics-statistics. that's not a typo. nineteen. he skipped years the way most people skip meetings.

after columbia he spent time doing long-term risk research at oxford's global priorities institute — this is the world of effective altruism, of people who think seriously and sometimes obsessively about the far future and how present decisions ripple into it. and that world had a very specific gravitational center in 2022: sam bankman-fried and ftx. aschenbrenner joined the ftx future fund, the philanthropic arm bankman-fried set up, in february of 2022. by the reporting, he was young, trusted, and close to the money — one account has him helping run a charity out of a penthouse in the bahamas. he resigned in november, right before ftx collapsed into one of the largest frauds in modern finance. he wasn't accused of wrongdoing. but it's the first time you see the pattern that follows him around: brilliant, early, adjacent to a spectacular blow-up, and out the door just before or just after the walls come down.

next stop was openai. in 2023 he joined the superalignment team — the group tasked with the almost science-fictional problem of how you control an ai system much smarter than yourself. and this is where he becomes a public figure, because in april of 2024, openai fired him. the company's stated reason was an information leak. aschenbrenner tells a different story, and he told it at length on the dwarkesh patel podcast. his version: he'd written a memo to openai's board arguing that the company's security was — his word — "egregiously insufficient" to stop foreign actors, china specifically, from stealing model weights and algorithmic secrets. he says the leak they cited was a harmless brainstorming document. he says that when he was let go, a lawyer questioned him about his views on agi, on whether the government should be involved, and about whether he and his team were loyal to the company. openai has denied that it fired him for raising security concerns. you don't have to pick a side to notice that the firing did something useful for him: it made him a dissident. and a dissident with a thesis is a much more compelling figure than an employee with one.

because he had a thesis. in june 2024, a couple of months after the firing, he published it. it's called "situational awareness: the decade ahead." it's about 165 pages across five essays, released free online, and dedicated to ilya sutskever, openai's former chief scientist. it argued, essentially, that superintelligence was coming this decade, that almost nobody was pricing it in, and that the people who could see it clearly — who had, in the military sense, situational awareness — were a tiny minority sitting mostly in san francisco. we'll get into the actual mechanics of the argument later, because they matter and they're more careful than the headline makes them sound. but the short version is: he wrote the bull case for ai as a physical, industrial phenomenon, and he wrote it with charts.

and then he did the thing that separates him from every other person who published a viral ai take that year. he put money behind it. in the second half of 2024 he launched a hedge fund, and he named it after the essay: situational awareness lp. he seeded it with about 225 million dollars, and the backers were not randoms. the collison brothers, patrick and john, who built stripe. nat friedman, the former ceo of github. the investor daniel gross. he took the title of chief investment officer and managing partner, and he co-managed the fund with carl shulman, another figure from that same long-termist intellectual world. a twenty-two-year-old with no professional investing record was suddenly running money for some of the most respected names in technology, on the strength of an essay.

and for a while — a remarkable while — it worked. in the first half of 2025 the fund returned about 47 percent after fees, well ahead of the s&p and the tech indices. by late 2025 it was managing around 1.5 billion dollars. and then it went vertical. through the first half of 2026, the fund reportedly returned 439 percent, net of fees. since inception, more than a thousand percent. jane street — the secretive quant trading firm that almost never backs an outside manager — came in as an investor. the assets ballooned from that 1.5 billion to, by the start of july 2026, a peak of roughly 45 billion dollars. he had become, on paper, one of the most successful new fund managers in a generation. the press started calling him the "nostradamus of ai." he got engaged to avital balwit, the chief of staff to dario amodei, the ceo of anthropic. the ai power couple. that's the top of the mountain. now here's the part that put him on everyone's screen this week.

at the end of july 2026, over a handful of trading days, the fund lost most of what it was. and it happened in the specific, mechanical, almost boring way that these things always happen — not because the thesis was disproven, but because of leverage and a margin call.

here's the sequence, as it's been reported by cnbc, the wall street journal, and the financial times, which reviewed the fund's investor letter around july 24th. situational awareness was running leverage of up to 400 percent — roughly four times its capital, financed with borrowed money. it was concentrated, and it was concentrated in ai infrastructure. then, in july, ai infrastructure stocks sold off hard. the philadelphia semiconductor index fell about 28.6 percent from its june 22nd peak. one momentum index of hot tech names dropped more than 50 percent. when you're levered four to one, you do not need a crash to get wiped out. a drawdown is enough. as the fund's holdings fell, its prime brokers — goldman sachs, jpmorgan, and bank of america, the banks that extend the leverage — issued margin calls. they wanted more collateral. and aschenbrenner didn't have it in the form they needed.

so on thursday, july 30th, he sold. not trimmed — sold. the fund offloaded its entire book of public stocks to ken griffin's citadel, in a single distressed block, at below-market prices. reporting puts the amount citadel scooped up at around 16 billion dollars in holdings. the fund's assets fell from that 45-billion-dollar peak to somewhere around 10 billion. call it 35 billion dollars of assets gone in days. what's left is essentially a large private stake in anthropic — reportedly around 5 billion dollars — and a fund that will now operate as a quieter, private vehicle rather than the levered public-markets rocket it had been. and the timing was almost unbelievably cruel: it happened the same weekend he was getting married, in carmel, california.

now, one detail that tells you why this hurt so much. the losses came from both sides of the book at once. on the long side, he owned the picks-and-shovels of ai — memory and chip names like sk hynix, micron, and sandisk, gpu clouds like coreweave and nebius, power plays like bloom energy, even bitcoin miners pivoting into ai data centers, names like iren and core scientific and applied digital. those got hammered in the selloff. and on the short side, he was betting against software incumbents — companies like adobe, on the theory that ai would erode them — hedged with something like 8.5 billion dollars of put options. those shorts moved against him. so the hedges didn't hedge. both ends of the trade lost at the same moment. that's the nightmare scenario for a concentrated book, and it's a very particular kind of pain, because it means the diversification you thought you had was an illusion. everything he owned was really one bet: ai capital expenditure keeps going up and to the right, without interruption. and for a few weeks in july, it didn't.

the reaction on wall street was less shock than i-told-you-so. one coaching-firm founder, jerry diao, put it bluntly: a lot of people, he said, saw this as a matter of not if, but when. and the sharper version of the critique is worth sitting with, because it's the whole story in one line — maybe his views on ai are right in the long run, but in the public markets you have to survive the short run. that's the gap this whole episode lives in. so let's actually go under the hood of what he believed, because the thesis is genuinely more rigorous than "ai go up," and understanding it is the only way to understand what did and didn't fail.

the intellectual engine of "situational awareness" is a move aschenbrenner calls "counting the ooms." oom means order of magnitude — a factor of ten. ten x is one oom, a hundred x is two ooms, and so on. his claim is that ai capability tracks something he calls effective compute, and that effective compute has been growing at a predictable rate that you can decompose into three separate streams, which multiply together.

stream one is raw physical compute — more chips, bigger clusters, more flops thrown at training. he pegs that at roughly half an order of magnitude per year. stream two is algorithmic efficiency — the fact that we keep finding better ways to train, so you get more capability out of the same hardware. also, roughly, half an order of magnitude per year. and stream three is the one that's easy to miss and, i'd argue, the most interesting: he calls it "unhobbling." this is the gap between what a model can do in principle and what it can actually do as a usable tool. a raw language model that predicts the next token is hobbled. wrap it in chat, in tool use, in the ability to plan and act over many steps as an agent, and you unlock capability that was already latent in the weights. unhobbling is the move from a chatbot that answers a question to an agent that does a job.

stack those three streams and, aschenbrenner argues, you get a very steep line. his anchor is the jump from gpt-2 to gpt-4. he characterizes that as going from roughly the ability of a preschooler stringing sentences together to something like a smart high-schooler — and that took about four years and several orders of magnitude of effective compute. then he extrapolates. he projects something like five more orders of magnitude — a hundred thousand times more effective compute — by 2027, which he claims would deliver another qualitative leap the size of the gpt-2-to-gpt-4 leap, stacked on top of gpt-4. and a jump that size from where gpt-4 already sat lands you, in his framing, at ai that can do the work of an ai researcher or engineer. that's his line — that agi by 2027 is, in his words, "strikingly plausible." and note the honesty in the phrasing. he doesn't say certain. he says the trendlines, extended, make it something you can't dismiss.

then comes the part that turns a forecast into an ideology. if you get ai that can do ai research, you don't get one more ai researcher — you get hundreds of millions of them, running in parallel, faster than humans. and they go to work on the very problem of making ai better. aschenbrenner argues that could compress a decade of algorithmic progress into a year — the classic intelligence explosion — taking you from roughly human-level to vastly superhuman in something like twelve months after you hit agi. that's the "from agi to superintelligence" chapter, and it's where the essay stops being an investment memo and starts being an eschatology.

and here's the bridge to the money, because this is the crucial move. if all of that compute has to physically exist, then somebody has to build it. he watched, in his words, the boardroom conversations in san francisco go from ten-billion-dollar compute clusters to hundred-billion-dollar clusters to trillion-dollar clusters — another zero every six months. these clusters need power measured not in megawatts but in gigawatts — ten gigawatts, then a hundred gigawatts, the output of large numbers of power plants. so the essay isn't really about software at all. it's about a physical build-out: chips, high-bandwidth memory, data centers, transformers, turbines, transmission lines, electricity. the bet writes itself. don't try to guess which chatbot wins. own the physical layer that every chatbot needs. that's the second-derivative trade — not ai itself, but the arms dealers to the ai war. and that is exactly the book the fund held. long the memory makers and the power companies and the gpu clouds. it is an intellectually coherent, even elegant, expression of the essay.

he was honest about the risks in the text, too, and this matters. he flagged the "data wall" — the problem that we may run out of high-quality human text to train on, which would require genuine algorithmic breakthroughs like synthetic data or self-play to get past. and he flagged the geopolitics — a us-china cold war over ai, the theft of model weights, the case for treating frontier labs like national security assets. which, remember, is the same concern that got him fired from openai in the first place. so the worldview is internally consistent from the security memo all the way through to the fund's positioning. it is one continuous argument.

so what actually broke? not the thesis, at least not yet. here's the distinction i'd urge you to hold onto, and it's the whole lesson. being right about a technology and being right about an asset over a given horizon are two completely different skills. the essay could be entirely correct that ai drives enormous demand for compute, memory, and power over the decade — and the fund could still be destroyed in a single month. because the fund wasn't a bet on the technology. it was a bet on the technology, times four, with no room to be early. leverage is a clock. it converts "you'll be proven right eventually" into "you have to be right by the margin call." and at four-to-one, a drawdown in the low twenties in percentage terms is enough to threaten your entire equity. the semiconductor index fell about 28 percent from its june peak. the math was never going to survive that, no matter how good the essay was.

so let me pull this all the way back to the "so what," because you didn't come here for a morality tale about a young man and some leverage. you came for what it means for how you build and run things. here's what i'd actually take from it if i were sitting in a cto's chair.

first: separate the map from the trade. aschenbrenner's map of ai as an industrial, power-and-silicon phenomenon is, i think, largely sound, and it's a genuinely useful frame for planning. the demand for memory, for data-center capacity, for electricity, for cooling — that's real, and the reasoning behind it — counting the ooms, the unhobbling of models into agents — is the kind of thing worth understanding deeply, because it shapes your compute costs and your vendor roadmap for years. the trade around that map is a separate object. this week didn't falsify the map. it falsified a particular, over-levered, over-concentrated way of expressing it. don't let the second thing make you cynical about the first.

second, and this is the operational one: watch the physical bottlenecks, because that's where his map is most likely to be right and where it bites you directly. the constraint on ai over the next few years is looking less like clever architectures and more like memory supply, power, and interconnect. if a 45-billion-dollar fund was betting the house on high-bandwidth memory and gigawatts of electricity, that tells you where the scarcity is. so when you're planning capacity — reserved instances, long-term compute contracts, where you site anything that needs serious power — assume that supply is tight and getting tighter, and that the price of compute is more volatile than your finance team wants it to be. this month's selloff also cuts the other way: capital that funds the build-out is skittish, and a pullback in ai capex financing can slow the very data-center buildout you're depending on. plan for both the scarcity and the whiplash.

third: the correlation lesson, translated out of finance and into engineering. aschenbrenner's hedges failed because everything he owned, long and short, was really the same bet on ai sentiment. his diversification was fake. you have the exact same risk in your architecture and your vendor stack. if your inference, your training, your tooling, and your key startup dependencies all rest on the same two or three suppliers and the same narrow slice of the market, you do not have a resilient system — you have one bet wearing the costume of many. the discipline is the same in both worlds: find the hidden single point of failure that your apparent variety is concealing.

fourth, on hiring and how you weigh people. this whole saga is a very expensive demonstration that being spectacularly right about the future and being able to execute under real-world constraints are different capabilities, and that the first one is much easier to fake and much easier to fall in love with. aschenbrenner is clearly brilliant. he saw the shape of the ai build-out earlier and more clearly than almost anyone, and he wrote it down. and he still got taken apart by risk management — by position sizing, by leverage, by the unglamorous discipline of surviving a bad month. when you're hiring your senior technical people, and especially when you're listening to the confident visionary in the room, ask who on your team owns the boring survival questions. what happens when this is down 30 percent. what's the blast radius when the dependency we all assumed was safe goes down. the visionary and the risk manager are rarely the same person, and the failure mode of a company — like the failure mode of this fund — is usually letting the first one operate without the second.

and fifth, the honest, uncomfortable one. this is at minimum a warning about the ai trade's fragility, and possibly an early tremor. jim cramer, of all people, called the forced selling a "clearing event" — the idea being that a big levered player getting flushed can mark a bottom rather than a top. maybe. or maybe it's the first crack. i don't know, and anyone who tells you they know for certain is selling something. what's worth holding in your head as a cto is that the enthusiasm funding a lot of the ai build-out is levered, concentrated, and emotional, and that the infrastructure your roadmap quietly depends on is being financed by exactly that kind of money. that's not a reason to stop building on ai. it is a reason to make sure your plans survive a scenario where the capital gets scared for a couple of quarters. build so that a funding winter is survivable, not fatal.

so where does that leave leopold aschenbrenner. he's twenty-five. he still holds a multi-billion-dollar stake in anthropic. he got married. and one financial times piece framed it neatly — that he got the future right and misread the past, misread the plumbing of how markets actually punish leverage. the essay may well still come true. the models may keep climbing the ooms, the clusters may keep adding zeros, agi may or may not arrive on his timeline. but the fund named after his certainty got taken out not by being wrong about ai, but by being unable to wait to be proven right. and if you build things for a living, that's the sentence to write on the wall. the market can stay skeptical, and the margin call can arrive, long before the future you're sure about finally shows up.