Dumb Money

THE NUMBER: 439%. That’s what Situational Awareness returned in the first half of 2026 — the best print on Wall Street in a generation, maybe ever at that size. Sit with the number the way an allocator sits with it, because it was never a bragging right. It was a warning label. Nobody compounds 439 percent in six months unlevered. The same borrowed money that made the return made the margin call, which means the return and the blowup were the same number wearing different clothes. Four weeks after that print, the fund’s prime brokers sold its entire public book to Ken Griffin’s Citadel before the Thursday open, and a $45 billion fund marked itself down to about $10 billion. Read the number the way the pros read it: when someone shows you a return that good, the first question is never “how do I get in?” It’s “what did they borrow to get it?”

There’s a stretch in the middle of Dumb Money — the 2023 GameStop movie — where Seth Rogen, playing Melvin Capital’s Gabe Plotkin, is watching his fund bleed out in real time while his kids play in the next room. The retail traders on the other side are being called the dumb money on every cable channel. And then the two most feared men in the industry, Ken Griffin and Steve Cohen, wire $2.75 billion into Melvin to stop the bleeding. The pros save the pro. The amateurs get their trading buttons turned off. That’s the movie, and that was basically 2021: the dumb money label got assigned by who had a Bloomberg terminal, not by who did something dumb.

Five years later Griffin’s phone rang again, and this version is better than the movie. The fund bleeding out wasn’t run by a grizzled short-seller caught leaning on a mall retailer. It was run by Leopold Aschenbrenner — 25 years old, Columbia valedictorian at 19, fired from OpenAI, author of the 165-page essay that became the intellectual scripture of the entire AI infrastructure trade. The man who told everyone, correctly, that the buildout was coming. And this time Griffin didn’t bail anyone out. He bought the book at the lows.

I spent a career on the other side of this table — analyst, long/short fund, credit work — and I want to walk you through what actually happened here, because the X version of this story is a conspiracy and the real version is more useful. The real version is a tale as old as time. People are right until they’re not.

🩳 The Signal: The First Forced Seller of the AI Era

The facts first, because they’re better than fiction. Situational Awareness launched in 2024 with about $225 million from a murderers’ row of backers: the Collison brothers, Nat Friedman, Daniel Gross, and — the detail that should have gotten more attention — Jane Street, a firm that almost never gives outside managers a dollar. The thesis was the essay: AGI is coming, therefore chips, memory, data centers, and power are structurally underpriced. Eight employees. Four investment professionals. By early July 2026 the fund was worth $45 billion and had returned 439 percent in six months, per the reporting at CNBC and the FT.

Then July happened. The AI infrastructure complex rolled over, and the fund’s four largest disclosed holdings — Nebius, Sandisk, Micron, CoreWeave — all fell more than 35 percent in a month. Nebius dropped roughly 48 percent from its peak, about $35 billion of market cap gone from one name. Sandisk lost 56 percent in about a month. Meanwhile the short book (legacy software names like Adobe) squeezed the wrong way. Long AI, short software: wrong on both sides at once, with borrowed money on both sides at once.

By Wednesday the FT reported the fund was calling investors and lenders trying to raise fresh capital, even offering pieces of the portfolio directly to its own LPs. That is not a sign of strength; that’s the sound of a bank run starting politely. By Thursday morning it was over: Bank of America, Goldman, and JPMorgan — the prime brokers who lent the money and the shares — had marketed the whole public book, roughly $16 billion of longs and shorts, and Citadel took the bulk of it in a block before the open. Millennium bid too. The Wall Street Journal noted the old benchmark for a trading disaster was Archegos, which lost $8 billion in ten days in 2021. If the current numbers hold, this one roughly tripled it.

What survived: a stake in Anthropic worth about $5 billion, which the fund was reportedly shopping as well. Hold that thought — it matters for the ending.

🔍 Anatomy of a Blowup (No Conspiracy Required)

Now the mechanics, because this is where the X discourse went sideways. The popular theory, running at 500K-plus views: Citadel talked up surprise-rate-hike fears into Tuesday’s Fed meeting, the market sold off, Leopold got margin-called, and Citadel scooped the book at the bottom. Diabolical genius, rug pull, etc.

Look — you don’t need it. Here’s the boring version, which happens to be the true version, which I have watched from the inside more than once.

If you run a fund that size, you file 13Fs with the SEC every quarter. Your longs are public with a lag, and this wasn’t just any book — it was the most talked-about trade of the year, attached to the most famous thesis in the industry. Long AI infrastructure, short old software. Everyone knew it. And the moment those positions started moving the wrong way, every desk on the Street could see exactly what he owned and exactly where it hurt. That’s when the wolves come. They lean on your longs. They buy your shorts. They don’t need inside information; they need your 13F, a screen, and patience, because the margin math does the killing for them.

The math is worth doing once, slowly. Say you’re running $40 billion and $10 billion of it is an illiquid Anthropic stake. Call it $30 billion of liquid capital. Lever that book — reports put it near four turns gross — and a 30 percent move against you doesn’t dent your equity. It deletes it. And long before you reach zero, the people who lent you the money and the shares stop being your partners and start being your liquidators. The primes step in and sell the book for you, in whatever order protects them. Orderly liquidation becomes disorderly liquidation becomes a run on the bank. We watched that sequence play out across tech trading desks over the past few weeks; Thursday was just the day it got a name attached.

Keynes wrote the epitaph a century ago: the market can stay irrational longer than you can stay solvent. Buffett wrote the punchline: when the tide goes out, you find out who’s been swimming naked. The tide went out in July. Leo was not wearing trunks.

And the cruelest detail is that he knew. “Obviously, not blowing up is task number one and two,” he told Dwarkesh Patel in 2024, back when the fund was new and the interviews were victory laps. He said the sequence of bets on the way to AGI was critical, that people underrated timing. He was right about that too. Knowing the rule and living the rule are different jobs, and the second one doesn’t care how smart you are.

🃏 Griffin Always Answers the Phone

One more thing the conspiracy theorists get backwards. Ken Griffin buying a blown-up book at the lows is not a plot twist. It’s the business model, and it’s at least the third time he’s run it.

In 2006, when Amaranth Advisors vaporized $6.6 billion on natural gas spreads — the record blowup of its era — Citadel and JPMorgan took over the energy book at distressed prices. In 2007, with the mortgage market seizing up, Citadel wired about $2.5 billion into a dying E*Trade and took its toxic securities portfolio at pennies on the dollar while everyone else was running for the exits. In 2021, Citadel and Point72 put $2.75 billion into Melvin. And now, in 2026, Citadel takes the Situational Awareness portfolio in a block. Four crises, four different asset classes, one trade: be the one with dry powder when the levered believer gets the call. The house doesn’t need to believe your thesis. The house needs you to need cash on a Thursday morning.

That’s why the Dumb Money title cuts both ways. Griffin has spent his whole career being the counterparty to conviction. He reportedly thinks the AGI fervor is overdone — and it didn’t matter, because at the right price you don’t buy the thesis, you buy the book. He got a few billion dollars richer this week off the true believers. That should roughly cover the pied-à-terre tax Albany has been trying to invent ever since he paid $238 million for the penthouse on Central Park South, with enough left over to buy out another block of Brickell for the Miami headquarters. Or both.

📈 What Didn’t Break

Here’s the part every “AI bubble popped” take will miss on Friday, and it’s the part that actually matters for you.

Nothing about the technology broke this week. Azure just closed a fiscal year past $100 billion in revenue, growing 43 percent. Google Cloud grew 82 percent. ChatGPT is closing on a billion weekly users — a faster climb than TikTok, Instagram, or YouTube. The demand curve that Aschenbrenner’s essay predicted is the one thing in this story performing exactly to thesis.

What broke is a financing structure. A four-person investment team ran a levered, crowded, publicly-visible portfolio against a four-week window, and the window won. We’ve seen this movie before — in 2000 the telecom bust wiped out the tourists while internet traffic kept doubling on schedule. The trade and the technology are different animals, and the market just spent July teaching that lesson at a $35 billion tuition.

But keep your eye on the slow-motion version of the same story, because it’s still running. Nvidia’s credit default swaps doubled this month to a record 82 basis points after reports it would backstop $250 billion of OpenAI data center leases. S&P has Oracle one notch above junk, with roughly half its $638 billion backlog resting on OpenAI. Microsoft’s backlog hit $678 billion — up 25 percent if you exclude OpenAI. As Tomasz Tunguz put it this week, the same borrowed dollar is showing up as backlog on more than one balance sheet. A week ago, in The Money’s Not Here, we said the AI economy’s paper value works until everyone wants it back on the same afternoon. Thursday was one fund’s afternoon. The corporate version doesn’t get margin-called in a morning — it gets repriced in basis points, quarter by quarter.

🧭 Two Investors, Same Facts

The most useful contrast of the week arrived, fittingly, in a tweet. Gavin Baker, a survivor of several cycles at Fidelity and now running Atreides, laid out his frame: the models themselves go to zero, data centers are commodity, energy is commodity, and only two things in the whole stack hold value — proprietary data and reinforcement learning. What he actually buys is utilization: take a GPU from 30 percent utilized to 60 and you’ve doubled the output of the AI factory without buying a single new chip.

Same facts as Leopold. Opposite portfolio. One man bought the buildout with borrowed money and a deadline. The other buys the efficiency of the buildout with discipline and no forced clock. You don’t have to agree with Baker’s survivor list (he keeps Google, xAI, and Anthropic; he’d tell you the rest is commodity) to steal the frame: don’t rent conviction with leverage. Own the things that compound whether or not the thesis arrives on schedule — your data, your workflows, your utilization.

That translates directly if you run a business instead of a book. Your “leverage” isn’t a prime broker; it’s the single vendor your operations can’t run without, the customer funded by venture debt, the comp plan denominated in AI stock. Your “utilization” is whether the AI you already pay for is actually working your workflows or sitting at 30 percent. The audit is the same one the primes ran on Leo. Better you run it than someone else.

🛠️ Don’t Weep for Leo

And truly, don’t. He’s engaged to Anthropic’s chief of staff. He still owns about $5 billion of Anthropic — the illiquid position, the one the primes couldn’t seize and sell by Thursday, which is its own lesson about which assets survive a run. He hedged his own life better than his book: either AGI comes and his ideas made him rich, or it doesn’t and he’s still in his twenties and brilliant. By his own framework, he’s fine either way. The LPs who compounded 439 percent and then round-tripped most of it might phrase it differently at the wedding.

The lesson isn’t that he was wrong. That’s what makes this one sting. He was right about the buildout, right that timing mattered, right that not blowing up was the whole job — and he blew up anyway, because being right and staying solvent are different skills, and only one of them is taught by essays. The tide went out, the book was crowded, the leverage was wrong, and the man with the cash and no opinions ate the whole thing before breakfast.

Being right was never the job. Staying solvent is.

— Harry and Anthony

Sources:

• CNBC — Star AI investor Leopold Aschenbrenner is unwinding trades after steep losses
• The Verge — The loss of Situational Awareness
• Bloomberg — Situational Awareness drops to $10 billion as Citadel steps in
• Axios — AI-focused hedge fund sells all of its stocks
• TNW — Aschenbrenner’s Situational Awareness fund sells out to Citadel
• Tomasz Tunguz — Microsoft Resells the Frontier
• Yahoo Finance — Nvidia’s rising CDS the talk of Wall Street
• The Information — ChatGPT nears 1 billion weekly active users
• stark0xbt on X — Gavin Baker’s framework
• Dividend Hero on X — the Aschenbrenner arc
• CO/AI — The Money’s Not Here (July 24, 2026)