Positions
$120B
A forecast, written the year he left OpenAI, arguing that almost everyone was underrating both the speed and the physical scale of what was coming.
The call was good. The constraint on AI isn't models — it's power, memory and concrete. So he bought the bottleneck and shorted the software that agents were going to eat.
439% net. A 165-page essay turned into several hundred million, then several billion.
Other funds copied the trades. So a good part of the run in neoclouds and semis was his own crowd buying in behind him — which is wonderful, right up to the morning everybody needs the same door.
None of that saved him. Hold onto that.
Leverage is borrowing money so you can bet more than you own. Four turns means four dollars of position for every dollar that is actually yours.
$10bn came in. It became $30bn of equity — his money, and the only money that can absorb a loss. He ran it at four turns, so he was steering $120bn of positions. $90bn of that was borrowed.
Up, it's beautiful: a 5% move on the book lands as +20% on him. Down, it's just arithmetic. A 25% move and he is not hurt. He is gone.
A margin call is not the market punishing you. It is your lender deciding your cushion got too thin, and demanding cash you do not have.
Here is the whole mechanism, in his actual numbers.
13F filings are public. Six names carried 85% of his disclosed book. He owned 8.2% of Core Scientific outright.
Read that as a regulatory filing and it's compliance. Read it as a shopping list and it's something else. You cannot leave a position that size quietly — there is no version of that exit nobody notices.
Both legs lost at once. Infrastructure sold off while the software shorts rallied. You cannot hedge being wrong in both directions on the same afternoon.
And once the street knows a man has to sell, it sells what he owns and shorts the rest. Nobody has to collude. Everybody can read.
On 24 July he wrote to investors. Up 439%. The sell-off, he said, had produced some of the most attractive opportunities since early 2025. He invited fresh capital from 1 August.
The book was liquidated on 30 July.
Six days. What he knew when he wrote it isn't on the record and I'm not going to pretend otherwise. The dates are on the record.
Citadel took the whole book at a rumoured 20–50% discount. That discount is the entire trade.
They're market-neutral, so they were already hedged against the thing that killed him. Nobody can hold $120bn of exposure forever — but they can bleed it out slowly, and he had until Thursday. One party had time. The other had a deadline. The price is just the difference between those two facts.
The episode likens Griffin to Buffett in 2008, to JP Morgan in 1907 — the man who steps in and absorbs the risk before it becomes everyone's problem. That comparison is fair, and it's worth saying plainly what it describes: in each case the man who stabilises the system is the man who ends up owning it cheap. Nobody rescued anybody. They priced it.
And note what didn't go.
The banks took the public book because the public book was the part they could sell. The Anthropic stake — around $5bn, from a May round valuing the company at $965bn, with an IPO possibly as soon as October — wasn't in the sale.
The illiquid asset survived precisely because it was illiquid. You cannot margin-call a private mark. So the investors' liquid equity is gone, and the position that might print money later is still there. He may yet have a decent year.
One question decides it: was the failing thing collateral for everyone else?
LTCM was — the currency market couldn't absorb the exit. Bear Stearns was — its paper sat in everyone's trades. Archegos wasn't; it was tech stocks on leverage, and it still took Credit Suisse down with it eventually.
This one isn't. A stock drawdown is not a credit event. Not the same animal.
CORRECTION — the episode dates LTCM to 1994 and has it causing the Asian financial crisis. 1994 is when LTCM was founded; it collapsed in September 1998, and the Asian crisis ran through 1997, ahead of it. The argument still holds. The chart uses the corrected dates.
With the forced seller gone, there's no mechanical reason left to short any of it. Nasdaq +2.8% Thursday. Korea ripped.
Everyone wants this to be the market calling time on AI. It isn't. It's a leverage story that happened to be wearing an AI-shaped book.
Four turns. Six names carrying most of the risk. Stakes too big to exit quietly. A short leg correlated to the long leg. Hedges built for the wrong disaster. A public filing telling everyone where to aim.
Those are construction failures. They'd have killed him just as fast betting on railways.
Nothing in the reporting alleges fraud. Being catastrophically wrong with borrowed money is not a crime. It is a Tuesday.
Note the sequence. The letter went out on 24 July, inviting capital from 1 August. The book was liquidated on 30 July. The fundraising window opened after the fund had already been sold.
And the Anthropic stake is still his. The banks took what they could sell; the private mark survived because nobody can margin-call it. His outcome and his investors' outcome are not the same outcome, and never were — management fees came off the top the whole way up.
He raises again. He is in his mid-twenties, no misconduct is alleged, and 439% is permanently on his record — the tape doesn't care that it was followed by a zero. The essay hasn't been refuted by any of this either; if anything the AI buildout kept validating it while his book was being sold. The pitch writes itself: right about the thesis, wrong about the sizing, and I have now had the education.
The precedent is exact. John Meriwether ran LTCM into a $3.6bn rescue by fourteen banks in 1998, brokered by the New York Fed. He opened JWM Partners in 1999 with $250m — the same strategy, less leverage. It lasted until 2009.
The people who don't get a second fund are the investors. That asymmetry isn't a scandal. It's the product.
Being right about the technology and staying solvent long enough to collect are two different jobs. He was exceptional at the first one.
Griffin was never smarter about AI than Aschenbrenner. He just never had to sell. In this business that is not a footnote — it's the entire skill, and it's the one nobody writes a 165-page essay about.
| Figure | Value | Source | Status |
|---|