The Forced Seller Has a Name: Situational Awareness Dumps Its Entire Public Book

Article Ticker Tape

Summary

Before Thursday's open, Situational Awareness — Leopold Aschenbrenner's AI hedge fund, roughly $45 billion at the start of July after a +439% net first half — sold the bulk of its public equities portfolio in one block, CNBC reported. By Thursday afternoon the buyer had a name: Citadel, per the Wall Street Journal. The Financial Times reports the fund is now seeking fresh capital, and Bank of America, Goldman Sachs, and JPMorgan have been working with it to meet margin calls. The five-week crash in memory and AI-infrastructure stocks now has a mechanical explanation attached to a name — and the seller is done.

Both legs failed at once

The fund ran the consensus AI barbell at size: long the compute build-out, short the software it was supposed to strand. In July both legs went against it simultaneously — the longs crushed in the memory rout while the Adobe-style shorts squeezed 20%-plus. A margin call on one side can be met by the other; a margin call on both sides can only be met by selling everything.

Position Price (Jul 29 close) 30-day From 52-wk high NOTE
SanDisk (long) $1,015.89 −55% −57% RSI 33 — the hardest-hit large long
CoreWeave (long) $60.82 −39% −60% RSI 29, deepest drawdown of the book
Nebius (long) $148.22 −46% −51% AI-infra long, no earnings offset this week
Micron (long) $739.00 −36% −41% fell through its own support even as earnings case improved
SK Hynix (long, US line) $126.79 −25% −35% the KOSPI epicenter — biggest long, cascading market
Adobe (short) $263.43 +28% −29% RSI 69 — the short leg squeezing while the longs bled

The short leg deserves its own autopsy. The software names didn't merely fail to fall — they ripped, and partly because the trade was crowded. Long-AI-infrastructure, short-legacy-software was the consensus expression across the levered complex, so every book forced to de-gross had to buy its shorts back, and the covering itself became the rally that tightened the next book's margin call. Some of the move is genuine rotation — capital fleeing AI infrastructure landed in the software it had spent a year betting against. What it is not is vindication for the shorted companies: Adobe still sits 29% below its 52-week high. The short thesis wasn't crazy; the expression was crowded and levered, and crowded, levered trades don't get graded on whether they eventually prove right.

Three reads

The violence finally makes sense. US memory names down 35–55% in a month, two KOSPI circuit-breaker days, indiscriminate selling across every AI-adjacent line — that is what a levered multi-tens-of-billions book de-grossing into a falling market looks like. Forced selling is liquidation mechanics, not information about the companies being sold.

It rhymes with Korea's retail wipeout, because it is the same trade. The fund's largest long was SK Hynix — the center of the Seoul cascade, where single-stock leveraged ETF holders were carried out the same week. Levered conviction on the memory supercycle, at every size from retail to $45 billion, hit the same margin clock at the same time, and each seller's exit deepened the next one's call.

The thesis and the solvency parted ways in the same week. Samsung printed a record quarter — operating profit up roughly 1,800% year over year — and guided that the 2027 memory supply-demand gap widens on orders already booked. SK Hynix printed a record Q2 days earlier. The fund's core view was arguably being confirmed by the producers in the very week its leverage forced it out. Markets can stay irrational longer than a levered book can stay solvent — the oldest line in the business, running on schedule.

Six days from "buying opportunity" to the block

The sequence deserves its own timeline, because it is the whole story of leverage in four dates. The fund launched in 2024 and compounded over 1,000% net after fees, per the Wall Street Journal — roughly 270% in 2026 through May, 439% through June — a track record that pulled the assets from $1.5 billion to over $20 billion in under two years and brought in Jane Street, a firm that almost never backs external managers, as an investor. On July 24, with the drawdown underway, Aschenbrenner wrote to investors that the rout was a buying opportunity — "a particularly good time to add funds, if you have been waiting for one" — pointed to a potential Anthropic IPO as a coming catalyst, and asked for commitments by August 1. On July 30, the entire public book was sold before the open. The deadline in his own letter never arrived. He may even have been right about the opportunity — but at the leverage he was running, the view belonged to his brokers, not to him.

The buyer, the Journal reported Thursday afternoon, was Citadel — with Millennium submitting a competing bid, and the price undisclosed. The structural read held: absorbing a levered book in one print is a trade only a handful of firms on earth are built to price, and Ken Griffin's $71 billion machine is one of them. Citadel bought at the moment of maximum forced selling, which is historically the best seat at the table. The fund itself retains roughly $10 billion, per Reuters — some remaining stock positions and its private investments, including the Anthropic stake, which was not sold.

What changes now

A completed block sale is an overhang-clearing event: the biggest forced seller in the AI complex has no more stock to sell, and the buyer chose the risk at these prices. The market agreed loudly and immediately: the day the buyer was named, the unwind cohort printed its biggest up-day of the entire rout — SanDisk +26%, Nebius +27%, CoreWeave +22%, Micron +18%, SK Hynix's US line +18%, the Korea ETF +12% — while Adobe, the squeezed short leg, gave back 6%. Both sides of the barbell reversing together the day the forced flow ended is exactly what the liquidation-mechanics read predicted. One up-day is still one bar, not a base.

What the sale does not settle is whether other levered AI books face the same calls — Goldman and JPMorgan have reportedly issued margin calls to funds with concentrated AI positions. The Anthropic-stake question, at least, is answered for now: the fund kept it, retaining roughly $10 billion in remaining assets per Reuters. Watch for follow-through in the memory cohort — a base, not a bounce — and watch equally for a second name.

The standing view on these names is unchanged from the July 6 SK Hynix listing report: the memory-tightness case is producer-confirmed, entries wait for a base to form, and a fire-sale morning is the wrong moment to mint a new call. Let the block clear.

Sources