AI Fund Collapse: How Leverage Wiped Out $35 Billion

A Balance Sheet Built on Belief (0:00)

Hey, chibis! I'm Aiwee, and today we're talking about a forty-five billion dollar AGI fund that collapsed in a single block trade. If you enjoy stories like this, hit the like button and subscribe if you haven't already — let's go! Imagine a balance sheet built almost entirely on one question. How soon will machines match human reasoning?

Now imagine that question leveraged four to one, wrapped in options, and split into two sides that look like a hedge but move like one trade. This is not a story about a wedding in California, though the timing overlapped. It is a story about what happens when conviction in artificial general intelligence is treated as a portfolio architecture, and when that architecture meets a volatility spike it never priced in. Over roughly two weeks in mid-twenty twenty-six, that collision erased an estimated thirty-five billion dollars.

From Millions to Billions (1:00)

Act One: The Architecture Situational Awareness Limited Partnership launched in late twenty twenty-four. It was a public equities fund that reportedly grew from roughly two hundred twenty-five million dollars to an estimated forty-five billion dollars in under two years, staffed by about eight people, and led by a founder in his early twenties with minimal professional trading experience. The strategy was described as long-short. On the long side, it concentrated in the physical backbone of the artificial intelligence build-out. Chip makers like SK Hynix, cloud and power providers like CoreWeave and Nebius, and memory names like SanDisk dominated the book.

The six largest long positions reportedly accounted for most of the exposure. On the short side, the fund targeted enterprise software companies such as Adobe and Salesforce, betting that generative tools would erode their margins. At first glance, this looks like balanced positioning.

The Hedge That Wasn't (2:00)

But both legs were driven by the same underlying assumption. That artificial general intelligence would arrive on a specific timeline and rewrite the economy immediately. When market sentiment around that single narrative shifted, the longs and the shorts did not offset each other. They moved in lockstep against the fund. That is not a hedge.

It is the same trade expressed twice. Act Two: The Builder To understand why this structure was funded, look at the founder. Leopold Ashenbrenner was born in Germany to two physicians, entered Columbia University at age fifteen, and graduated in twenty twenty-one as valedictorian at nineteen. The economist Tyler Cowen had already publicly described him as an economics prodigy around age seventeen.

The Leaked Security Memo (3:00)

His early career included a philanthropic role at FTX, the crypto exchange that collapsed in late twenty twenty-two. He then joined OpenAI's Superalignment team, departing in April twenty twenty-four. OpenAI stated the departure followed leakage of an internal brainstorming document to outside researchers. Ashenbrenner has offered a different account, tracing the conflict to an internal memo he wrote about security risks, which he shared with board members. Two months after his exit, he published a roughly one hundred sixty-five page essay titled Situational Awareness: The Decade Ahead.

Its central thesis, presented as his own forecast rather than established fact, was that artificial general intelligence could arrive by twenty twenty-seven and that global preparation was inadequate. The essay was amplified by Tim Ferriss on The Tim Ferriss Show, where Ferriss referred to Ashenbrenner with a Nostradamus comparison and directed listeners to the document.

Conviction Attracts New Capital (4:00)

That attention helped attract Silicon Valley backers including the Collison brothers behind Stripe, former GitHub chief executive Nat Friedman, and investor Daniel Gross. The fund was named after the essay, and the pitch extended its argument. If the timeline is real, then traditional market pricing is wrong, and liquid equities offer the only venue large enough to deploy capital fast enough. Act Three: The Capital Flywheel The fund's growth from a few hundred million to tens of billions in under twenty-four months reveals a split in how capital evaluates risk. Silicon Valley limited partners wrote nine-figure checks based on thesis alignment and pedigree.

But when Ashenbrenner reportedly sought institutional capital in New York, the reception was cooler. Blackstone reportedly took a meeting and passed. At least one wealthy New York investor asked a basic risk management question.

Conviction Replaces Stress Testing (5:00)

What happens if the thesis is early, or if adoption slows, or if the timeline slips by even a few quarters? The answer, according to reporting, was not a detailed stress model but a restatement of conviction. This cultural gap is central to the episode. One ecosystem treated the fund as a bet on a worldview. The other looked for survival mechanics if that worldview took longer to materialize.

The choice of public equities over private venture deals also reflected scale ambition. As one investor reportedly explained, artificial general intelligence at that magnitude required markets where billions could be moved in and out quickly.

When Liquidity Turns Dangerous (5:45)

The irony is that liquidity became the mechanism of destruction. You can exit fast, but so can your prime brokers. When sentiment turned, that speed became a hazard. Act Four: The Trigger and the Mechanism The drawdown began in July twenty twenty-six. The catalyst was the release and rising credibility of lower-cost Chinese open-weight models, perceived as competitive with leading American frontier systems.

Markets repriced the scarcity value of the infrastructure stack. By the end of July, all six of the fund's largest long positions had fallen between roughly fifty percent and seventy-eight percent from their recent peaks. At the same time, the enterprise software shorts rose instead of falling. The supposed hedge had inverted.

The Overnight Block Sale (6:45)

Because the fund was reportedly running about four to one leverage, with options adding further exposure, the equity base could not absorb the synchronized shock. Prime brokers including Bank of America, Goldman Sachs, and JPMorgan issued margin calls. Around July thirtieth, roughly sixteen billion dollars of the fund's publicly traded stock portfolio was sold in a single overnight block trade. The buyer was Citadel, which reportedly won the package in an open auction at a discount of more than ten percent below market value. The estimated peak-to-trough loss was roughly thirty-five billion dollars, taking the fund from about forty-five billion dollars to about ten billion dollars in roughly two weeks.

The fund survived in part because an estimated five billion dollar stake in Anthropic was privately held and therefore illiquid. It was excluded from the liquidating block. That stake also carried a personal connection. Ashenbrenner's then-fiancée, Avital Balwit, is cited in reporting as chief of staff to the chief executive of Anthropic.

Collapse Beside a Wedding (7:45)

Days after the block trade, the founder reportedly sent a letter to clients acknowledging what he called market ructions, and noted that new subscriptions would be accepted beginning August first. The same week, he was married in Carmel, California at a multi-day event with panels and breakout sessions attended by Silicon Valley figures. The personal milestone and the professional collapse were simultaneous. Act Five: The Lessons The mechanical lesson is about pseudo-hedges. A long-short structure only reduces risk when the two legs respond to different factors.

Here, both legs were sensitive to the same variable. Artificial intelligence narrative intensity. When that variable reversed, correlation broke down in the worst possible direction. The mathematical lesson is about leverage and time.

Leverage Magnifies the Loss (8:45)

Even if the directional thesis eventually proves correct, a levered book can be destroyed by volatility before the thesis pays out. At four to one leverage, a fifty percent drop in an underlying name does not halve your equity in that position. It can erase it. And volatility does not scale linearly with returns. The deeper problem is that expected returns double when you double exposure, but the risk of ruin grows faster.

If your lenders force liquidation after a drawdown, you do not get to wait for the recovery. The cultural lesson is the distance between Silicon Valley conviction and Wall Street risk discipline. One rewards the strength of a worldview. The other demands proof that the fund can survive if that worldview is slightly wrong or slightly late. There is also a macro footnote.

AI Debt Reaches the Bond Market (9:45)

At the time of the drawdown, thirty-year United States Treasury yields were at their highest level since two thousand seven. This reflects a separate but related channel. The same artificial intelligence capital expenditure and borrowing cycle that funded companies like CoreWeave and Nebius also contributed to pressure on the long end of the yield curve. The fund's collapse and the bond market signal are not the same story, but they share a common root. An industry borrowing aggressively against a future that has not yet arrived.

Situational Awareness Limited Partnership did not fail because its founder was wrong about the future. It failed because its construction assumed the future would arrive on schedule and without turbulence. The fund's continued operation after the drawdown was not a vindication of the strategy. It was a contingent outcome dependent on a private asset being illiquid at the exact moment the public book was liquidated.

Belief Is Not a Portfolio (10:45)

That is luck, not architecture. For anyone watching the artificial intelligence trade, the episode is a reminder that a directional view and a tradable view are different instruments. You can believe in the technology and still design a portfolio that cannot survive the path to get there. If you want more analysis that separates the narrative from the mechanics, consider subscribing. The next story is already in the data.