← Back to feed

Leopold Aschenbrenner’s hedge fund, Situational Awareness, reported a 67% loss in July after gaining 439% in the first half of 2026. The fund’s net position is still up roughly 80% on the year, but only because the H1 gain was large enough to absorb the July collapse. The math reconciles only if you were at +439% entering the month.

Aschenbrenner graduated as Columbia’s valedictorian and spent time on OpenAI’s superalignment team before writing a widely-cited essay on AI timelines and building a fund around the physical infrastructure thesis: the companies that will supply the machinery behind AI — SK Hynix, CoreWeave, Nebius, Micron, Bloom Energy — would compound as AI scaling continued. That bet worked for the first six months.

What Broke in July

The fund was running approximately four-to-one leverage, borrowing three to four dollars for every dollar of investor capital. At that ratio, a 25% decline in the underlying positions eliminates the cushion. In July, the holdings fell 35-47%.

The directional problem was structural. The fund was long AI infrastructure and short software broadly. That sounds like a hedge, but both sides required the same world-view to hold: AI capital expenditure continues outpacing software revenue. When the market questioned that narrative in July, both legs moved against the position simultaneously.

The leverage meant collateral problems. Banks including Goldman Sachs, JPMorgan, and Bank of America made margin calls as positions declined. Selling to meet those calls pushed prices lower, which triggered more calls. Aschenbrenner’s letter to investors describes it as a bank-run dynamic — known vulnerability attracting more pressure, not a single shock.

An emergency fundraise failed. Millennium and Jane Street examined the portfolio and declined. Citadel’s Ken Griffin acquired the entire distressed book in a single transaction, deliberately structured to prevent rival funds from picking off individual positions. The discount was above 10% on roughly $16 billion.

The One Position That Could Not Be Seized

One holding survived: Aschenbrenner’s stake in Anthropic. Private company equity has no daily mark, so it cannot be posted as collateral, and it cannot be seized when collateral is insufficient. The asset he was unable to sell is the one the banks could not take.

The trade resolution itself was briefly bullish. Removing a forced seller from the market pushed US tech stocks to their best single-day performance in four months. Korean chip names rose 18%. He sold at the bottom.

What the Episode Reveals About AI Infrastructure Investing

The thesis — AI infrastructure spending creates durable value in the companies that supply it — has not been disproved by this episode. The specific hardware names Aschenbrenner held have generated real revenue growth through 2026. What the episode demonstrates is the distance between a directional thesis being correct and a leveraged fund surviving long enough to collect.

The stocks required a stable market view for an extended period. Four-to-one leverage on volatile semiconductor and data-center names does not give a fund that stability. If other market participants know the position and the leverage, they also know the exact conditions that force liquidation. That information becomes actionable.

Key Numbers

• H1 2026 return: +439%
• July 2026 return: -67%
• YTD net: approximately +80%
• Leverage ratio: ~4x (3-4 dollars borrowed per dollar of equity)
• Portfolio decline in July: 35-47% across holdings
• Citadel acquisition price: above 10% discount on ~$16B
• Surviving position: Anthropic equity (private, not collateralized)