$450 Billion Hedge Fund Liquidated Following AI Stock Collapse

South Korea's KOSPI index triggered circuit breakers consecutively this week. At its lowest point, the index erased nearly 40% of its gains from the historical high of 9,380 set over a month ago. The Seoul market lost $2.18 trillion in market capitalization.

On May 27, nine weeks ago, South Korea launched single-stock leveraged ETFs. These funds promised to deliver twice the daily volatility of individual company stocks, prompting retail investors to pour into the two products tracking Samsung and SK Hynix.

On July 29, both stocks suffered heavy losses: SK Hynix plunged as much as 20% intraday and closed down 9.6%; Samsung fell as much as 14% and closed down 5.2%. Since these two companies account for over half of the KOSPI index’s weight, their sharp declines dragged the entire index significantly lower. To maintain double the intraday exposure, these funds typically increase positions after rallies and reduce them after declines, generating further volatility and cascading risks.

In addition to this month’s comprehensive reassessment of AI trading, China’s largest DRAM producer, CXMT, surged 466% on its first day of listing in Shanghai on July 27, becoming the most valuable publicly listed company on the Chinese mainland. Although CXMT still lags behind Korean manufacturers in scale and advanced memory technology, this listing has sparked concerns that rising Chinese production capacity could depress traditional memory prices. Looking further out, the KOSPI has risen over 50% year-to-date, making it the best-performing major market.

On July 31, South Korea’s Minister of Economy and Finance, Ku Jun-chul, announced that South Korea plans to launch a new sovereign wealth fund next year dedicated to long-term investments in strategic domestic industries such as AI, semiconductors, robotics, defense, and biotechnology. Seoul will contribute 20 trillion won (approximately $14 billion) to the fund, with 16 trillion won coming from its holdings in state-owned financial institutions and 4 trillion won from inheritance and gift taxes.

On July 27, NVIDIA and dozens of partners launched the Open Secure AI Alliance; six days earlier, OpenAI disclosed an internal network evaluation in which an agent escaped and infiltrated Hugging Face.

Closed AI tools hindered parts of Hugging Face’s forensic work, while the self-hosted open-weight model GLM 5.2 reviewed over 17,000 operations, helping to contain the breach. Open weights make the trained model parameters publicly available, allowing organizations to run and inspect them on their own infrastructure. Hugging Face joined this coalition, while leading closed labs such as OpenAI, Anthropic, and Google did not.

During the same period, three other documents presented differing positions. NVIDIA’s letter dated July 24, “Open Weights and U.S. AI Leadership,” initially had 25 signatories and quickly grew to over 230. Anthropic remained absent and responded on July 27, stating that it opposes a complete ban on open weights but supports controls on advanced chips, restrictions on replicating industrial-scale models, and mandatory safety testing for sufficiently powerful models.

On July 28, over 1,100 industry professionals signed a letter titled “Slow the Frontier,” calling on Washington to establish an international mechanism to slow the development of autonomous AI. Dario Amodei, CEO of Anthropic, and Jakub Pachocki, Chief Scientist at OpenAI, signed the statement, and both companies endorsed it.

These documents focus on how to govern and deploy powerful models, with recent emphasis on open-weight models. Starting in August, an executive order issued on June 2 directed the Department of the Treasury, the Department of Defense (through the NSA), and the Department of Homeland Security (through CISA) to develop classified benchmarking procedures and a voluntary pre-release framework for covered frontier models.

In late 2024, Leopold Aschenbrenner launched the Situational Awareness Fund with hundreds of millions of dollars in seed funding, naming it after his articles on artificial general intelligence as a concentrated long position on AI stocks.

The fund achieved a post-fee return of 439% in the first half of 2026, with peak reported assets under management reaching $45 billion. According to reports, the fund’s total exposure peaked at four times its capital. At this level, every $100 of investor equity supported $400 in positions. A 10% adverse move in the portfolio would erase approximately $40, or 40% of the fund’s capital. As losses erode this buffer, prime brokers may require additional collateral or force liquidation.

Following the sale of AI stocks this month, on July 30, Leopold sold the majority of its publicly traded stock portfolio to Citadel. Situational Awareness still holds approximately $10 billion in assets, with its largest holdings remaining private companies such as Anthropic.

Author: Chamath Palihapitiya; Translated by Shenchao TechFlow