Situational Awareness Lost 67% in July. Anthropic Is Headed for a Screen Price.

Former OpenAI researcher Leopold Aschenbrenner's Situational Awareness fund lost 67% in July and sold most of its public-equities portfolio to Citadel after leveraged AI trades triggered margin calls.

The fund was launched in 2024 by Aschenbrenner after OpenAI fired him over an alleged information leak - a dismissal Aschenbrenner has linked to an internal memo criticizing the company's security practices. He had no professional investment experience, but the fund returned more than 1,000% since inception, according to the Wall Street Journal, and reached its peak of $45 billion at the start of July 2026, according to CNBC. Thirty days later it was down to about $10 billion.

The fund was long AI infrastructure and short software, levered as high as 400%. In July, the semiconductor and AI-infrastructure complex sold off sharply. Sandisk fell nearly 47% in July, Micron dropped almost 29%, and several of the fund's other disclosed holdings suffered similarly severe drawdowns. Short positions in software names including Adobe moved against it at the same time.

Goldman Sachs, JPMorgan Chase, and Bank of America, the fund's three prime brokers, had been working with it to meet margin requirements or reduce positions in an orderly fashion. By the week of July 27, they were already marketing both sides of the book (longs and shorts), and attempts to appeal to investors for further capital were unsuccessful. On July 29, the FOMC held rates in a 9 to 3 vote, and several of the fund's core holdings posted intraday yearly lows. Before the next morning's open, most of the fund's public book, longs and shorts together, was sold to Citadel in a single negotiated block. The price wasn't disclosed, but there were other potential bidders - including Millennium and Jane Street.

Bank of America's Brian Moynihan addressed it publicly on August 5. "These are all warning shots," he told CNBC. "Valuations get out, leverage in the system gets there. You have to be careful." It is the only on-the-record comment from any party to the trade.

The private book was not sold - Situational Awareness kept its stakes in private AI companies, including a position in Anthropic. Aschenbrenner's decision to lever the public fund by up to 4x essentially created a ticking time bomb during selloffs. At roughly 4x gross exposure, a 30% drawdown in the long book works out to something like a 120% hit to equity before any offset from hedges. In July the shorts added to the losses rather than cushioning them. Public stocks get reevaluated every day at close, and when the market falls consistently, the lender asks for collateral.

Private assets, by comparison, have no daily screen price. Those investments are harder to call, because market turmoil is hidden behind infrequent pricing.

This absence of a daily mark is normally considered a disadvantage - certainly for sellers that's the case. Illiquidity discounts are real and a stale mark is a bad mark. But it also provides insulation. A private company's valuation is set by its last round or last transaction, not by public sentiment's volatile reaction to intra-day market news. Private assets can still be financed, marked down or used as collateral, but they are not continuously repriced on an exchange. That generally insulates them from the daily margin-call mechanics that can force the liquidation of listed securities.

One caveat to all of this: every account of the sale is sourced to people familiar with it. Neither Citadel nor Situational Awareness has commented on the record.

One of the assets that made up Situational Awareness's private market portfolio - Anthropic - is heading for a screen price.

Anthropic confidentially submitted a draft Form S-1 to the SEC on June 1, 2026. No public S-1 has appeared on EDGAR, and the company has not announced a date, a price range, a ticker, an exchange, or underwriters. Its last private mark was a $65 billion Series H at a $965 billion post-money valuation in May. Secondary-market trading has implied something closer to $1.2 trillion.

October is the consensus expectation for listing among reporters and prediction markets - but it is by no means publicly scheduled. Prediction markets put completion by year-end somewhere in the 68 to 77 percent range.

What's interesting here is that we've seen similar experiments run already in 2026. SpaceX listed on June 12 and raised more than $85 billion. The stock reached a post-IPO intra-day high of around $225, but has steadily reduced in value since that point and now sits at $110, fairly close to its all-time low.

Signs of doubt are starting to show in the market. OpenAI, which filed confidentially about a week after Anthropic, is reported to be aiming for 2027 rather than the initially expected 2026 IPO, following SpaceX's debut. Bloomberg has reported Anthropic still appears on track for a potential October listing, with executives beginning investor meetings.

The same property that broke Aschenbrenner's public book, a price that resets every day whether you agree with it or not, is what Anthropic is ostensibly about to acquire.

A private valuation is a negotiated headline. In late-stage private financings generally, investors may accept a higher headline valuation in exchange for liquidation preferences, downside protection, or other structural rights. Without the underlying terms, the quoted post-money figure does not necessarily describe the economics of every share equally. And they stick - Anthropic's secondary market has implied something nearer $1.2 trillion, but $965 billion is still the number everyone cites. Private marks do not move when someone disagrees with them, but rather when a new round prices or a holder decides to write one down.

A public valuation is the marginal trade. Thousands of decisions instead of a handful, resolved continuously, and nobody has to agree to the number for it to become the number. All you need is a motivated buyer and a motivated seller, with much less friction in the way of getting a deal done.

SpaceX shows what that costs and who pays it. The company priced at $135 and raised what it set out to raise. Four days later the stock touched $225.64. By July 28 it was at $107.01. The offering worked exactly as designed, but the offering valuation proved provisional.

Deal teams live inside negotiated headline numbers every day, in earnouts, in stock-for-stock structures, in equity comp for target employees. As more AI-era private companies edge toward public listings, the gap between a company's last private mark and what the market actually pays for it becomes a live underwriting question, not an academic one. Know which number you're actually working with, and what it would take for the market to disagree with it.