The Chip Bet That Exposes AI Fund Risk Gaps

SAN FRANCISCO — Leopold Aschenbrenner’s Situational Awareness is still writing large AI checks. This week the embattled fund put roughly $400 million into Source Foundry, a Sequoia-backed Stanford-founded startup pitching faster, cheaper chip manufacturing — and, eventually, equipment that could challenge ASML’s lithography franchise. Combined with an earlier stake, the position now sits near $500 million. The money is real. So is the lesson beside it: AI thesis risk and hedge-fund risk are not the same spreadsheet.

Last month, Situational Awareness sold most of its public equity book to Ken Griffin’s Citadel after a violent unwind in AI infrastructure names. Assets that had swollen into the tens of billions compressed toward roughly $10 billion in some tallies. The fund kept private holdings, including Anthropic. Then it doubled down on Source Foundry. That sequence reads like resilience. It is better read as a forensic exhibit.

Conviction Survived. The Risk Model Did Not.

Aschenbrenner built Situational Awareness in 2024 as a former OpenAI researcher with no trading résumé and a blunt view that AI infrastructure would devour capital. Early returns reportedly rewarded concentration in semiconductors, cloud GPU landlords, and related hardware — often with heavy leverage — while shorting software names cast as disruption victims. In July, the longs cracked and the shorts rallied. The hedge failed in both directions. Prime brokers called for collateral. Forced sales followed. Crowding turned a drawdown into a funding spiral.

That is the gap. Risk systems that measure single-name volatility or neat long/short pairs miss correlated crowding and prime-broker liquidity as a single factor. When everyone owns the same AI hardware basket, the “hedge” against software becomes another crowded bet — not insurance. Culled has tracked the same prove-it mood when Wall Street shrugged at Nvidia’s best quarter and when the dot-com echo warned that narrative speed outruns physical and financial plumbing. Situational Awareness lived the plumbing failure.

Private Silicon Is Not a Risk Reset

Moving from leveraged public AI equities into private hard-asset silicon looks like maturity: less mark-to-market violence, more “real” capacity. Source Foundry sits in the manufacturing-tool layer that custom-silicon wars and hyperscaler capex both ultimately need. A $5 billion-class valuation on a stealth challenger in one of the most entrenched industries on earth is still a venture lottery ticket, not a completed risk model. Illiquidity hides volatility until the next capital call or secondary markdown. It does not erase it.

The utility question remains open too. If apps cannot justify the buildout, private chip tooling inherits the same demand cliff — only with longer lockups. Geopolitical and cyber overlays add non-market shocks that crowded AI books rarely stress-test cleanly, from Five Eyes warnings on AI attacks to China’s state-scale AI industrial bet.

The actionable principle is blunt. Treat thematic AI leverage as one factor, not a diversified stack. Price funding liquidity and hedge correlation before you price the next wafer story. Source Foundry may yet matter. The July unwind already does.