This Was Situational Awareness's Top 5 Holdings at the End of Q2. They're All on Sale Now

The article revisits Leopold Aschenbrenner’s Situational Awareness, which “blew up” in late July after leveraged AI bets triggered margin calls when AI stocks tumbled, with the fund selling more than $10B of its AI portfolio to Citadel. It highlights key AI holdings from the end of Q2—Micron (+600% YoY) and Sandisk (+2,700% YoY) tied to memory tightness, plus Bloom Energy with revenue up 166% to $1.07B and GAAP operating income of $182.2M—while flagging higher-risk neocloud names (Nebius, CoreWeave) running triple-digit growth but massive losses and billions in debt. Overall, it frames AI as still investable, but warns leverage and volatile exposure can be toxic.

Analysis

The real signal here is not “AI is broken,” it’s that the factor exposure inside AI is splitting between asset-light bottlenecks and balance-sheet-intensive compute landlords. Memory and advanced packaging names should keep catching flows because they monetise every incremental watt of hyperscaler capex with minimal financing risk; the market is still underpricing how quickly HBM scarcity can translate into gross-margin durability for MU and, secondarily, TSM. The unwind also clears a visible supply overhang: forced sellers in crowded AI names can temporarily depress multiples, but they do not fix the structural shortage in memory or foundry capacity.

The weak link is leverage. CRWV and NBIS are exposed to a regime where higher funding costs, slower customer prepayments, or even a small pause in AI capex can compress EV/revenue multiples faster than operating losses improve; that risk is not linear because debt maturity and covenant optics matter more than headline growth. BE is a different story: useful as a “picks-and-shovels” energy proxy, but the stock only works if contracted backlog converts into bankable cash flow, not just strategic announcements.

Contrarian read: the market may be over-penalizing the AI complex as a monolith after one leverage blow-up. If NVDA demand remains strong and hyperscaler capex commentary stays firm, the better trade is to own the supply bottlenecks and fade the weakest funding models, not short the entire theme. Falsifiers are clear: a sequential deceleration in hyperscaler capex, a sharp widening in high-yield spreads, or a memory ASP rollover would invalidate the long MU/TSM thesis within 1-3 months.

Key Decisions for Investors

• Long MU vs short CRWV into the next AI-capex/earnings cycle: pair expresses the view that scarcity economics outrun leveraged compute economics; target 15-25% relative outperformance over 1-3 months, stop if CRWV can refinance cheaply or MU guides down memory pricing.
• Add TSM on weakness as the highest-quality ballast in AI infrastructure; 6-18 month hold with lower drawdown risk than the neoclouds, and a clean falsifier is any evidence of foundry share loss or unexpected capex slowdown.
• Buy SNDK only tactically on pullbacks, not chasing strength: tighter supply can keep the stock bid over 1-3 months, but it is the most cyclical of the scarcity trades and should be trimmed if spot memory pricing stops rising.
• Avoid or hedge NBIS/CRWV ahead of financing windows; if you need exposure, use put spreads rather than outright shorts because sentiment can stay extreme, but the downside accelerates if credit markets tighten.
• Treat BE as a watchlist name, not a conviction long, until contract backlog converts into visible free cash flow; otherwise the stock remains narrative-driven and vulnerable to multiple compression.