Tracing the gas trail back to the genesis block—the 13F filing of the Situational Awareness Fund, submitted on August 14, 2026, reveals a portfolio that looks like a smart contract written by a brilliant but reckless architect. The holdings snapshot as of June 30 shows a staggering $20.24 billion in equities, with 55.5% crammed into two storage chip stocks—SanDisk and Micron. The rest is a distributed mess of AI cloud providers, fuel cells, and a cluster of Bitcoin miners pivoting to AI data centers. Then the market corrected. The fund collapsed. Citadel stepped in to take over the ‘problematic portfolio.’ The 13F is not a warning—it’s an autopsy. And the cause of death is written in plain sight: structural fragility masked as high conviction.
Context: The Man, the Thesis, the Leverage Leopold Aschenbrenner is not your typical hedge fund manager. Former OpenAI superalignment team member, author of the widely-read ‘Situational Awareness’ essay on AI geopolitics, turned macro investor. He launched this fund with a thesis that AI compute is the hard currency of the future—and the bottlenecks are physical: storage, power, data center real estate. The portfolio is a direct implementation of that worldview. As of June 30, the fund held: SanDisk ($5.674B, 28.0%), Micron ($5.574B, 27.5%), Bloom Energy ($1.902B, 9.4%), TSMC ADR ($1.254B, 6.2%), Nebius ($1.034B, 5.1%), CoreWeave ($0.954B, 4.7%), Core Scientific ($0.578B, 2.9%), Applied Digital, IREN, Riot Platforms, CleanSpark, and others. That’s a combined CR2 of 55.5% and CR7 of 84.3%. For context, a typical institutional fund’s top-10 holdings rarely exceed 30-40% of assets. This is a 2-3x concentration multiplier.
By July 2026, AI-related stocks experienced a sharp drawdown. The fund was forced to sell most of its public equity holdings, with market reports citing ‘leverage pressure.’ Citadel, the prime broker, took over the troubled portfolio. The 13F, filed a month later, is a snapshot of a portfolio that no longer exists. But the structural errors are frozen in time.
Core: The Invariant Breakdown Let’s analyze this portfolio as a protocol. The core invariant is a leveraged bet on the physical bottlenecks of AI compute. The code is written in capital allocation: high conviction, narrow diversification, no hedging. The invariant assumes that the demand for AI compute continues to grow exponentially, and that storage, power, and data center capacity are the binding constraints. That’s a reasonable hypothesis—but as any DeFi auditor knows, a protocol’s security depends on the worst-case assumptions, not the best-case.
Entropy increases, but the invariant holds. Except when it doesn’t. Let’s trace the entropy.
Storage Layer: The 55% Syndrome SanDisk and Micron are both NAND flash and HBM memory suppliers. Micron is one of three HBM suppliers (alongside SK Hynix and Samsung) and is heavily exposed to AI server memory. The logic is sound: AI training and inference require high-bandwidth memory, and HBM has been supply-constrained. But storage is a cyclical industry with a history of boom-bust cycles. The thesis banks on the current tightness persisting indefinitely. That’s a fragile invariant. Capacity expansions take 18-24 months. By 2027, HBM supply could outpace demand. The fund’s storage concentration is a leveraged exposure to a single cyclical narrative.
Power Layer: The Bloom Energy Bet Bloom Energy makes solid oxide fuel cells. The thesis: AI data centers need reliable, clean power, and Bloom’s technology can provide on-site generation. But Bloom Energy is a pre-profit growth company with a history of cash burn. Its stock is volatile. The fund allocated 9.4% to a single energy company—again, a high-conviction, high-risk bet. No diversification into other power sources (natural gas, nuclear, renewable credits). That’s a single point of failure.
Cloud Layer: CoreWeave and Nebius CoreWeave is a GPU cloud provider that has signed massive contracts with OpenAI. Nebius is a European AI cloud. These are direct plays on GPU-as-a-service. The logic: as AI models scale, they need more GPU compute, and cloud providers are the gatekeepers. But these are also capital-intensive businesses with thin margins and high leverage to Nvidia GPU supply. If Nvidia’s supply chain improves, the scarcity premium evaporates. The fund held both—a 9.8% combined allocation—but these are correlated assets, not hedges.
Miner Layer: The AI Transition Bet The fund held approximately 7% in Bitcoin miners: Core Scientific, Applied Digital, IREN, Riot, CleanSpark. The rationale: these miners own large power capacity, data center infrastructure, and are pivoting to AI hosting. Core Scientific has signed AI hosting contracts. IREN has been expanding its data center footprint. This is a clever proxy for ‘AI power and real estate.’ But it introduces a second correlated risk: Bitcoin price. If Bitcoin drops, these miners’ core business suffers, even if AI hosting grows. The fund did not hedge Bitcoin exposure. The miner stocks are also illiquid—when the fund needed to sell, they would have been the hardest to exit. In a forced liquidation, these positions likely suffered the worst slippage.
Missing Layer: AI Applications The portfolio has zero exposure to AI application layer companies—no OpenAI, no Anthropic (though they are private), no software companies leveraging AI. This means the fund is betting on the infrastructure without any hedge on the demand side. If AI model companies reduce capex, the entire infrastructure chain suffers. This is a classic ‘picks and shovels’ bet, but the picks and shovels are all in the same basket.
Leverage: The Unseen Reentrancy The 13F does not disclose leverage ratios, derivative positions, or prime broker agreements. But the fact that Citadel had to take over the portfolio suggests that the fund was using structured financing—likely total return swaps or margin loans. Given the high concentration, the loan-to-value ratio would have been extremely sensitive to price drops. The July drawdown in AI stocks (likely 15-20% for the portfolio) would have triggered margin calls across multiple positions. The miners, being illiquid, would have been the first to be liquidated, causing a cascade. This is a classic reentrancy attack: the margin call on one asset triggers a forced sale of another, which depresses prices further, leading to more margin calls.
In the absence of trust, verify everything twice. The fund’s thesis was built on trust in the long-term AI narrative. But the invariants of risk management were not verified. The portfolio lacked a break-glass mechanism: no stop-loss, no hedging, no diversification. The leverage was the hidden reentrancy vulnerability.
Contrarian: The Blind Spots The conventional critique of this fund is that it was too concentrated. That’s obvious. The contrarian angle is that the fund’s very thesis—the ‘situational awareness’ of AI infrastructure—created a blind spot for its own risk. Aschenbrenner’s worldview is that compute is the decisive factor in AI development. That worldview correctly identifies the bottlenecks but incorrectly assumes that the bottlenecks are stable and that the market will reward them proportionally. The market is a second-order system. When everyone piles into the same ‘bottleneck’ trade, the trade itself becomes the bottleneck. The fund’s high conviction led to overconfidence in the persistence of the scarcity premium.
Another blind spot: the miners as AI transition plays. The market has priced these miners as AI proxies, not as crypto proxies. That means their valuation is driven by AI narrative, not Bitcoin fundamentals. If the AI narrative cools, these stocks could lose 50-70% of their value, as they lose both the AI premium and the crypto floor. The fund did not account for this double-loss scenario.
Furthermore, the fund’s lack of diversification across the AI stack (hardware, software, services) meant that it was effectively long a single factor: AI infrastructure capex. That factor is correlated with interest rates, regulation, and geopolitical tensions. The fund had no hedge for rate hikes, export controls, or an AI winter. The portfolio was a high-beta, high-conviction, no-alpha bet.
Code is law until the reentrancy attack. In traditional finance, the ‘code’ is the investment mandate. The fund’s mandate was ‘bet on AI infrastructure bottlenecks.’ That code looked solid until the liquidity crisis triggered a reentrancy spiral. The fund’s collapse is a textbook case of invariant failure: the invariant that ‘AI capex will grow forever’ was stress-tested and found lacking.
Takeaway: The Vulnerability Forecast The Situational Awareness Fund’s collapse is not an isolated incident. It is a warning for all concentrated, leveraged bets on AI infrastructure. The market is now aware that the ‘AI trade’ can deleverage violently. The next time a fund with similar concentration hits a drawdown, the prime brokers will be less forgiving. The leverage in the AI infrastructure ecosystem is opaque—13F filings don’t show it. But the risk is real.
Entropy increases, but the invariant holds. The invariant here is that high conviction must be paired with high risk management. The fund failed to do that. The lesson for DeFi and traditional finance alike: trust no single thesis, verify every risk parameter. The 13F of the Situational Awareness Fund is a snapshot of a failure. The next one might not even get a filing.