The SEC filing arrived on August 14. It documented holdings as of June 30. By then, the fund was already dead.
Two weeks earlier, market reports had confirmed what the filing would eventually prove: AI-linked equities had fallen, leverage was pressing, and Citadel was taking over a “problem portfolio.” The 13F is an autopsy, not a warning. It is a transparency mechanism that arrived 45 days too late to matter.
The numbers are stark. SanDisk position: $5.674 billion. Micron: $5.574 billion. Combined, 55.5% of a $20.24 billion portfolio. Top seven positions: approximately 84.3%. A typical institutional fund runs a CR10 of 20–40%. This portfolio was two to three times more concentrated, wrapped in undisclosed leverage, and stapled to small-cap Bitcoin miners with terminal liquidity risk.
We build the rails, then watch the trains derail.
That phrase has always applied to protocols. It applies equally to portfolios positioned as infrastructure. The only difference is the failure mode—and the forensic trail left behind.
Context: A Worldview, Quantified
The fund is Situational Awareness LP, founded by Leopold Aschenbrenner. Former OpenAI superalignment researcher. Author of the widely circulated essay that made “compute is the currency of the AI era” a mainstream institutional talking point. The 13F is the quantifiable expression of that worldview, filed under penalty of perjury with the SEC.
This portfolio is not a tech allocation. It is a vertical infrastructure spine. Storage layer: SanDisk and Micron. Foundry layer: TSMC ADR at 6.2%. Cloud and GPU layer: CoreWeave and Nebius at 9.8%. Power layer: Bloom Energy at 9.4%. And a miner-conversion layer—Core Scientific, Applied Digital, IREN, Riot Platforms, CleanSpark—at roughly 7% combined.
The logic is coherent. HBM supply is constrained; SanDisk and Micron are primary memory vendors to every AI server builder. Data center power is a genuine bottleneck; Bloom Energy sells the fuel cells that keep racks alive. CoreWeave and Nebius are the physical layer of GPU rental. Bitcoin miners, rebranded as AI data center hosts, offer optionality on power contracts and facility footprints.
But this is a case study in timescale error. The fund bet on a narrative that was directionally correct yet temporally fragile—and it levered that fragility.
Core: The Forensic Dissection
Start with the concentration math.
A CR2 of 55.5% means the fund’s survival depended on two cyclical semiconductor names. Memory chips are violently cyclical. In an upcycle, this structure prints alpha. In a downcycle, it compounds losses. The portfolio held no software layer, no application layer, no hedge, no uncorrelated anchor. Every position was a function of the same underlying variable: AI capital expenditure growth.
That is correlation by construction. Storage, power, and compute clouds do not diversify each other when the shared driver is a single CapEx wave. They move in lockstep. Across my years auditing liquidation engines in DeFi lending protocols, I have watched this failure mode repeat: protocols that calibrate safety margins on an upward-only oracle discover, exactly once, how the downside correlation matrix behaves.
Second, the leverage structure. The 13F discloses equity positions. It does not disclose financing. It does not disclose short positions. It does not disclose transactions after June 30. The $20.24 billion figure is the visible tip of a risk iceberg.
Citadel’s language matters. It took over a “problem portfolio.” That is not a typical margin call. That phrasing suggests a structured position—a total return swap, a margin loan package, or an options book—being unwound through counterparty agreement rather than exchange liquidation. If a TRS was involved, effective leverage could be substantially higher than the two-to-three-times that public reporting implies. The July implosion was not a market event. It was a contract event.
Third, the miner thesis. Including Bitcoin miners inside an “AI infrastructure” portfolio is intellectually clever and structurally dangerous. The AI-transition narrative treats miners as power-and-facility holders, converting proof-of-work revenue into compute-hosting rental income. But these entities retain Bitcoin price exposure, equity dilution risk, and counterparty concentration—Core Scientific’s and CoreWeave’s GPU lease obligations being the clearest examples.
When the AI narrative retreats, miners lose both valuation pillars at once. The downside is a double-decrement. And in a forced liquidation, the small-cap miners are the tail that suffers maximum price impact. They are the subordinated tranche of this structured product. I have seen this exact dynamic in DeFi’s leveraged farming collapses: the illiquid tail absorbs the largest percentage loss precisely because it cannot exit without moving the market against itself.
The market spillover deserves attention too. When a $20 billion fund deleverages in weeks, the flow itself becomes news. The July drawdown in AI infrastructure equities was not purely fundamental. It was partly this fund’s forced selling, amplified by other levered funds running similar “AI bottleneck” strategies. The 13F thus documents a coordinated-chain reaction in hindsight.
What is the transferable lesson? High conviction does not substitute for liquidity planning. I have seen “code is law” protocols die the same way: they optimize the bull case, ignore the down-correlation matrix, and discover too late that their safety parameter was calibrated on a bull-market oracle.
Contrarian: The Filing Is a Solvency Illusion
Here is the counter-intuitive angle: the 13F makes the fund look more solvent than it was.
The June 30 valuation of $20.24 billion tells us almost nothing about liquidation value in late July. After the drawdown, after forced selling, after leverage costs, the realized losses are likely far larger than public data will ever permit reconciliation. The filing is a snapshot of a patient before the cardiac arrest.
The second counter-intuitive point: the underlying thesis was correct—on the wrong timescale. Physical bottlenecks appear real during a CapEx surge. They evaporate when capacity catches up. The historical record in semiconductors says shortages resolve in 18 to 36 months. When they do, the binding-constraint narrative unwinds, and both earnings and multiples compress simultaneously. The market treated “hardware shortage” as permanent. It never is.
And the deepest irony: code is law, until the oracle lies. Here, the oracle was market consensus on AI CapEx growth. When that oracle repriced, the fund’s internal logic—concentrate on the bottleneck, lever the conviction—executed exactly as written. Self-destruction was not a bug. It was the specified behavior under the specified condition.
Takeaway: The Rail Gauge Was the Risk
Watch Citadel’s subsequent 13F filings. That is the liquidation tracker. How quickly the problem portfolio unwinds reveals both the realized leverage ratio and the true scale of counterparty losses. If large block trades appear in miner names, assume the damage is still propagating.
The structural lesson for DeFi and TradFi is identical: leverage plus concentration plus an illiquid tail is a kill switch. The rails were correctly built. The train derailed because the engineer forgot that gravity—like leverage—is always on. In a bear market, survival is the only benchmark that matters. This fund did not survive. Its filing is the headstone.