The numbers don't lie. They rarely do, when you can find them. A fund peaks above $45 billion in July, then contracts to roughly $10 billion by late summer โ a drawdown north of 78%. The portfolio manager, Leopold Aschenbrenner โ former OpenAI researcher turned architect of the Situational Awareness fund โ was forced to dump nearly every publicly traded equity position to Castle Securities at a discount. By September, he was back. Not with stocks. With options. Hundreds of millions of dollars' worth, concentrated across SK Hynix, SanDisk, AMD, Bloom Energy, CoreWeave, and a DRAM-themed ETF.
This is not a crypto story. But the mechanics โ the leverage, the narrative, the opacity, the second-act comeback โ are identical to what I have watched play out in DeFi since 2020. Let me explain why this matters to anyone trading AI-crypto crossover names, decentralized compute protocols, or simply trying to read the signal from the noise.
Context: From Thesis to Trauma to Re-Entry
Aschenbrenner built his reputation on a 165-page treatise called "Situational Awareness," published in mid-2024, predicting that Artificial General Intelligence would arrive by 2027 and that the United States needed industrial-policy-level focus to maintain its lead. The thesis attracted capital โ significant capital, from funds and family offices seeking concentrated AI exposure outside the Magnificent Seven wrapper.
By July 2025, the fund had reportedly swelled past $45 billion. Then came the unwind. AI infrastructure stocks โ the same names the fund had loaded up on โ sold off sharply. The leverage that amplified gains during the rally now amplified losses during the reversal. Aschenbrenner was forced to liquidate nearly his entire public-equity book to Castle Securities, a large broker-dealer known for absorbing distressed blocks. The fund's AUM reportedly shrank to roughly $10 billion. He retained his Anthropic private position โ illiquid, but theoretically high-upside.
Now the re-entry. According to CNBC's reporting, Aschenbrenner is buying options on AI infrastructure and memory: SK Hynix (HBM leader), SanDisk (NAND), AMD (the NVIDIA alternative), Bloom Energy (on-site power generation for data centers), CoreWeave (neocloud GPU rental), and the Roundhill Memory ETF. The thesis: AI capex bottlenecks are migrating from GPU scarcity to memory and power scarcity. Whether that thesis is sound, and whether options are the right instrument to express it, is the question.
Core: Stress-Testing the New Positions
I built my first real trading edge by reverse-engineering the 2020 Compound oracle exploit from raw mempool data. The pattern then was identical to what I see now: a high-conviction narrative, leveraged exposure, structural assumptions that may or may not hold, and a thin layer of public information covering a much thicker layer of private positioning. Let me break down what the trades actually imply, position by position.
The Memory Thesis Is Real, But Cyclical. HBM โ High Bandwidth Memory โ is the structural bottleneck beneath every AI accelerator. SK Hynix controls the dominant share of HBM3 and HBM3E supply. NVIDIA's H100 and H200, AMD's MI300 series, all require it. The bottleneck is genuine. But here is the mechanical issue: HBM is a product category inside a cyclical industry. DRAM pricing cycles have historically delivered 18-24 month upswings followed by brutal corrections. SK Hynix's stock has already priced in a substantial portion of the HBM scarcity premium. Options give asymmetric exposure โ limited downside to the premium, unlimited upside if the cycle extends. But the premium itself is now elevated. Implied volatility on memory names sits at multi-year highs. SanDisk's NAND exposure adds a second leg, but NAND cycles have been even more violent historically. The Roundhill Memory ETF basket spreads the bet across multiple suppliers, reducing single-name risk while maintaining thematic exposure.
Power Is the Silent Bottleneck. Bloom Energy's inclusion is the most structurally interesting leg. Data center power constraints in the U.S. โ particularly Northern Virginia, Phoenix, and Atlanta โ are now measured in gigawatts, not megawatts. Grid interconnection queues stretch 3-7 years. Bloom Energy's solid oxide fuel cells enable behind-the-meter generation, bypassing utility interconnect delays. This is a real thesis, validated by hyperscaler RFPs I have seen referenced in vendor filings. But the market cap of Bloom Energy versus the actual addressable data center power market implies the stock has already priced a significant share of the secular growth. Options on Bloom Energy reflect this: elevated implied vol, steep skew, and a premium that consumes real capital.
CoreWeave Is a Leveraged Play on GPU Rental. CoreWeave began as an Ethereum mining operation, pivoted to GPU cloud, and now runs one of the largest neocloud footprints in North America. Its contract structure with Microsoft and OpenAI has been disclosed in S-1 filings โ long-duration, take-or-pay, with pricing power tied to H100 spot rates. But the stock has roughly halved from its post-IPO peak, and short interest remains elevated. Aschenbrenner's option bet here is a volatility play: if AI capex resumes its upward trajectory, CoreWeave's earnings power revives; if it does not, the company faces refinancing risk on its GPU-backed debt. Anyone holding CoreWeave exposure through the 2025 unwind knows the path-dependency is brutal.
AMD as the Second Source. Including AMD in the basket signals diversification away from pure NVIDIA exposure. MI300X and MI325X have credible design wins โ Microsoft, Meta, Oracle โ but software ecosystem maturity (ROCm versus CUDA) remains the gating factor. Based on my audit experience reverse-engineering EigenLayer's restaking contracts in 2023, I learned that theoretical compatibility rarely survives contact with production workloads. Options on AMD here likely express a view that the second-source premium will widen as hyperscalers seek to avoid single-vendor dependency. Whether ROCm matures fast enough to make that bet pay is an open question.
The Mechanics of the Re-Entry. Here is where it gets dangerous. Options are not equity. A $200 million option premium can control $2-4 billion in notional exposure, depending on strike and expiry. If Aschenbrenner is running concentrated, long-dated call positions, his notional AI infrastructure bet may be larger now than it was in July โ funded by a fraction of the capital. This is leverage on leverage. Structure defines value; chaos destroys it. The structure here is built on implied volatility, not balance sheet. We saw the same pattern in DeFi in 2022. Three Arrows Capital ran concentrated bets on stETH, Luna, and GBTC. When the curve flattened, the leverage unwound in days. The current AI infrastructure trade carries the same risk profile: high narrative conviction, high concentration, leveraged exposure, and limited public visibility into fund-level mechanics.
The Opacity Problem. CNBC's reporting relies on anonymous sources familiar. No fund disclosure. No AUM confirmation. No position-by-position transparency. No detail on whether these are calls, puts, or spreads; no strike prices; no expiry dates. This is the same epistemic problem we face in DeFi: protocol teams claim TVL and yield figures that may or may not survive stress testing. I have spent years building local testnet environments to verify what smart contracts actually do versus what their documentation claims. The hedge fund equivalent โ a true position-level audit โ is impossible from the outside. Retail investors watching this narrative have no way to verify whether the trade is real, whether the notional is as large as implied, or whether the fund is selling premium rather than buying it. The signal is noise-filtered through marketing.
Contrarian: What the Crowd Misses
The retail reading of this story will be: "Smart money is back in AI, so AI is back." That is wrong on two counts. First, the smart money just demonstrated it is not, in fact, smart enough to avoid a 78% drawdown. The fund's July-to-September performance suggests risk management was inadequate for the leverage assumed. Calling a fund "battle-tested" after a near-total liquidation is a stretch. Second, the re-entry via options is a low-conviction expression compared to outright equity ownership. Options are conditional bets. They expire. They can be closed without disclosure. The signal is weaker than the headlines suggest.
What retail misses: the timing. Options expire. If the AI infrastructure thesis takes 6-12 months to play out โ which it likely will, given the lead times on HBM capacity expansion and data center commissioning โ short-dated options will decay before the thesis resolves. Aschenbrenner may be playing a volatility game, not a directional game. He may be selling premium, not buying it. The article does not say. Without disclosure, retail has no edge in interpreting the trade.
What smart money sees: capital structure fragility. CoreWeave's debt load. Bloom Energy's revenue concentration. AMD's software moat. SK Hynix's customer concentration on NVIDIA. These are not problems options solve. They are problems time and execution solve. The fund's re-entry, if anything, accelerates the timeline for these questions to be answered in price action.
Takeaway
We do not predict the future; we hedge against it. Aschenbrenner's comeback tells us one thing with certainty: someone with a track record, a thesis, and access to capital still believes the AI infrastructure trade has 2-3x of upside from here. Whether that belief is correct will be tested by the data โ HBM shipment volumes, DRAM pricing, data center power consumption, hyperscaler capex guidance โ over the next four quarters.
For crypto traders, the signal is simpler: AI-crypto crossover names โ decentralized compute protocols, AI agent tokens, RWA plays tied to GPU financing โ will move with this narrative. When the AI infrastructure trade accelerates, these tokens catch a tailwind. When it cracks, they fall twice as hard. Position accordingly. Stress-test your assumptions. And never confuse a comeback with a confirmation.
