The numbers do not lie. A fund that managed $45 billion in July 2024 has been reduced to approximately $10 billion. The mechanism was predictable: high leverage on concentrated AI-related positions, followed by a margin call cascade that forced the liquidation of nearly all publicly traded holdings at distressed prices to a single counterparty called Castle Securities. Now, according to sources familiar with the matter, the same fund manager is back in the market, purchasing options worth hundreds of millions of dollars on AI infrastructure and storage-related names. The tickers include SK Hynix, SanDisk, AMD, Bloom Energy, CoreWeave, and the Roundhill Memory ETF (DRAM).
This is not a comeback story. This is a stress test of whether redemption arcs exist in quantitative finance, or whether the market is simply watching a sophisticated entity repeat the same structural errors under different market conditions.
I have spent twenty years dissecting financial engineering in digital asset markets. I have seen leverage hide inside structures that looked conservative on paper. I have watched funds describe their risk management as "robust" in investor letters while their actual portfolios resembled a casino floor with the house edge removed. The Situational Awareness fund, managed by a former OpenAI researcher, presents a particularly instructive case study in how narrative authority can substitute for operational discipline, and how the halo of technological credibility can mask fundamental portfolio construction failures.
The purpose of this article is not to celebrate or condemn. It is to dissect.
The AI Infrastructure Thesis: What the Fund Got Right (And What It Did Not)
Before examining the fund's structural vulnerabilities, it is worth acknowledging what the underlying thesis has merit. The AI infrastructure buildout has created genuine physical bottlenecks that extend well beyond GPU availability. High-bandwidth memory (HBM), which is critical for training large language models, has seen demand outstrip supply since 2023. SK Hynix, the dominant HBM supplier, has been operating at or near capacity while Samsung and Micron attempt to close the gap. Enterprise SSD and DDR5 demand have been pulled upward by AI server deployment, creating a secondary demand surge that traditional PC and mobile markets cannot generate alone.
Theēµå bottleneck is equally real. Data centers require reliable, high-capacity power feeds. The interconnection queues for gridę„å „ in major markets have lengthened considerably, with some regions reporting waiting periods measured in years rather than months. Bloom Energy, which provides on-site power generation solutions, has positioned itself as a beneficiary of this constraint. The logic is straightforward: if a hyperscaler cannot connect to the grid quickly enough, on-site generation becomes the only viable path to capacity expansion.
CoreWeave represents a different but related play. As a GPU cloud provider, it rents compute capacity to AI developers who cannot afford or do not want to build their own infrastructure. During periods of GPU shortage, GPU cloudē§čµ rates increase, benefiting providers with existing inventory. CoreWeave has been aggressive in acquiring H100 and H200 clusters, positioning itself as a direct beneficiary of the training compute arms race.
These are coherent investment theses. They have fundamentals behind them. A fund that holds positions in these names during a period of sustained AI capital expenditure growth would be making a defensible directional bet. The problem is not the thesis. The problem is the vehicle and the leverage.
The Structural Problem: What $45 Billion in AUM Actually Hides
Let me be precise about what the reported figures mean. When a fund claims $45 billion in assets under management at peak, several questions immediately follow that the media coverage does not answer. What is the composition of those assets? How much is proprietary capital versus investor capital? What is the leverage multiplier applied to the underlying positions? What are the redemption terms for the investor base?
The fact that the fund was forced to liquidate "nearly all" publicly traded positions at a discount to a single counterparty tells me something specific: the fund had a liquidity mismatch problem that was not disclosed or was not adequately managed. When a fund must sell assets quickly to meet margin calls, it does not have the luxury of price discovery. Castle Securities presumably acquired those positions at a discount that reflected the distress, not the fundamental value. This is how wealth transfers from the fund's investors to the counterparty that has the liquidity to absorb the sale.
The retention of Anthropic private equity positions is telling. Private equity positions are illiquid. They cannot be easily sold to meet margin calls. In a liquidation scenario, the fund would naturally hold onto illiquid assets while selling liquid ones first. This creates a perverse dynamic: the most speculative, hardest-to-value assets remain on the balance sheet while the liquid positions are haircut and sold. From a due diligence perspective, this is a red flag. The fund's stated strategy may have been "concentrated, high-conviction positions," but the actual behavior during stress suggests inadequate liquidity stress testing.
The new options positions add another layer of complexity. Options are not direct equity purchases. They represent leveraged exposure to underlying assets without requiring the full capital outlay of a spot position. When a fund is already operating with high leverage, adding option positions can amplify both gains and losses in non-linear fashion. The reported "hundreds of millions" in option premiums suggests that the fund is still operating with significant leverage, just in derivative form rather than direct borrowing.
I want to be specific about what I mean when I say this is a problem. In my experience auditing smart contract systems and DeFi protocols, I have learned to distinguish between leverage that is transparent and leverage that is hidden. Hidden leverage kills funds. Transparent leverage can be managed. The Situational Awareness fund appears to have had leverage that was either undisclosed or inadequately stress-tested under adverse conditions. The result was a 78% reduction in AUM within months.
The Contradiction at the Heart of the Comeback
Here is the core tension that the bullish narrative ignores: a fund that just experienced a catastrophic liquidity event is making concentrated leveraged bets on the same category of assets that caused the original collapse. The AI infrastructure names that the fund held during the summer of 2024 presumably contributed to the drawdown. Whether through direct positions in GPU cloud providers, memory stocks, or power infrastructure plays, the same macro conditions that are now being bet on again were present during the crash.
The difference, according to the narrative, is that the summer decline was an overreaction. The AI capex cycle is intact. The physical bottlenecks are real. The positions were sold in a panic, not because the thesis was wrong, but because the market was risk-off and leverage amplified the selling.
This argument has merit. Market volatility can separate price from fundamental value. A fund with a long time horizon could legitimately view the summer selloff as an opportunity. But there is a crucial difference between a long-term investor accumulating during a dip and a highly leveraged fund re-establishing concentrated positions after a margin call cascade. The former has dry powder and patience. The latter has a damaged balance sheet, potentially skittish investors, and a psychological incentive to recover losses quickly.
The unknown that concerns me most is the investor base composition. When a fund loses 78% of its AUM in a short period, the investor base typically fragments. Some investors redeem. Some stay and wait for recovery. Some negotiate special terms. The fund that emerges from this process is not the same fund that entered it. The capital base may be smaller, more concentrated, and more tolerant of risk. Or it may be more conservative, having learned painful lessons about leverage. Without regulatory filings or investor disclosures, there is no way to know which version of the fund is making the new option purchases.
The reference to the manager being a former OpenAI researcher is doing significant narrative work in the coverage. It is positioning technical credibility as a reason to trust the investment thesis. This is a logical fallacy that I encounter frequently in digital asset markets. Being a technical expert in one domain does not confer expertise in portfolio construction, risk management, or market timing. The same person who understands transformer architectures and GPU memory hierarchies may not have the same depth of experience in managing leverage during a liquidity crisis. These are different skill sets. The market's tendency to conflate them has contributed to several high-profile failures that I have analyzed over the years.
What the Market Is Actually Pricing
Let me address the storage narrative specifically, because it represents the most interesting element of the new positions. The DRAM ETF (ticker DRAM) tracks an index of memory and storage companies. SK Hynix and SanDisk are individual positions that fit within this theme. The thesis is that memory is transitioning from a cyclical commodity to a structural growth asset driven by AI demand. This thesis is not wrong, but its execution is more complicated than the narrative suggests.
HBM production requires specialized manufacturing processes that SK Hynix has mastered ahead of competitors. This technical lead has translated into pricing power and margin expansion. However, Samsung and Micron are investing aggressively to close this gap. By 2025 and 2026, new HBM capacity will come online from multiple suppliers. The supply-demand dynamic that currently favors producers will normalize, potentially compressing margins. A fund that buys HBM exposure via options is betting on timing as much as direction. The direction of AI memory demand is likely positive. The timing of when that demand translates into sustained margin expansion rather than just revenue growth is uncertain.
The NAND market, where SanDisk operates, has a different dynamic. NAND is used for storage rather than compute. The AI demand pull for NAND is real but more diffuse than HBM. Enterprise SSD demand is growing, but NAND suppliers have historically struggled to maintain pricing discipline during demand surges. New capacity comes online, inventories build, and prices correct. This cyclical behavior is well-documented and has historically punished investors who bought at peak margins.
The DRAM ETF, as a diversified basket, provides exposure to this theme without single-company concentration risk. This is actually a more conservative choice within the thesis. But options on the ETF introduce their own dynamics. The ETF's liquidity profile is different from individual stocks. The option pricing will reflect broader market volatility in the semiconductor sector, not just the memory sub-sector.
The CoreWeave Question: Private Market Valuation Meets Public Market Scrutiny
CoreWeave is the most interesting position in the reported portfolio because it occupies a strange space between private and public markets. CoreWeave is a private company that has raised significant venture and private equity capital. Its valuation has been marked up substantially as GPU cloud demand has grown. But it has not gone public. The options positions referenced in the coverage presumably relate to public proxies or synthetic instruments, not direct CoreWeave equity.
This raises a question that the media coverage does not address: what is the fund actually buying when it buys CoreWeave exposure? If it is using public proxies like the DRAM ETF or semiconductor indices, the correlation to CoreWeave's actual business is indirect. If it is using OTC derivatives or private market instruments, those are not publicly disclosed. The hedge fund that wants GPU cloud exposure but cannot access the underlying company directly may be making a correlated bet rather than a direct bet. The risk profile of a correlated bet is different from a direct investment. The thesis may be correct while the instrument may be wrong.
In my work analyzing DeFi protocols and blockchain infrastructure companies, I have learned to distinguish between the underlying economic thesis and the specific instrument used to express it. These are not the same thing. A correct thesis executed through the wrong instrument can produce a loss. An incorrect thesis executed through a well-structured instrument can produce a gain. The instrument matters independently of the thesis.
The Leverage Problem Will Not Resolve Itself
I want to return to the leverage question because it is the thread that connects the fund's summer collapse to its current positioning. The reported liquidation to Castle Securities was reportedly conducted at a discount. This means that assets were sold for less than their fair value to generate liquidity. The fund lost wealth not just because positions fell in price, but because the mechanism of liquidation extracted additional value as a cost of speed and certainty.
If the fund is now using options to re-establish exposure, the question is whether the premium being paid reflects the current volatility environment accurately. Options pricing is not static. It adjusts based on implied volatility, time to expiration, and the risk-free rate. In a period of elevated market uncertainty, options premiums are higher. The fund is not just betting on the direction of AI infrastructure stocks. It is betting that the current implied volatility is too high or too low relative to realized volatility. These are different bets.
The additional complexity is that the fund's recent history may affect its ability to trade on optimal terms. Prime brokers and counterparties will be aware of the fund's recent stress. They may widen spreads, demand higher margin, or limit the types of strategies that can be employed. The fund that was managing $45 billion in July 2024 may have had more negotiating leverage with counterparties than the fund that is managing approximately $10 billion now. This is not a trivial consideration. Counterparty terms affect the effective leverage of every position.
What Bulls Get Right: The Demand Is Real
I want to be fair to the underlying thesis, because dismissing it entirely would be intellectually dishonest. The AI infrastructure buildout is not a narrative. It is a capital expenditure cycle that is visible in the earnings reports of NVIDIA, the capacity announcements of hyperscalers, and the order books of data center construction firms. The demand for compute is not fictional. The demand for memory to feed that compute is not fictional. The demand for power to run those data centers is not fictional.
The Storage is real. The bottleneck is real. The capital flow is real. A fund that correctly identified these structural trends in 2023 and positioned accordingly was making a fundamentally sound decision. The problem was not the thesis. The problem was the execution vehicle and the leverage profile. The fund conflated a correct directional thesis with a correct assessment of how to profit from that thesis. These are separable judgments. You can be right about the direction of a market and wrong about the instrument, the timing, or the leverage required to express the view profitably.
What Bulls Miss: Execution Is Not Separable From Thesis
The contrarian angle that the market is underpricing is not whether AI infrastructure demand is real. It is whether a fund that just experienced a leverage-driven collapse is structurally capable of exploiting that demand more effectively than it did in the previous cycle. The answer to that question depends on whether the fund has changed its risk management, its leverage profile, its liquidity management, or its investor base composition. None of these changes are visible from the outside.
The market is being asked to trust that the same manager, with a smaller capital base and a recent track record of catastrophic drawdown, has somehow fixed the structural problems that caused the collapse. The evidence presented for this trust is the manager's technical background and the coherence of the underlying investment thesis. Neither of these is sufficient. Technical expertise does not prevent liquidity mismatches. A coherent thesis does not prevent margin calls.
The Takeaway: Watch the Leverage, Not the Narrative
The Situational Awareness fund's comeback will be written about as a test of whether the AI infrastructure thesis survives a liquidity event. The more important question is whether the fund has addressed the specific mechanism that caused the liquidity event. Until regulatory filings, investor disclosures, or observable behavior demonstrate a change in leverage profile and risk management, the default assumption should be that the structural vulnerabilities remain.
The code compiles, but the reality bankrupts. A correct thesis does not prevent a broken vehicle from arriving at the wrong destination. I do not trust the narrative; I trust the leverage ratio. And that number is not disclosed.
The transaction is permanent. The mistake is not. The question is whether the fund's investors understand what they are holding, or whether they are trusting a story that has already cost them seventy-eight cents on every dollar they entrusted to this vehicle. Illusion has a price tag. Truth has none. The market will eventually discover which one this fund is selling.
