A single fund, down roughly 78% from its peak, is quietly rebuilding its book through call options β and the shape of that bet tells you more about the AI infrastructure complex than any earnings call this quarter.
Peak assets north of $45 billion. Present assets near $10 billion. The gap is not a rounding error; it is a crime scene. The fund is Leopold Aschenbrenner's Situational Awareness, and the instrument is no longer leveraged spot. It is now a ladder of call options with expiry dates β a deliberate insertion of a time stop into a thesis that was previously sold to limited partners as a decade-long conviction trade. Five names are said to sit at the center of the new structure: AMD, SK Hynix, SanDisk, CoreWeave, Bloom Energy. Compute, memory, storage, cloud, energy. Read that list twice. It is not a portfolio. It is a wiring diagram.
The reporting behind this β CNBC attributing to unnamed sources, a personal-investor tweet from Shay Boloor doing the amplification, an SEC subpoena against the fund's counterparty banks running in the background β would not survive a single line of my own diligence checklist. And yet the trade construction itself is legible, and that legibility is the story. Because when a fund that markets itself on AGI arriving in 2027 re-prices its public book toward infrastructure bottlenecks, it is telling you β between the lines, in the only language that costs money β that the smartest money has stopped buying the model layer and started buying the plumbing.
Context: Who is betting, and with whose money
Strip the drama and you are left with a footnote-grade biography. Aschenbrenner is a former OpenAI Superalignment researcher, the author of the long-form "Situational Awareness: The Decade Ahead" series, and a public holder of an extreme timeline: artificial general intelligence somewhere around 2027. He converted that reputation into capital, raised a fund, and let the fund express the thesis with leverage. Through July, that worked β right up until it did not. According to the reporting, the fund's leveraged book amplified a decline in its equity positions, triggered margin calls for additional cash, and forced a liquidation that took assets from the $45 billion neighborhood down toward $10 billion. Ken Griffin's Citadel reportedly acquired some of the distressed positions at a discount. That is vulture-capital behavior. It tells you the fund was in a passive unwind, not a controlled rotation.
Now the same manager is back, this time without leverage in the spot sense β call options carry a defined premium, a capped loss at the position level β but with an explicit expiry attached. The instruments are said to be dated. That single structural fact is worth more than the entire headline. A spot position says "I will hold this belief indefinitely." A dated call option says "I must be right by a specific Friday." The conversion from one to the other is the conversion of a religion into a trade.
The reported targets map cleanly onto the physical stack an AI cluster demands. AMD supplies the compute silicon as NVIDIA's only meaningful scale challenger. SK Hynix supplies high-bandwidth memory, the packaging-adjacent component that has become a hard constraint on accelerator output. SanDisk, freshly separated from Western Digital, supplies NAND β the flash tier where training corpora land and inference caches live. CoreWeave supplies rented GPU capacity, a pure-play AI cloud bound tightly to NVIDIA's ecosystem. Bloom Energy supplies on-site power via solid-oxide fuel cells, aimed squarely at the data-center electricity problem.
Five names, five chokepoints, one underlying driver. I will come back to that driver and why it is the most dangerous quality of the trade. But first, the newsroom framing is wrong in a way that matters. This is being reported as "AGI researcher bets on AI stocks after wipeout." That framing invites you to copy the picks. Do not. Copy the structure instead. The picks are a reflection of a thesis; the structure is a confession about its lifespan.
The full-stack trade is really a single-factor exposure wearing a costume
Here is where I part company with nearly every commentator I have read on this. The consensus read is that Aschenbrenner has "diversified across the AI value chain." That is a category error, and it is the exact error that produced the July margin call.
Five holdings across five sub-sectors is not diversification when the sub-sectors share a single demand variable. AMD's accelerator revenue, SK Hynix's HBM pricing, SanDisk's NAND mix, CoreWeave's utilization, and Bloom's data-center backlog all move on the same variable: aggregate AI capital expenditure. If hyperscaler capex slows, all five compress together. The portfolio is not five bets. It is one bet, expressed five times, each expression carrying its own idiosyncratic risk on top.
This is the same analytical failure I watched play out in DeFi in 2021, and I have the scar tissue. I joined a Series A startup in Melbourne that year as a junior researcher and found, buried in the analytics, that roughly 70% of a headline liquidity pool was actually composed of governance tokens from the same protocol β self-referential collateral dressed as yield. When the underlying governance token fell, the "diversified" pool fell in lockstep. I wrote a memo proposing a pivot toward real-world-asset tokenization instead of speculative yield; leadership rejected it; I later anonymized the memo and published it. The lesson generalizes perfectly here. A basket whose members share a common risk factor is not a basket. It is a lever with extra steps.
And the July event was exactly that lever clearing. When the AI narrative wobbled, five assets fell as one, leverage multiplied the fall, and the margin desk did the rest. The new options structure reduces the loss ceiling per position. It does nothing to reduce the correlation. If anything, buying dated calls on five highly correlated names is a concentrated bet on the timing of a synchronized move β a bet that becomes worthless not if you are wrong on direction, but simply if you are early.
Read the energy leg: it is the most forward-looking line in the book
Of the five, the one I find genuinely informative is Bloom Energy, and not because I like fuel cells. I find it informative because of what it implies about where this investor thinks the binding constraint will migrate.
Every prior cycle of AI-infrastructure enthusiasm has been framed around one bottleneck at a time. First it was GPUs β "we cannot buy enough of them." Then it was HBM and advanced packaging β "the accelerator is fine, but we cannot get the memory stacked or the die packaged fast enough." The choice of SanDisk extends that logic one tier further, to flash: the storage layer where a training corpus actually lives, and where inference-time key-value caches consume enormous memory bandwidth. Choosing Bloom Energy jumps to an entirely different layer: the electricity itself, and specifically distributed, on-site generation rather than the centralized grid.
That is not a hedge-fund impulse. That is an engineering judgment, and it happens to be correct on the physics. A grid interconnection queue in most developed markets runs three to five years. A data hall can be commissioned in months and filled with accelerators in weeks. The mismatch is structural and getting worse, not better. When the deployment speed of compute outruns the delivery speed of electrons by an order of magnitude, the constraint stops being the chip and starts being the substation.
I have spent this year framing crypto as the operating system for autonomous economic entities, and the energy leg is where that framing stops being a thought experiment. Self-sovereign AI agents that transact, meter their own compute, and settle payments will need metered access to power at a granularity the utility model was never designed to price. A hedge fund buying a fuel-cell maker is, perhaps unwittingly, pricing the same bottleneck my own research has been circling: the physical substrate under agentic economies.
The omission that says the most: no NVIDIA, no TSMC
What is absent from the reported basket is, to my eye, more revealing than what is present.
Aschenbrenner did not reportedly buy NVIDIA. He bought NVIDIA's memory supplier, NVIDIA's cloud partner, NVIDIA's accelerator rival, and the power vendor selling into NVIDIA's deployments. That is a "picks-and-shovels around the picks-and-shovels" construction β the sell-side of the sell-side. It is a deliberate refusal to pay the incumbent's multiple, and it is internally coherent only if you believe the incumbent's valuation already fully embeds the demand cycle. In other words, the new book is not a bull bet on NVIDIA. It is a bet that the incremental return in the AI complex has rotated to the challenger and the input-supplier layer.
The absence of TSMC is the lazier omission. If you believe the chokepoint thesis β that AI value accrues to whoever holds the physical constraint β then the advanced-packaging monopoly is the single most obvious expression, and it is missing. The benign reading is that the position is already held through a regional vehicle, or that the ADR lens makes it awkward. The uglier reading is that the basket is not a systematic chokepoint map at all. It is a collection of names with the best call-option convexity β high beta, high implied vol, storylines that move β assembled by someone who needs to be right fast, not right thoroughly.
I will not pretend to know which. But I do know that a genuine full-stack thesis with a monopoly supplier excluded is a thesis with a hole in it, and holes in theses are where leverage finds its way in.
The crypto angle nobody in the sell-side piece will write
Here is the part that matters to a cross-border payments researcher, and it is the connective tissue every mainstream report on this event skipped.
The AI infrastructure complex and the crypto settlement layer are colliding on the same problem: how autonomous software pays for scarce compute and power without a human in the loop. The five-name basket is a bet on who owns the physical capacity. But capacity is only half the system. The other half is the rail that meters and clears machine-to-machine expenditure, and that rail is being built right now, largely outside the public-market equity complex.
In 2020, in my final year of a computer-science master's, I built a Python simulation pitting legacy correspondent-banking fees against early ERC-20 stablecoin transfers across 10,000 mock transactions. The cost disparity came out at roughly 40%, and the interesting part was not the headline gap β it was the variance. Legacy rails had high, stable, spread-based costs; the tokenized rail had low median cost with fat tails driven by gas conditions. That fat tail is the whole governance problem of autonomous payments in miniature. A human treasurer tolerates settlement variance. An agent making a million micro-decisions a minute does not.
This is why I keep hammering that the settlement layer is the unmodeled leg of every AI-infrastructure equity thesis I read. The equity basket prices compute, memory, storage, cloud, and power. It prices zero of the clearing mechanism. And that mechanism is where the friction actually lives β where a stablecoin corridor, a metered API, and a power contract have to reconcile against each other in real time.
I first put a version of this in writing in 2025, proposing a "Proof-of-Workload" consensus framing in which compute providers post verifiable workload attestations that double as the settlement primitive for AI-to-AI payments. The paper circulated, I got a Consensus speaking slot out of it, and the practical effect was to make me insufferable at conferences. But the underlying point survives the ego: if AI agents are to become primary liquidity providers in DeFi β my standing call is that they will, within this cycle β then the settlement rail has to carry both the payment and the proof that the underlying work was actually performed. A stablecoin can carry the payment. It cannot, today, carry the proof.
So when I look at a basket of five AI-infrastructure names, the loudest thing I hear is the silence around the sixth leg. The compute, memory, storage, cloud, and power layers are all being priced by public markets. The settlement layer is being priced by nobody β or by token markets that public funds will not touch and that the bankruptcy-remote structures of the fund's limited partners almost certainly prohibit. That asymmetry is either the biggest mispricing in the complex or a very good reason it remains untouched by serious allocation. Which one it is depends entirely on whether the rail can be made regulator-legible, and I will come to that.
The regulatory floor underneath all of it
Last year, in a mid-level role at a global fintech consultancy, I led a team of three analyzing how the European Union's Markets in Crypto-Assets framework would reshape Asian remittance corridors. To get anywhere, I had to negotiate with compliance officers to obtain non-public audit trails, and what those trails showed was not flattering to the industry's self-image: about 60% of nominally "decentralized" exchanges were still leaning on centralized custodians somewhere in the stack. My report ended up cited by two Australian banks and shifted their outsourcing strategies. It also permanently recalibrated how I read the word "decentralized" in any pitch deck.
That experience is the lens through which I read the SEC subpoena against the banks that traded with the fund. A subpoena is not a charge. But it is a signal about where enforcement attention is pooling, and it is pooling at the seam between a flashy strategy and the boring custodians who actually hold the assets. That seam is exactly where the next phase of the AI-crypto convergence will get tested β not at the model layer, not at the equity layer, but at the custody and settlement layer, which is where the 60% figure lives.
An AI-agent payments system built on rails that secretly terminate in a handful of centralized custodians is not a decentralized economic operating system. It is a faster version of the correspondent-banking layer I simulated against in 2020, with better branding. And here is the uncomfortable corollary: an equity basket betting on AI infrastructure consumes that infrastructure's output, but it has zero exposure to whether the output gets settled in a way that any regulator accepts. The fund is long the physical layer and short β implicitly, unknowingly β the legal layer.
The contrarian cut: the time premium is the real position
Everyone is reading this event as "will he be right on AI?" That is the wrong question. The right question is "what does the option structure reveal about his confidence in the timeline?"
A dated call option does not merely cap loss. It imports a theta clock into a thesis that, until July, was marketed as a decade-long hold. You do not attach an expiry to a belief you intend to hold forever. You attach an expiry to a belief you must convert to cash within a window. The migration from leveraged spot to dated options is therefore best read not as de-risking but as a tacit admission that the original time horizon was unaffordable β that the fund no longer has the luxury of waiting out a drawdown, because its limited partners, its post-liquidation balance sheet, and the SEC all impose a horizon shorter than the AGI timeline it sells.
And that is the genuinely contrarian point, the one the crypto-friendly commentariat will resist: the belief that AGI arrives in 2027 has now been converted, by its most prominent holder, into a trade that expires before 2027 does. If you take the structure seriously, the strongest signal in the whole episode is not optimism about AI. It is a quiet repricing of the timeline by the person with the most reputational capital staked on it. When the prophet's own instruments start carrying expiration dates, you should update on the instruments, not the sermons.
The gap I keep returning to is the one that has no issuer. The equity complex prices the physical stack; the token complex prices a settlement layer nobody institutional will touch; and the AI-agent economy that both are supposedly building toward needs the two to meet in the middle β metered compute, provable work, cleared payment, all within one legally legible rail. No public-market basket expresses that convergence, and no token express it under current disclosure standards either.
Which leaves a question I cannot answer and will not pretend to. When the machines start paying each other for power, memory, and compute, who is the counterparty of record β the fuel-cell vendor, the cloud, the stablecoin, or the custodian none of them will admit is holding the assets? The first fund to price that answer will not need a decade. But it will need something the current book conspicuously lacks: a settlement leg that survives an audit. Watch whether the next subpoena lands there, because that is where the trade actually lives.