A fund that peaked above $45 billion in July shrank to roughly $10 billion, and it exited almost every public equity position at a discount, transferring the inventory to Castle Securities. That is the number. Not the comeback headline. The number.
The same manager is now back, buying several hundred million dollars of options tied to AI infrastructure and storage: SK Hynix, SanDisk, AMD, Bloom Energy, CoreWeave, and the Roundhill Memory ETF. Anonymous sourcing. No strikes. No expiries. No notional. No confirmation that the capital is new rather than residual.
Here is why this belongs in a crypto publication. In the same window, a basket of tokenized compute, storage, and GPU-leasing instruments repriced against the identical narrative โ with thinner float, weaker disclosure, and no Castle Securities standing by to absorb the inventory when the margin call arrives. The ledger balances, but the architecture bleeds.
Setting the plumbing first, because a great many people who look impressive need the basics stated plainly.
Situational Awareness is built on a specific engineering claim: that the binding constraint in AI is migrating away from the accelerator itself and into the physical layer around it โ high-bandwidth memory, conventional DRAM, NAND, grid interconnection, and on-site power generation. That claim is not frivolous. HBM capacity has run tight through 2023โ2025, HBM wafer allocation crowds out conventional DRAM supply, and data-center interconnection queues in several regions now exceed hardware lead times. A reasonable person could underwrite that thesis without embarrassment.
On the crypto side, the identical claim got compressed into tokens. There are DePIN GPU marketplaces pricing rental hours, compute tokens pricing a governance claim, tokenized capacity products pricing a revenue-share agreement drafted by an offshore entity, and RWA wrappers pricing a slice of data-center or energy exposure sitting behind a custodian, a legal opinion, and a redemption calendar. None of these instruments hold an HBM stack. None hold a megawatt. They hold claims about claims.
I am treating the underlying reporting carefully because the sourcing is weak. It rests on people familiar with the matter, offers no AUM breakdown, no leverage multiple, no funding source, and no option terms. There is also a date conflict inside the source material โ a 2023 stamp sitting alongside post-2024 instruments โ which I am flagging rather than resolving. Lower your confidence accordingly. What survives the sourcing problem is a structural description, and structure is what I can audit.
Start with the shape, because the shape is the tell. A fund that loses roughly 78% and then re-enters the same factor through options has not changed its risk model. It has changed its instrument. The July collapse was not caused by being wrong about AI. It was caused by expressing a correct directional view through a structure that could not survive the path. Options do not repair path dependence. They amplify it, because the premium decays on a clock that does not care about your thesis.
Now look at what was retained and what was surrendered. Public equities โ liquid, marked hourly, trivially priceable โ went out the door at a discount. Private equity in Anthropic โ illiquid, mark-to-model, no observable print โ stayed on the books. That is a liquidity mismatch, and it is the single most diagnostic detail in the entire story. It tells you the fund optimized for reported volatility rather than survivability. A mark-to-model asset cannot be sold into a drawdown, which is precisely the moment you need it to be sellable. This is not a moral failing. It is a balance-sheet fact.
Map that onto chain and the mismatch reproduces at industrial scale. A DePIN compute protocol typically carries three buckets: a liquid governance token, a vesting allocation held by insiders and the treasury, and a hardware or contract book with multi-year depreciation. That is the same architecture as liquid options paired with illiquid private equity, minus the discount buyer of last resort. When the factor reprices, the treasury's liquid bucket falls first, the vesting schedule keeps printing supply into weakness, and the hardware book stays marked at cost because nobody is marking it at all.
Run the stress test, since stress tests are the only honest form of price prediction. Assume AI capex expectations reset 40% lower across two quarters. Not a crash. A reset. Under that path, the fund's options decay toward premium, margin requirements tighten, and the illiquid Anthropic position becomes collateral of last resort โ the one asset you structurally cannot sell into weakness. That is not a forecast about any specific fund. It is the arithmetic of the balance sheet as described.
Apply the same 40% to a representative compute token with a mid-cap float and three mechanisms fire in sequence. First, staking collateral โ frequently the very same token, accepted as margin โ falls 40%, repricing every loan written against it. Second, unlock schedules calibrated to a bull-market price continue releasing on schedule, because a vesting calendar is a calendar, not a signal. Third, market makers widen, order-book depth thins, and the oracle feeding the lending market lags the venue that actually clears.
On that third mechanism I have direct professional exposure.
In 2026 I led a security audit for a prominent AI-agent protocol integrating with Ethereum, and the critical flaw I identified was not in the agent logic. It was in the oracle data verification path, and the gap permitted up to $12 million in potential exploit value. The agent consumed compute-price and utilization feeds. The feeds were signed. The signing key was sound. What was absent was any binding relationship between the reported metric and a settled event. The oracle was reporting a number about a market rather than the outcome of one.
That distinction is the entire game, and it is being ignored across an asset class that now prices itself on precisely such numbers. When I published the findings on AI-crypto bridge vulnerabilities, three regulatory bodies in Singapore and Europe adopted the framework as a compliance baseline. They adopted it for a boring reason: the framework was never about AI. It was about verification. A price feed that cannot be reconciled to settlement is not data. It is testimony.
Return to the memory thesis, because the thesis itself deserves respect even where the instruments do not.
The engineering case is specific and defensible. HBM stacks consume wafer area disproportionately relative to conventional DRAM, so capacity migration toward HBM simultaneously tightens DDR5 and enterprise SSD supply while AI servers pull demand for both. That is a genuine supply-side squeeze with a real cost curve, and it explains why a sophisticated manager would want memory exposure rather than pure accelerator exposure. It is also โ and this is the part that gets dropped from the pitch โ a cyclical squeeze. Memory has been a commodity cycle for four decades. Reclassifying it as a growth asset does not delete the cycle. It means the cycle now reprices against a growth multiple, which makes the drawdown steeper rather than shallower. Valuation is a fiction; exposure is the reality. The fund is buying options on the exposure. The chain is selling tokens on the valuation. Those are not the same trade, and in a drawdown they are not even the same direction.
The second-supplier logic deserves its own paragraph, because it is carrying more weight than most readers realize. Including AMD alongside memory names implies a bet on the non-incumbent accelerator path โ a bet on share shift rather than on market size. Those are different risk profiles. A memory squeeze pays you if AI capex grows at all. A second-supplier thesis pays you only if capex grows and procurement diversifies, which is a two-condition trade with a software-ecosystem dependency layered on top. On-chain, that nuance vanishes entirely. Compute tokens do not distinguish between accelerator generations, suppliers, or contract structures. They price a single blended narrative and call it a sector.
The power leg is where the on-chain representation is weakest and the physical constraint is strongest. Data-center interconnection queues run years in some jurisdictions. On-site generation and storage become economically rational at a certain queue length, which is why an entity like Bloom Energy enters the trade as an infrastructure solution rather than an energy company. But grid interconnection is a regulatory process, not a token standard. You cannot wrap a queue position in a smart contract and have it mean anything to a regulator, a utility, or a bankruptcy court. What you can do is issue a token that references an entity that holds a position in a queue. That is three layers of abstraction between the buyer and the megawatt.
Which brings me to something I have watched for three years and will keep repeating in print. RWA on-chain has been a storytelling exercise. The obstacle is not technology, and it never was. The obstacle is that institutions do not need a public chain to hold a claim on a data center. They need a custodian, a legal wrapper, and a bankruptcy-remote vehicle โ all of which exist, all of which predate the chain, and none of which become cheaper because the settlement layer is permissionless. What institutions actually want from a chain is narrow and unglamorous: atomic settlement of specific instruments, and auditability. What they get from most AI-infrastructure tokens is a governance claim on a narrative with a two-year unlock tail.
So when the same week produces a headline about an options comeback in memory and power, and a set of tokens repricing on that headline, be precise about the linkage. Nothing on-chain holds the megawatt. Nothing on-chain holds the HBM stack. Nothing on-chain holds the wafer allocation. The linkage is sentiment, and sentiment is the one input that has never once cleared a margin call.
There is a further forensic detail worth extracting from the July event, because it is the kind of thing that never gets recorded as a valuation input. When a forced seller transfers inventory at a discount to a single counterparty, the discount itself is a data point about market depth. It tells you how much size the venue could not absorb and at what clearing level. The fund's counterparty acquired the position at a concession, which is a transfer of value from the seller's investors to the buyer, executed in one transaction with no public print. Read that alongside the retained Anthropic stake and the pattern is unambiguous: the liquid assets subsidized the illiquid ones. Now perform the same exercise on-chain and notice that the equivalent event โ a liquidation cascade โ is also not recorded as a valuation input by any oracle. The liquidations are public in the sense that they are visible. They are not public in the sense that they are incorporated into the price of the collateral accepted by the next lending market. That lag is the systemic seam.
Then there is composability, which converted a single bad factor into a system event in 2020 and has only concentrated further since.
When I modeled the dependency chains of Compound and Aave during the 2020 DeFi Summer, the finding was that at a 50% drop in a major collateral asset, roughly 80% of leveraged positions would fall below their collateralization thresholds. That conclusion contradicted the prevailing sentiment of the period and was subsequently cited by three institutional hedge funds. The method was not exotic. It was a recursive walk through every market that accepted the same asset as margin, counting the independent exit paths available to each position.
Run that walk today on the AI-compute complex. Count how many lending venues accept the same two or three collateral tokens. Count how many of those tokens share a single upstream driver: capex guidance from a small number of hyperscalers, delivered on a quarterly calendar. Count how many liquidators depend on the same oracle update interval. Count how many of those venues use the same staking derivative as secondary collateral. Then recognize that the correlation you are underwriting is not an assumption you are making. It is a documented dependency you are choosing to ignore.
That is how I found the fracture line before the quake struck โ not by predicting the event, but by counting the number of independent paths out of the position. In a properly mapped correlated collateral system, the honest answer is frequently zero, and zero is a mathematical property rather than a sentiment.
One more structural note, because it is the thread running through both the fund story and the token story. Using options instead of spot changes the payoff shape but not the underlying exposure. Directional conviction expressed through leverage converts a thesis into a survival problem, and a survival problem is decided by liquidity, not by being right. The fund survived July by selling liquid assets to a buyer with pricing power and no obligation to be generous. On-chain, the equivalent event is a liquidation engine selling into the same bid โ except the seller has no negotiation, no discount agreement, no counterparty relationship, and no ability to wait for a better print. The mechanism is faster and the outcome is strictly worse. Minted in haste, seized in cold logic.
Here is what the bulls got right, and it is not nothing.
The bottleneck migration is real, and the people who identified it early were correct before they were wrong. Memory and power genuinely are the constraints. The engineering is directionally sound, and a reader who dismisses the entire thesis because one levered vehicle blew up is committing the identical error in reverse โ confusing the instrument with the underlying. Skepticism aimed at a balance sheet should not be mistaken for skepticism aimed at a supply chain.
Second, on-chain markets price this faster than private structures do, for all their faults. Perpetual funding rates, liquidation prints, and unlock calendars are observable in real time. When the compute complex repriced, the funding curve moved before any fund confirmed anything to any journalist. That is genuine information gain, and it belongs to anyone with a dashboard and the patience to read it.
Third, and this is the uncomfortable one: the second entry may be directionally correct. The problem with a 78% drawdown is not that the view was wrong. It is that the structure could not wait for the view to be right. That is a survivability judgment, not a directional one, and conflating the two is how intelligent people end up explaining a correct thesis from a liquidated account.
The shared blind spot on both sides is identical. Everyone is debating whether AI needs memory and power, when the operative question is who can hold the position long enough to be paid.
The thesis is not in question. The instruments are. Before you buy the on-chain expression of a physical bottleneck, perform the arithmetic the marketing deck omitted: how many independent paths out of this position exist, and how long can you hold it if the answer is one. Watch the unlock schedules. Watch the oracle update intervals. Watch whether the collateral accepted by your lending market is the same collateral accepted by every other lending market, because if it is, you are not diversified โ you are early to the same liquidation.