When a fund loses roughly seventy-eight percent of its assets and then re-enters the same trade, the interesting question is not what it bought. It is what it deleted. Leopold Aschenbrenner's Situational Awareness reportedly rebuilt exposure this month through call options on AMD, SK Hynix, SanDisk, CoreWeave and Bloom Energy — five tickers that, read literally, are not a portfolio. They are a bill of materials for a datacenter. One computes. One carries memory bandwidth. One holds cold data. One rents the cluster. One keeps the lights on. My reflex — the same reflex that once had me decompile a DePIN contract just to prove that “decentralized GPUs” were a queue with a logo — was to check whether the crypto market is pricing the identical stack. It is. It is pricing it more confidently, and with far less engineering.
Context first, because these numbers circulate without provenance. Aschenbrenner is not a generic AI bull; he is a former OpenAI Superalignment researcher and the author of the Situational Awareness series, which argues AGI arrives around 2027. The fund, named after the thesis, reportedly peaked above $45 billion and collapsed to roughly $10 billion when leveraged longs met a drawdown and margin calls followed. Citadel reportedly bought the distressed positions at a discount. The SEC has subpoenaed the fund's dealings with Wall Street banks. The rebuild swaps leverage for options. The sourcing is thin — a CNBC report citing unnamed people, a Twitter thread, no 13F filing — so treat every figure as directional, not audited. I flag this because the crypto press, which carried the story, has an incentive to launder a distressed AI fund into a bullish AI narrative.
Now the mechanics, which are more interesting than the narrative.
Map the five targets. AMD is the compute layer, the only accelerator vendor with the volume to challenge NVIDIA. SK Hynix is HBM, the high-bandwidth memory that feeds the accelerator and remains genuinely supply-constrained. SanDisk is NAND, the substrate for training data and inference caches. CoreWeave is GPU cloud, the wholesale layer that NVIDIA's ecosystem underwrites. Bloom Energy is solid-oxide fuel cells — on-site power for a machine that now draws like a small city.
Here is what a “he is bullish on AI” headline buries: these five positions share a single driver — AI capital expenditure. They are not diversified. They are one risk factor written five times, in five dialects. When the AI narrative repriced, all five fell together, and leverage converted correlation into insolvency. I keep finding the same error on-chain: a protocol ships five “independent” security modules, then wires all five to one admin key. Modularity is an entropy constraint, not a diversification strategy, unless the failure domains are genuinely disjoint. Aschenbrenner's basket fails that test, and so does most of the AI-compute token sector.
This is not a theoretical concern. Run the correlation. Five positions whose cash flows all hinge on the same capex line behave, in a drawdown, like a single position at five times the notional — and a portfolio levered on top of that behaves like one leveraged position, which is precisely the structure that produced the original loss. The option rebuild does not repair the concentration; it only caps the per-position damage. The correlation is still sitting under the wrapper, waiting for the next capex revision.
Margin is the mechanism worth naming here. A leveraged spot book has no expiration, so it survives a bad week and dies on a bad month — the broker, not the thesis, sets the horizon. That is the failure mode that took the fund from its peak to a fraction of it: not a wrong thesis, a thesis funded with borrowed money and no clock. The option rebuild replaces an involuntary liquidation schedule with a chosen one. That is an upgrade in structure, not in conviction.
Consider the DePIN pitch, the crypto-native version of this trade. Render, Akash, io.net and the newer verification-first compute networks all argue that tokenized GPU supply will undercut CoreWeave on price. The argument has a real kernel: idle silicon exists, and payment rails are the friction. But the kernel is buried under three engineering problems that tokenomics cannot price away. First, verification — if you cannot cheaply prove that a remote worker actually ran the job, you have not decentralized compute, you have decentralized trust in a stranger. I have audited enough of these modules to know the soundness errors hide in proof aggregation, not the happy path. Tracing the gas leak in the untested edge case is the entire job, and almost nobody funds it. Second, bandwidth. Training and inference are not embarrassingly parallel the way the whitepapers imply; checkpointing, gradient sync and KV-cache movement are bandwidth-bound, and bandwidth is physical. Latency is the tax we pay for decentralization, and it is not deductible. Third, slashing — every marketplace eventually needs a penalty for non-delivery, and every penalty clause is a new attack surface. The code is a hypothesis waiting to break.
And a large share of on-chain compute supply is emission-funded, which makes its headline utilization a subsidy artifact. Stop the incentives and the miners leave. The reflex that hollowed out liquidity mining applies one layer up, undiluted.
This is where the Aschenbrenner basket and the on-chain compute sector diverge, and the divergence is the whole story. He paid for HBM, storage and power — the layers that are actually scarce and actually measurable. The token sector pays for the narrative of compute, priced in a unit with no strike and no expiry. One is a claim on a bottleneck. The other is a claim on a sentence.
Which brings me to the part the coverage skipped. The stated lesson of this episode is “distressed legend returns to a winning trade.” The structural signal is the option wrapper. When you rebuild with call options, you do not merely cap your downside at the premium — you inherit a clock. Every contract carries a strike and an expiry, and those two parameters encode a claim about when the thesis pays, not whether it is true. A leveraged spot book can hold a belief in perpetuity and simply bleed margin. An option book cannot. The belief has been time-bounded, and the wrapper admits it. Conviction with an expiry is a different instrument from conviction without one.
The crypto read inverts this. On-chain AI tokens are, almost without exception, perpetual claims: no strikes, no delta, no carry, no theta, no forced reckoning. Their holders have taken the Aschenbrenner thesis, deleted the time constraint, deleted the leverage discipline, and deleted the accountability that a quarterly filing imposes. If his options are a bounded bet on a real bottleneck, the tokenized version is an unbounded bet on a vibe. When AI capex next reprices — and it will, because capex is lumpy and depreciation is not — the assets with no expiry and no cash flow are marked down first, not last.
There is a second blind spot, quieter. Note what he did not buy. Not NVIDIA, whose valuation already encodes the consensus; he bought its memory supplier instead — selling shovels to the shovel sellers, a wager that the bottleneck migrates downstream into HBM. Not a datacenter REIT, but a fuel-cell vendor — a wager that the binding constraint is power, not floor space. Both are challenger bets: AMD against CUDA, Bloom against the grid. Challenger baskets carry convexity and fragility in equal measure, and crypto has been trading that exact reflex for two years, rotating from one “ETH killer” to the next. The challenger trade is the same trade in every asset class: you are paid for being early and punished for being wrong about the timing. Options let you price the timing. Tokens do not.
So watch the term structure, not the headline — strikes and expiries tell you what the fund believes about when, which is the only falsifiable part of any AGI thesis. Watch the energy arithmetic, because grid interconnection queues run three to five years while clusters deploy in months, and that mismatch is the most durable bottleneck in the stack; a fuel-cell position is a bet on that specific spread. Watch SEC escalation, a confidence discount no amount of conviction offsets. And watch the crossover: if AI capex finally uncouples from crypto liquidity, the tokens that borrowed the AI story without borrowing its cash flows get repriced first. Debugging the future one opcode at a time means reading the instrument, not the press release.
