When the code bleeds, the ledger keeps the truth. Leopold Aschenbrenner learned this lesson in the most expensive way possible.
The former OpenAI superalignment researcher watched his hedge fund, Situational Awareness, hemorrhage from a peak exceeding $45 billion to roughly $10 billion in assets under management. The culprit: leverage amplifying drawdowns, triggering margin calls, forcing a tactical retreat. This was not a valuation fluctuation. This was a systemic risk management failure disguised as a market correction.
Now he's back. The same five positions. The same AI infrastructure thesis. But the instrument has changed.
The Tactical Pivot That Reveals Everything
The original positions were built on leverage. Buying power concentrated in a handful of AI infrastructure names. Infinite time horizon. Conviction as collateral.
The new positions use call options with defined expiration dates. The difference is not cosmetic. When you hold leveraged spot, you can wait. Time is your ally. When you hold options, time works against you. Theta decay is a daily tax on your thesis.
This structural shift tells me Aschenbrenner has narrowed his window for conviction expression. He's no longer betting on AGI arriving "eventually." He's betting on AGI arriving by a specific date or his capital expires.
The Five-Position Black Box
AMD. SK Hynix. SanDisk. CoreWeave. Bloom Energy.
Decode the logic. Each position maps to a critical node in the AI compute stack:
- AMD → The only scale challenger to NVIDIA in AI acceleration. The MI300/MI325 trajectory suggests competitive momentum, but ROCm versus CUDA remains a software ecosystem gap that requires ongoing investment to close.
- SK Hynix → HBM bandwidth is the memory wall bottleneck. AI training workloads are fundamentally memory-bandwidth-bound. Hynix holds technical leadership, but capacity expansion creates a supply-side question that matters for 2026+ valuations.
- SanDisk → NAND storage underpins both training dataset handling and inference caching. Underappreciated as an AI beneficiary because it lacks the headline glamour of GPU names.
- CoreWeave → Pure-play GPU cloud. No enterprise software diversification. No legacy business to hide behind. The most concentrated AI infrastructure bet in the portfolio, and therefore the most volatile.
- Bloom Energy → The outlier. Fuel cell technology for distributed data center power. This is not a traditional utility play. This is a bet on the electricity bottleneck becoming the binding constraint for AI compute deployment within the next 18-24 months.
Together, these five names constitute a thesis about AI infrastructure as a system, not a collection of isolated winners. The chain breaks if any single node fails.
What the Market Is Not Seeing
Here is the uncomfortable arithmetic: five positions, one risk factor.
All five names derive value from a single assumption—that AI capital expenditure continues expanding at exponential rates. When the narrative shifts, they move together. This is not diversification. This is concentration expressed through sector exposure.
The July 2024 wipeout demonstrated this precisely. As sentiment rotated away from AI pure-plays, all five names compressed simultaneously. Leverage amplified the damage because there was no cross-sector hedge embedded in the portfolio structure.
The options structure changes the loss ceiling but does not change the thesis correlation. A concentrated bet on AI infrastructure delivered through options still carries concentrated AI infrastructure risk.
The Citadel Signal Nobody Is Discussing
Ken Griffin's Citadel acquired distressed positions from this fund at a discount. This is vulture capital behavior. It signals the fund was in forced liquidation territory, not orderly rebalancing.
Aschenbrenner and Citadel, acting on the same underlying AI thesis, reached opposite conclusions about the fund's solvency. One sold at a discount. One bought the discounted assets. Someone misjudged the timeline.
The question is not whether AI infrastructure will appreciate. The question is whether this specific fund survives long enough to participate in that appreciation.
The SEC Variable
Regulatory subpoenas targeting the fund's interactions with Wall Street banks introduce a dimension that most commentary ignores. A formal investigation does not require a conviction to damage a fund's operation. Prime brokerage relationships become complicated. Counterparty confidence erodes. LP redemption requests accelerate regardless of underlying thesis quality.
The fund retained Anthropic private equity positions while liquidating public market exposure. This is rational portfolio segmentation—long-duration private equity conviction separated from medium-duration public market trading. But it also signals the fund's public market activity is under constraints that prevent unlimited thesis expression.
The Energy Bet: Signal or Noise?
Bloom Energy deserves separate attention. Selecting fuel cell technology over traditional power utilities reveals a specific hypothesis: the next AI infrastructure bottleneck is electricity, not compute.
GPU supply is normalizing. HBM supply is expanding. But data center power infrastructure takes 3-5 years to scale versus months for GPU deployments. This creates a temporal mismatch that Bloom Energy exploits through distributed, deployable generation capacity.
If this thesis is correct, traditional utility stocks remain structurally inadequate for AI power demand. If it is incorrect, Bloom Energy becomes an expensive bet on a problem that grid operators will solve through conventional means.
What This Means for AI Infrastructure Positions
The Aschenbrenner case offers three lessons that apply beyond this specific fund:
First, conviction without risk management is a liability, not an edge. The 78% drawdown did not result from a wrong directional call. It resulted from position sizing that left no room for thesis timing variance.
Second, the shift to options reveals thesis timelines matter as much as thesis direction. "AI will transform infrastructure" and "AI will transform infrastructure within 18 months" are fundamentally different trades requiring different position structures.
Third, the full-stack infrastructure thesis is intellectually coherent but operationally concentrated. Investors replicating this theme need explicit cross-sector hedges or explicit acknowledgment that they are making a single-factor bet on AI capital expenditure duration.
The positions will be tested. The SEC investigation will resolve. The options will expire or be exercised. Whatever the outcome, the structural choices made after the wipeout—the move from leverage to options, the retention of private equity conviction, the energy sector addition—will determine whether this is a comeback story or a slow-motion capitulation.
Markets do not care about conviction. They care about capital structure and time horizons.
The code does not lie.