From Essay to Portfolio
Leopold Aschenbrenner’s June 2024 essay [Situational Awareness: The Decade Ahead]1 is the unifying document for the fund. It argues five things:
- Straight lines on log graphs — current scaling trends extrapolate to AGI by ~2027 and to systems materially smarter than humans by ~2030.
- The constraint is industrial, not algorithmic. Training the next generations of frontier models requires data centers that consume tens to hundreds of GW of dedicated power, fab capacity to produce tens of millions of leading-edge accelerators per year, and the power, cooling, fiber, and land to host them.
- The dollar amounts are unprecedented. Aschenbrenner explicitly forecasts cumulative AI capex on the order of trillions of dollars across the late-2020s — comparable to the largest infrastructure buildouts in modern industrial history.
- The US must onshore production. Concentration of leading-edge logic at TSMC in Taiwan is treated as a national-security risk; Intel Foundry, Tower, and US power buildouts gain strategic value beyond their standalone economics.
- The model labs themselves are not necessarily the best public-equity expression — they are mostly private (OpenAI, Anthropic, xAI), or capture only a slice of value (Google, Meta). The picks-and-shovels — power, data centers, optics, chips — capture the spend regardless of which lab “wins.”
How That Translates Into Positions
The Q4 2025 13F maps cleanly onto the essay:
| Essay claim | Portfolio expression |
|---|---|
| Hundreds of GW of new power demand | Bloom Energy, EQT (gas feedstock), Solaris, PSIX, Babcock & Wilcox, Liberty Energy, ProPetro |
| Data centers will be the binding constraint | Core Scientific, IREN, Applied Digital, Cipher, Riot, Hut 8, Bitdeer, Bitfarms, CleanSpark, WhiteFiber |
| GPU-cloud capacity is the marginal product | CoreWeave (common + leveraged via calls) |
| Onshoring leading-edge logic | Intel (calls), Tower Semiconductor |
| Networking inside and between DCs is a bottleneck | Lumentum, Coherent |
| Memory walls bind training | Sandisk |
| AI displaces labor-arbitrage IT services | Infosys puts |
What is conspicuously absent is just as informative:
- No NVIDIA long. Q3 2025 actually held NVDA puts (closed by Q4). The fund expresses NVDA exposure indirectly via CoreWeave (the largest non-hyperscaler buyer of NVDA) and via the data-center hosts.
- No hyperscaler longs (no MSFT, GOOGL, META, AMZN). The thesis is that capex flows through the hyperscalers to the suppliers below them.
- No model labs. Press references private exposure to Anthropic, but it is not in the 13F.
- No TSMC long. Briefly held TSM puts in Q3 2025; closed by Q4. Onshoring expressed via Intel and Tower instead.
The Leverage Signal
The fund’s reported AUM (~5.5 B. Two factors explain the gap:
- Listed call options on CoreWeave, Intel, Bloom Energy, and EQT are reported at the value of the underlying — not the premium. The fund’s actual cash outlay is a fraction of the headline notional.
- Concentration with leverage. The top three line items (BE, CRWV calls, INTC calls) account for ~43% of the 13F. Aschenbrenner has publicly said he has “most of his net worth” in the fund, and the book is run as a high-conviction expression of the essay rather than a diversified portfolio.
Quarter-Over-Quarter Posture Shifts
| Quarter | Reportable value | Net posture |
|---|---|---|
| Q4 2024 | $254.8 M | Initial book — utility/hyperscaler-adjacent (CEG, VST, TLN, VRT, MOD, MRVL) |
| Q1 2025 | $1,005.6 M | Pivot to data-center operators (CORZ, IREN, APLD), first INTC calls |
| Q2 2025 | $2,123.0 M | Concentration; large SMH ETF puts as macro hedge |
| Q3 2025 | $4,138.4 M | Major expansion — added BTC-miner-pivot complex, optics, storage, broad puts on AI semi names |
| Q4 2025 | $5,516.8 M | Closed most semi puts (NVDA, AVGO, MU, TSM, SMH); doubled down on power (BE became #1) and on the BTC-miner-pivot complex |
The Q3→Q4 rotation is the key tell: the fund moved from a hedged AI-semi book in Q3 to an unhedged, concentrated power + data-center + optics + Intel-call book in Q4. That is consistent with conviction that the bottleneck has shifted from chips (which Q3 puts implied could correct) to power and physical capacity (which Q4 longs imply will run).