The Aschenbrenner Thesis: AI Infrastructure as the Trade of the Decade
Situational Awareness LP | Q4 2025 13F Analysis | March 2026| Metric | Value |
|---|---|
| AUM (13F equity) | $5.52 Billion |
| Inception AUM | ~$255 Million |
| Q4 2025 Holdings | 29 positions |
| H1 2025 Return* | +47% (after fees) |
01 Who Is Leopold Aschenbrenner?
Leopold Aschenbrenner, 24, is a former OpenAI Superalignment researcher fired in mid-2024 who published a 165-page manifesto titled Situational Awareness: The Decade Ahead shortly after leaving the lab. The essay argued that AGI was arriving faster than most appreciated, and that the real economic prize would fall to those who built or owned the physical infrastructure enabling it. He immediately translated the thesis into a hedge fund, Situational Awareness LP, seeded by Patrick and John Collison (Stripe), Nat Friedman, and Daniel Gross.
The fund started with roughly $255M in commitments at end of 2024. By Q4 2025, twelve months later, its reported 13F long equity book stood at $5.52B. That is not AUM in the traditional sense (shorts, derivatives, and international positions are excluded from 13F filings), but the trajectory is unambiguous.
He has almost all of his net worth in the fund. He is not a tourist.
02 The Core Thesis (From Situational Awareness)
The fund's investment logic follows three sequential bottleneck claims, each driving the next phase of capital deployment:
| Phase | Bottleneck Identified | Investment Expression |
|---|---|---|
| Phase 1 (2024) | GPU supply is the binding constraint. AI labs cannot train without more chips. | NVIDIA, Broadcom, TSMC, Micron |
| Phase 2 (late 2025) | GPUs becoming commoditized. New bottleneck is electricity and physical data-center capacity. | Bloom Energy, CoreWeave, Core Scientific, Bitcoin miner repurposing, Kilroy Realty |
| Phase 3 (emerging) | Optical networking and memory become chokepoints. Connectivity between GPUs matters as much as the GPUs themselves. | Lumentum, Coherent, Tower Semiconductor, SanDisk |
The overarching macro frame: AGI by ~2027 is the operating assumption. If that is even directionally correct, the capex required to support it dwarfs anything the grid and data-center base can currently handle. Every investment is a bet on that capex being real.
03 Portfolio Breakdown: Q4 2025 13F
Top Holdings (13F Long Equity, as of 31 Dec 2025)
| Company | Ticker | Est. Position | Thesis |
|---|---|---|---|
| Bloom Energy | BE | ~$876M (16%) | Behind-the-meter solid-oxide fuel cells powering AI data centers off-grid. Modular, deployable fast. ~$20B demand backlog. Revenue +34% in 2025; guiding +40% in 2026. |
| CoreWeave | CRWV | ~$800M+ (15%) | Premier neocloud. GPU-as-a-service at scale for AI labs and enterprises. Added $300M+ in Q4 on top of ~$500M in Q3. Long-term HPC hosting contracts with hyperscalers. |
| Intel | INTC | Calls (large) | Contrarian national-security play. Only US-based leading-edge foundry. Government took 9.9% stake Aug 2025. |
| Lumentum | LITE | Significant | Optical networking for data interconnect layer inside and between data centers. |
| Core Scientific | CORZ | ~$419M | Bitcoin miner pivoting to AI hosting. Has permits, grid access, physical real estate. 12-year HPC hosting contracts with CoreWeave. Leopold holds ~9.4% (activist stake). |
| IREN | IREN | Meaningful | Bitcoin miner repurposing sites for AI/HPC. Buy the entity that already has grid access instead of queuing for permits. |
| EQT Corp | EQT | Added Q4 | Natural gas production. Upstream exposure to the energy thesis. |
| Riot Platforms | RIOT | ~$78M | Bitcoin miner with substantial power footprint pivoting to AI data center hosting. |
| Hut 8 | HUT | ~$40M | Miner-to-AI hosting thesis. Owns owned and operated data centers. |
| Kilroy Realty | KRC | New in Q4 | Data center REIT. Physical real estate with existing power and cooling infrastructure. |
| SanDisk / Storage | N/A | New build | Storage and memory demand driven by AI inference and training data. Emerging Phase 3 bet. |
| Coherent Corp | COHR | Added Q4 | Optical components. Same connectivity thesis as Lumentum. |
| WhiteFiber | WFBR | ~$28M | AI infrastructure provider. Q3 revenue +64% YoY. Phase 3 connectivity exposure. |
Key Exits: What He Sold and Why
The Q4 2025 filing shows a deliberate exit from the GPU cycle trade that defined 2024:
- NVDA (NVIDIA) Exited equity; held $300M PUT position (profited as NVIDIA corrected). Thesis: GPU value is market-priced in.
- AVGO (Broadcom) Exited. Same logic: the chip-designer value capture window has closed.
- TSM / MU (TSMC & Micron) Exited. Semiconductor supply chain now priced for perfection.
- INTC equity (Intel) Sold shares but retained massive CALL options. Freed liquidity while keeping upside on the national-security thesis.
Short Positions (Inferred / Reported)
- INFY (Infosys) Large short. IT outsourcing firms whose business model (cheap labor arbitrage for IT work) is being automated by Claude Code, Codex, and similar tools. Structural call on software eating the delivery layer.
- SMH (VanEck Semiconductor ETF) Large PUT position as a hedge against broad semi exposure while maintaining conviction in Intel and Broadcom specifically.
04 Mock AI Infrastructure Portfolio
Hypothetical illustrative portfolio applying the framework. Not financial advice.| Bucket | Ticker | Weight | Rationale |
|---|---|---|---|
| LAYER 1: POWER (30%) | |||
| Bloom Energy | BE | 15% | Best-in-class distributed fuel cell generation. $20B backlog. High conviction. |
| EQT Corporation | EQT | 8% | Largest US natural gas producer. Upstream exposure to fuel cell demand. |
| Solaris Energy | SEI | 7% | Modular, distributed power solutions. Same grid-bypass need at smaller scale. |
| LAYER 2: COMPUTE (25%) | |||
| CoreWeave | CRWV | 15% | Dominant GPU-as-a-service neocloud. Multi-year HPC contracts. |
| Applied Digital | APLD | 6% | AI data center hosting. Earlier stage, higher risk, higher upside. |
| WhiteFiber | WFBR | 4% | Smaller HPC build-out. Revenue +64% YoY. |
| LAYER 3: REAL ESTATE (20%) | |||
| Core Scientific | CORZ | 10% | Best-positioned BTC miner pivot. 12-yr CoreWeave HPC contracts. |
| IREN Ltd | IREN | 5% | Miner pivoting to AI/HPC hosting. Expanding footprint. |
| Kilroy Realty | KRC | 5% | Data center REIT. Defensive exposure. |
| LAYER 4: CONNECTIVITY (15%) | |||
| Lumentum | LITE | 8% | Optical components for AI cluster interconnects. |
| Coherent Corp | COHR | 7% | Diversified optical/laser components. |
| LAYER 5: WILDCARD (10%) | |||
| Intel (calls) | INTC | 6% | National security bet. Asymmetric via calls. |
| Tower Semiconductor | TSEM | 4% | Specialty analog/mixed-signal foundry. |
05 Steelman & Counter-Arguments
The Steelman
- Track record. Sequencing is correct. GPU scarcity was real in 2023-2024. He exited before saturation hit NVIDIA in early 2025.
- CapEx is real. Google, Meta, Microsoft, and Amazon collectively announced $650B+ in AI capex. That spend has to go somewhere physical.
- Grid constraint is documented. US grid has not added meaningful baseload capacity in years. AI data centers need 10-50x more power per sq ft than typical commercial real estate.
- Liquor-license hack is elegant. Bitcoin miners already have grid connections, permits, industrial real estate, and power contracts. Repurposing for AI hosting is a 12-24 month shortcut vs 5-year permitting.
- Infosys short is logical. IT outsourcing is directly in the path of agentic AI.
- Revenue exists. Bloom sells into real purchase orders. CoreWeave has multi-year contracts. Cash-flow businesses, not story stocks.
The Counter-Arguments
- Single-thesis concentration risk. Essentially one idea: AI capex keeps growing. If hyperscaler spending slows, every position correlates to the downside simultaneously.
- The thesis may be priced in. Bloom's market cap reflects the AI power narrative. CoreWeave IPO'd at a steep multiple. A copycat portfolio in March 2026 enters at a different risk/reward.
- AGI timeline risk. AGI by 2027 underpins everything. Polymarket puts probability of OpenAI announcing AGI before 2027 at ~13%.
- Regulatory and permitting risk. Permitting for new power generation, data center zoning, and potential DOE intervention could introduce multi-year delays.
- Manufacturing execution risk (Bloom). Concentrated at ~16% of reported equity. Fuel cell tech has execution risk at scale.
- Manager track record is short. 24 years old, 12 months of track record. Performance could be macro tailwind.
- Incomplete picture from 13F. We do not know the full derivatives book, leverage levels, or international positions.
06 What Aschenbrenner May Have Missed
- Healthcare and biotech AI. AI deployed in life sciences, drug discovery, materials research. Not captured by infrastructure plays.
- Semiconductor capital equipment. ASML and other lithography companies remain long-term critical.
- Water / cooling infrastructure. AI data centers consume enormous amounts of cooling water. Water rights and dry-cooling solutions could be Phase 4 bottlenecks.
- International exposure. European energy transition, Southeast Asian data center buildouts, UAE/Saudi AI megaprojects.
- AI cybersecurity layer. Security vendors targeting AI/cloud workloads are an adjacent bet.
- Consumer/enterprise AI adopters. Labor-intensive industries that can do more with fewer people become demand-side plays.
07 Executive Summary
| Core Thesis | GPUs are priced in. Energy + physical data center capacity is the next binding constraint. Buy the bottleneck, not the hype. |
| Key Positions | Bloom Energy (power), CoreWeave (neocloud), Core Scientific (miner to AI hosting), Lumentum (optics), Intel calls (national security). Short: Infosys (AI automation of IT outsourcing). |
| Bull Case | $650B+ hyperscaler capex is real and accelerating. Grid is genuinely constrained. Miner-to-AI hosting hack is durable. |
| Bear Case | Single-thesis concentration. Positions may already reflect the narrative. AGI timeline slip crashes the whole model. |
| Gaps | No exposure to water/cooling, international AI buildouts, healthcare AI, or cybersecurity for AI systems. |
| Bottom Line | Thesis is structurally sound and validated in real time. The question is whether entry at current valuations still offers asymmetric upside. |