# The $5.5B Bet on AGI Infrastructure Source: https://x.com/Eli5defi/status/2044346674269888648 Kiwi News: https://news.kiwistand.com/stories/The-dollar55B-Bet-on-AGI-Infrastructure?index=0x69e1ecba9a1b3379842d30bacb492d4fc9a68089a84bf4973c0e51db03e13199ad6f8aef Submitted by rvolz.eth on 2026-04-17 UTC, 1 upvote, 0 comments ## Summary Leopold Aschenbrenner, a former OpenAI researcher who predicted AGI by 2027 in his 165-page essay Situational Awareness, now runs a hedge fund whose Q4 2025 portfolio is concentrated in power and physical infrastructure rather than chips. The article reports the fund reached $5.5 billion in gross equity exposure within 18 months and returned approximately +47% net of fees in H1 2025, against +6% for the S&P. Its top five positions make up 65.3% of exposure, led by Bloom Energy at 20.6%, and the fund exited Nvidia, Broadcom and TSMC hedges and longs during 2025. ## Article The 24-year-old former OpenAI researcher who predicted AGI by 2027 now runs one of 2025's best-performing hedge funds. His portfolio tells a different story than his essay. @leopoldasch is 24. He graduated Columbia as valedictorian at 19. He worked on OpenAI's superalignment team until April 2024, when he was fired after writing a memo to the board arguing that OpenAI's security was insufficient to protect model weights from foreign actors. Two months later he published Situational Awareness, a 165-page essay arguing AGI arrives by 2027. Then he started a hedge fund. In 18 months he went from zero to $5.5 billion in gross equity exposure. H1 2025 returns: approximately +47% net of fees, against +6% for the S&P and +7% for the average tech hedge fund. His backers are Patrick and John Collison (Stripe), Nat Friedman (ex-GitHub), and Daniel Gross. His research director previously worked at Peter Thiel's Clarium Capital. His fiancée is chief of staff to Anthropic CEO Dario Amodei. The interesting part isn't his biography. It's that his Q4 2025 13F reveals a portfolio that's quietly walked away from the consensus AI trade, and it's worth understanding why. The thesis, in plain English Leopold's argument rests on three things compounding at once: Compute. Training runs get larger every year. GPT-2 to GPT-4 represented roughly a 10,000x jump. The trajectory continues. By 2027, another 100x on top. Algorithms. We keep doing more with less. GPT-4o is 6x cheaper than original GPT-4. Gemini 1.5 Flash is 85x cheaper. This trend holds. Unhobbling. Scaffolding, tools, chain-of-thought, longer context. An RLHF'd small model can match a raw model 100x larger. GPT-3.5 with proper scaffolding beats vanilla GPT-4 on coding. Stack them together and you get roughly 5 orders of magnitude of effective compute by 2027. That's 100,000x. His framing: "It's completely wrong to just imagine 'GPT-6 ChatGPT.'" Once AGI arrives, it automates AI research itself. A GPU fleet running the equivalent of "100 million automated Alec Radfords" at 100x human speed compresses a decade of algorithmic progress into a year. Superintelligence by 2030. Economic growth at 30%+ annually. You don't need to accept all of this. The investment-relevant point is that if even half of it is directionally correct, the capex required to build the physical infrastructure is unprecedented. Where the thesis gets specific Leopold's scaling table is the part that matters: 2022: ~10K H100-equivalent GPUs, ~$500M, ~10 MW 2024: ~100K GPUs, billions, ~100 MW 2026: ~1M GPUs, tens of billions, ~1 GW (Hoover Dam-scale) 2028: ~10M GPUs, hundreds of billions, ~10 GW (small US state) 2030: ~100M GPUs, $1T+, ~100 GW (>20% of current US electricity) Industry capex projection: $150B/yr today, ~$2T/yr by 2028, ~$8T/yr by 2030. And the line that explains his portfolio: "Power is the binding constraint, not chips or investment." Nuclear takes too long. Solar and wind don't concentrate well for this use case. Natural gas from Pennsylvania shale is buildable in 1-2 years with enough capacity for 200 GW continuous. He argues every available voltage transformer is being contested right now. He rates AI lab security at "level 0" (specifically Google). He expects the US government to eventually nationalize frontier AI into a Manhattan Project-style consortium. Take the thesis seriously and the chokepoints aren't models or chips. They're electrons, transformers, optical networking, cooling, and physical data center buildout. The Q4 2025 portfolio The fund files 13Fs with the SEC. Here's the verified book in descending weight: Top 5 positions = 65.3% of exposure (with Google Finance Link): Bloom Energy (BE) — 20.6%, ~$875M. Fuel cells deployed next to data centers. Power without waiting for the grid. CoreWeave (CRWV) — 18.2%, calls. Nvidia-backed GPU cloud with a $12-16B OpenAI contract. Lumentum (LITE) — 11.2%. Optical transceivers that connect millions of GPUs together. Core Scientific (CORZ) — 9.8%, plus a 13D filing disclosing 9.4% ownership of the entire company. Bitcoin miner pivoting to AI/HPC hosting. IREN — 7.7%. Another Bitcoin miner pivoting to AI. The rest: Applied Digital (APLD, 6.5%), SanDisk (SNDK, 5.9%), Cipher Mining (CIFR, 3.6%), EQT natural gas (EQT, 3.1%), Coherent photonics (COHR, 2.1%), Tower Semiconductor (TSEM, 2.0%), Kilroy Realty data centers (KRC, 1.2%), plus five more Bitcoin miners (RIOT, HUT, BTDR, CLSK, BITF), Babcock & Wilcox (BW), Liberty Energy (LBRT), and ProPetro (PUMP) for oilfield services. One notable short: puts on Infosys (INFY). The thesis is that AI coding tools disrupt traditional IT outsourcing. The pivot worth understanding In early 2025, the portfolio resembled a conventional AI bull book. Intel (INTC) calls, Broadcom (AVGO), TSMC (TSM), Nvidia (NVDA) exposure, VanEck Semiconductor ETF, Vistra (VST) and Constellation Energy (CEG) for power. By Q4 2025, he had exited: → Nvidia puts ($299M unwound) → Broadcom puts ($76M) → TSMC puts ($75M) → Micron (MU) puts ($50M) → Vistra ($252M) → Constellation Energy → VanEck Semiconductor ETF puts ($196M) He took off chip hedges and chip longs simultaneously. He exited nuclear utilities. He concentrated into fuel cells, natural gas, optical networking, and Bitcoin miners converting to AI hosting. The read: he believes the chip supply bottleneck is resolving while the power and physical infrastructure bottleneck is tightening. Most investors are still fighting the last war around Nvidia earnings prints. Leopold's portfolio is positioned for the scarcity migrating downstream to whoever can deliver megawatts and optical bandwidth. Right or wrong, that's the call. Thematic map for readers building their own framework If you want to build around his thesis without cloning his book, here's the skeleton (click underlined for direct Google Finance link): Power (the actual bottleneck): GE Vernova (GEV), Siemens Energy (ENR), Bloom Energy (BE), EQT (EQT) and gas E&Ps, Constellation (CEG), Vistra (VST), Talen (TLN) for nuclear, Cameco (CCJ) and Kazatomprom (KAP) for uranium, Oklo (OKLO) and NuScale (SMR) for SMR exposure. Grid and transformers: Eaton (ETN), Quanta Services (PWR), Hubbell (HUBB), Schneider Electric (SU), Hitachi (6501), Prysmian (PRY). Cooling: Vertiv (VRT), nVent (NVT), Munters (MTRS), Super Micro (SMCI). Optical networking: Lumentum (LITE), Coherent (COHR), Arista (ANET), Ciena (CIEN), Cisco (CSCO). Data center real estate: Equinix (EQIX), Digital Realty (DLR), Iron Mountain (IRM), Kilroy (KRC). HBM memory: SK Hynix (000660), Samsung (005930), Micron (MU). Fabs and equipment: TSMC (TSM), Samsung (005930), Intel (INTC), ASML (ASML), Applied Materials (AMAT), Lam Research (LRCX), KLA (KLAC), Tokyo Electron (8035), Advantest (6857), Disco Corp (6146), ASM International (ASM), BE Semiconductor (BESI). Private AI through public proxies: Microsoft (MSFT) for OpenAI (~27% stake, ~$135B embedded). Google (GOOGL) for Anthropic (~14%). Amazon (AMZN) for Anthropic (~7.8%). ASML (ASML) for Mistral (11%). SoftBank (9984) for OpenAI (~$30B invested). Destiny Tech100 (DXYZ) and ARK Venture Fund (ARKVX, interval fund) for diversified private AI if you accept the premium. The risks Timeline dependency. The capex thesis hinges on AGI-ish capability arriving roughly on schedule. If GPT-6 disappoints, the power demand curve deflates and infrastructure names derate. Leverage. $5.5B of 13F exposure against approximately $1.5B in AUM implies ~3.5x leverage. Small-cap miners and neoclouds with thin float are liquidity-sensitive on the way down. Q4 2025 is estimated at -21%, which is what that leverage looks like in a bad quarter. Concentration. Top 5 positions make up 65% of the book. Bloom Energy alone is 20%. Single-name risk is material. Geopolitics. His thesis explicitly warns against Middle Eastern AI infrastructure, but G42 and MGX are buying anyway. Taiwan disruption breaks the supply chain story regardless of his framing. Crowded pivot trade. Seven Bitcoin miners pivoting to AI hosting is a real story, but also a crowded one. The thesis works if AI hosting demand is durable. It unwinds fast if it's a narrative trade. Regulatory risk. If "The Project" materializes the way he imagines, it involves government coordination and potentially forced consolidation of AI labs. Bullish for defense primes, bearish for speculative AI cloud positioning. You can check the complete dashboard above here Why this matters even if he's wrong Most AI analysis gives you vibes. Leopold provides a falsifiable thesis, a dated timeline, a capex table with specific power numbers, and a portfolio that reflects stated beliefs. If scaling continues, the framework gets vindicated and infrastructure names rerate. If scaling plateaus, markets will know by 2027. The most useful takeaway isn't the specific picks. It's the observation that few investors are pricing the gap between AI demand curves and physical infrastructure capacity. That gap is the center of his thesis. One additional data point: he has invested almost his entire net worth in the fund. Most AI commentators have not. The chip cycle had its moment in 2023-2024. The electron cycle is having its moment now. Whether the intelligence explosion cycle gets its moment in 2027, or the whole framework unwinds before then, is the open question. NFA, DYOR. Check My Other Channels: Substack →https://eli5defi.substack.com Telegram →https://t.me/eli5definews Disclaimer: This information is for educational purposes only and does not constitute professional financial or tax advice. Some content may be developed in collaboration with third parties, and we may hold positions in the assets mentioned. We strongly recommend conducting independent research and consulting with a qualified professional before making any financial or tax-related decisions.