Leopold Aschenbrenner, the former OpenAI safety researcher who left the lab in April 2024 and launched a hedge fund by July, filed his first 13F with the SEC last week. The portfolio holds $47 million across six public equity positions, with three of them Bitcoin mining operators: MARA Holdings, Riot Platforms, and Core Scientific. The balance of the book tilts toward data center REITs and utility-scale power infrastructure. No software. No hyperscalers. No model labs.
The filing confirms a thesis Aschenbrenner outlined in essays and interviews throughout 2024: that the constraint on frontier AI is not talent or capital, but physical infrastructure—specifically, the ability to source, site, and operate gigawatt-scale power for training clusters. Bitcoin miners, uniquely, hold permitted power contracts, substation access, and operational know-how for high-density electrical load in rural jurisdictions where utilities have headroom. MARA operates facilities in Texas and North Dakota with 1.1 gigawatts of contracted capacity. Riot controls a 400-megawatt site outside Austin. Core Scientific has pivoted from bankruptcy to hosting arrangements with hyperscalers, converting mining rigs into AI inference pods. Aschenbrenner is long the optionality embedded in those power purchase agreements.
The intelligence-desk read is that Aschenbrenner believes the next phase of the AI capital cycle is a land grab for energy, not for chips or models. His fund launched during the narrowest equity market in two decades, while power infrastructure remains under-indexed and misunderstood. Bitcoin miners trade at steep discounts to book value because the market still prices them as commodity producers tied to coin volatility. Aschenbrenner is repricing them as infrastructure plays with embedded real options on AI compute demand. The 13F suggests he views the conversion value of mining sites into AI hosting as materially mispriced. It also signals that he expects hyperscalers to face permitting bottlenecks faster than the market anticipates, likely in the 2025-2026 window when announced cluster builds reach the interconnection queue.
Operators and allocators should track three follow-on events. First, whether MARA or Riot announce hosting contracts with hyperscalers or sovereign AI buyers in Q1 2025—those would validate the infrastructure arbitrage thesis with public pricing. Second, any moves by Aschenbrenner to raise a second vehicle or pivot into private infrastructure, which would confirm the public positions are scouts for a larger deployment. Third, whether other AI-native allocators—particularly those with lab exposure—begin rotating into the same power layer. The market currently treats AI infrastructure as a chip story. If it reprices as an energy story, the capital flows shift.
The filing lands while Microsoft, Google, and Meta each plan to double their capex in 2025, with the majority earmarked for data center construction. The permitting layer is already failing them. Aschenbrenner is long the choke point.