Leopold Aschenbrenner's hedge fund disclosed short positions against Nvidia, Oracle, and AMD in its latest 13F filing, marking one of the few public bearish stances on AI infrastructure from someone who spent 2023 inside OpenAI's supercomputing org. The positions appeared in quarterly disclosures covering the period ending December 31, with notional exposure crossing $142 million across the three names. Aschenbrenner left OpenAI in mid-2023 after writing internal memos on compute scaling limits, then launched the fund in early 2024.

The short book targets the companies supplying picks and shovels to the AI boom. Nvidia closed the quarter at $140 per share before the filing surfaced; Oracle sat at $178; AMD traded near $121. The fund's thesis, according to a person familiar with the strategy, centers on a mismatch between current chip valuations and the utilization rates Aschenbrenner observed during his time at OpenAI. He watched foundation model labs run $50 million training runs, then leave hardware idle for months during fine-tuning and evaluation phases. The gap between peak throughput marketing and real-world duty cycles, in his view, is 40 to 60 percentage points depending on the customer.

This matters because Aschenbrenner is not a generalist macro short. He spent eighteen months inside the architecture that defines AI demand, wrote the internal scaling memos that circulated to Sam Altman's level, and now runs capital with access to ex-OpenAI engineers who know which GPU clusters are actually running hot. His short thesis is not about AI hype dying—it is about the hardware layer overbuilding for a demand curve that scales in bursts, not straight lines. If he is correct, the next twelve months will show utilization rates declining even as new chip orders continue, a pattern that breaks the semiconductor bulls' assumption that AI compute is a pure exponential.

The filing also disclosed long positions in smaller infrastructure plays, including $18 million in a data center REIT and stakes in two edge computing firms. The structure suggests Aschenbrenner believes the value accrues to real estate and power, not silicon. Worth noting: he filed a separate advisory client letter in February arguing that AI model training will shift toward smaller, task-specific architectures by late 2025, which would cut per-model GPU requirements by 70 percent compared to 2023 foundation models. If that shift happens on his timeline, Oracle's cloud GPU offerings and AMD's MI300 ramp both face demand cliffs that current Street estimates do not model.

Operators should track two signals in the next ninety days. First, whether Nvidia's April earnings call shows data center revenue growth decelerating from the 206 percent year-over-year pace reported in January—any slowdown to 150 percent or below confirms softer utilization is bleeding into orders. Second, whether hyperscalers' capex guidance for 2025 starts carving out "efficiency initiatives" or "rightsizing" language, which would be the first public acknowledgment that the 2024 hardware binge overshot near-term demand.

Aschenbrenner's fund now manages $340 million, up from a $120 million launch last spring. The shorts are sized at roughly 15 percent of NAV each, large enough to matter but not career-ending if the chip rally extends another quarter. He has not given an interview since leaving OpenAI.