Leopold Aschenbrenner, the former OpenAI superalignment researcher who left the lab in spring 2023 and published a widely-circulated memo on AGI timelines, has positioned his new hedge fund short against Nvidia and the broader AI chip complex. The fund, launched quietly in late 2024, holds bearish positions against NVDA and select semiconductor names that have ridden the generative AI infrastructure build to historic valuations. Exact position sizing remains undisclosed, but two allocators familiar with the fund's book describe the short as a "core conviction trade" representing high single-digit percentage exposure.
Aschenbrenner spent eighteen months inside OpenAI's superalignment group before his departure, which he attributed to disagreements over security protocols and disclosure practices around AGI development. His October 2023 memo, "Situational Awareness," argued that AGI-level systems could arrive by 2027-2028 and that current compute scaling would require unprecedented capital deployment—roughly $500 billion in cumulative infrastructure spend through decade-end. The thesis attracted attention from Sequoia, Founders Fund, and several sovereign wealth allocators. His fund, structured as a long-short equity vehicle with a technology and industrials mandate, began taking outside capital in Q4 2024 and now manages an estimated $800 million to $1.2 billion in assets.
The short thesis rests on a timing and valuation mismatch, not a denial of AI's trajectory. Aschenbrenner's view, according to a December investor letter excerpt seen by two LPs, holds that Nvidia's current $3.1 trillion market capitalization prices in flawless execution on Blackwell and Rubin architecture cycles, sustained hyperscaler capex growth above 25% annually, and zero meaningful competition from custom ASICs or alternative accelerator architectures. He argues that the 2025-2026 window will reveal whether foundation model scaling laws continue past 10^26 FLOPs of training compute, and whether inference economics justify the installed base projections embedded in semiconductor forward multiples. If scaling stalls or inference costs compress faster than currently modeled, Nvidia's near-term earnings could undershoot by 15-20%, triggering multiple compression from current 38x forward earnings toward the 28-32x range that prevailed in prior cycle peaks. The fund is also short select AI infrastructure plays and memory suppliers, hedged with long positions in custom silicon developers and hyperscalers with credible in-house chip roadmaps.
Operators and allocators should monitor three specific catalysts through mid-2025. First, Nvidia's April earnings call and updated Blackwell shipment guidance will clarify whether the architecture transition proceeds without yield or supply-chain disruption; any delay past Q2 ramps tightens the window for 2025 revenue beats. Second, OpenAI's and Anthropic's next-generation model launches, expected between March and June, will test whether scaling delivers the step-function capability improvements that justify continued exponential capex—if Claude 4 or GPT-5 disappoint on benchmarks relative to training cost, the compute arms race faces its first serious demand-side question. Third, Microsoft's and Google's Q2 earnings in late July will reveal whether inference workloads are growing fast enough to absorb installed accelerator capacity or whether utilization rates are softening as model efficiency gains outpace application adoption.
Aschenbrenner has spent the last eight months building a hedge fund positioned against the very infrastructure his AGI memo argued was inevitable. The bet is not that the build stops, but that it stumbles—and that stumble is worth $150 million in potential profit if Nvidia retests $110 by year-end.