Leopold Aschenbrenner, the former OpenAI superalignment researcher who departed in 2024, disclosed bearish positions against Nvidia and Oracle through his hedge fund Situational Awareness in the firm's latest 13F filing. The positions, spread across put options and short equity stakes, total approximately $1.2 billion notional exposure against the sector's dominant infrastructure players. Aschenbrenner launched the fund in mid-2024 after publishing a widely-circulated memo predicting AGI timelines shorter than consensus and criticizing OpenAI's security posture.
The filing shows concentrated short interest in Nvidia at roughly $780 million notional, Oracle at $310 million, and Advanced Micro Devices at $140 million. The fund simultaneously holds long positions in Anthropic-adjacent private credit instruments and select cloud software names, suggesting a thesis that model training costs will compress faster than hardware revenue can adjust. Aschenbrenner has not filed a letter explaining the positions, but his public writing from Q3 2024 argued that algorithmic efficiency gains—specifically test-time compute scaling and sparse expert routing—would reduce incremental capex per unit of model capability by 40-60% before 2026. That thesis now has capital behind it.
The market implication is straightforward: a former insider with access to frontier lab roadmaps is betting that the current AI hardware buildout is mispriced relative to coming efficiency curves. Nvidia's forward P/E sits at 28x on expectations of sustained datacenter GPU demand, while hyperscalers have guided to $200 billion in combined AI capex for 2025. If Aschenbrenner is correct that model labs will extract more capability per dollar of compute—whether through architectural improvements, better data curation, or synthetic data pipelines—then the incremental revenue assumptions embedded in chip stocks are too high. Oracle's position is particularly pointed: the company has pivoted to GPU cloud resale and AI infrastructure partnerships, booking revenue on long-term capacity contracts. A demand slowdown would hit contracted backlog conversion rates and cloud gross margins simultaneously.
The counterargument is that Aschenbrenner left OpenAI before seeing GPT-5 training runs and may be underestimating the compute requirements for post-training reinforcement learning at scale, which remains brutally inefficient. Anthropic's Claude 3 Opus required roughly 3x the training compute of Claude 2, and inference costs for agentic workflows—where models call themselves recursively—are rising faster than training costs are falling. If the next capability jump demands larger context windows and multimodal pre-training, hardware demand could stay elevated regardless of algorithmic gains. The market has priced in a 15-20% annual increase in datacenter GPU shipments through 2026; Aschenbrenner is implicitly betting that figure will miss by half.
Allocators should watch Nvidia's April earnings call for any mention of order push-outs or extended delivery timelines, which would signal cooling demand. Oracle's June quarter results will clarify whether GPU cloud bookings are accelerating or plateauing. The next Anthropic or OpenAI model release—expected between now and September—will settle the compute efficiency question empirically. If the new models show meaningful capability gains without proportional training cost increases, Aschenbrenner's thesis tightens. If they require 10x more compute for 2x better benchmarks, the short unwinds quickly.
The filing also confirms that Situational Awareness is managing institutional capital, not just Aschenbrenner's personal book. The fund's ADV filing from January listed $340 million in regulatory AUM, which means the positions disclosed represent roughly 3.5x leverage or structured derivative exposure. That is a serious bet, not a hedge.