Leopold Aschenbrenner, the former OpenAI researcher who departed the company over classified information disputes, has positioned his hedge fund against Nvidia and several leading AI chipmakers according to recent regulatory disclosures. The filing marks one of the first significant public moves from a fund manager whose previous employer sits at the center of the AI infrastructure buildout those same chip companies supply.
The fund's bearish positioning targets the sustained rally in semiconductor equities that has carried Nvidia shares up more than 170% over the past twelve months and lifted the broader VanEck Semiconductor ETF by 48% in the same period. Regulatory filings do not disclose position sizing or specific strike prices, but the disclosure threshold suggests material capital allocation against a sector that has absorbed nearly $780 billion in net inflows across equity and derivative instruments since January 2024. Aschenbrenner's fund appears to be betting that current valuations price in growth assumptions the underlying demand cannot support at scale.
The timing matters because the short position arrives as Nvidia's forward price-to-earnings ratio sits at 28.4x, down from a peak of 42.1x in mid-2024 but still elevated relative to the 19.7x historical average for large-cap semiconductor manufacturers. The company reports earnings in three weeks. Consensus estimates call for $0.84 per share on revenue of $37.9 billion, representing sequential growth of 6.8% in a quarter where hyperscaler capital expenditure is expected to decelerate for the first time in seven quarters. If Aschenbrenner's thesis holds, the market has mispriced the gap between AI infrastructure investment announcements and actual deployment timelines—a gap his former colleagues at OpenAI would understand better than most public market participants.
What makes this positioning noteworthy is the source. Aschenbrenner spent two years inside OpenAI's superalignment team before his exit in April 2024, giving him direct visibility into how leading AI labs consume compute resources and how those consumption patterns translate into chip orders. His public writings since departure have focused on the distinction between research-phase AI development and production-scale inference workloads, arguing that the latter requires different silicon architectures than the training-focused H100 and H200 chips that drove Nvidia's recent margin expansion. If production inference shifts toward custom ASICs or alternative architectures, the current Nvidia multiple assumes a market structure that may not persist.
Allocators should watch three developments over the next ninety days. First, Nvidia's January 26 earnings call for any commentary on inference-specific chip demand versus training chip demand—management has historically bundled these figures. Second, capital expenditure guidance from Microsoft, Meta, and Alphabet in their respective January and February earnings, specifically any line-item separation between data center expansion and AI-specific infrastructure. Third, any material shifts in Broadcom's custom ASIC order book, which would signal hyperscalers moving workloads away from general-purpose GPUs faster than the market expects.
The filing arrives the same week Taiwan Semiconductor reported December revenue up 58% year-over-year, confirming strong chip demand at the foundry level but offering no breakdown between AI and non-AI wafer starts—precisely the data point that would validate or invalidate Aschenbrenner's position.