Regarded as one of the most aggressive AI investors globally, Leopold Aschenbrenner is shorting NVIDIA, ASML, and Oracle in the public markets with a nominal position of about $9 billion, while shifting funds into deeper AI infrastructure and model assets.

His positioning logic can be broken down into three directions, one judgment:
Short chips. His hedge fund, Situational Awareness LP, has built approximately $8.46 billion in put options on the semiconductor sector, covering NVIDIA, Broadcom, AMD, Oracle, Micron, TSMC, and others. He mentioned in a paper that by the end of 2025 or early 2026, the market will have essentially fully reflected the value of GPUs. Chips are the most easily commoditized layer in the entire AI stack, with Amazon's Trainium, Google's TPU, and Cerebras' new architecture all competing in this space, which could compress profit margins.
Long AI labs. Anthropic is the fund’s largest single holding, accounting for about 20%. In his framework, labs are "the demand-owning layer." Additionally, he has a heavy position of around $300 million in CoreWeave, a new cloud service provider that specializes in helping AI labs deploy GPUs, manage power supply, and cooling.
Long infrastructure. Many AI labs face the problem that while GPU capacity is sufficient, the existing power grid is designed for humans and cannot handle AI’s electricity demands. He has allocated 20% of his position into Bloom Energy, which produces fuel cells, and has also bought into a batch of companies transitioning from Bitcoin mining to AI data centers: CleanSpark, Riot Platforms, Applied Digital, IREN. The bottleneck has shifted from "who makes chips" to "where to put the chips."
For the first time in history, short positions exceed long positions, with an $8 billion short exposure, which is 40 times the fund’s net worth 18 months ago.
He is not betting that AI will fail; he is betting that profits will shift from chip manufacturing to demand-driven labs and infrastructure bottlenecks.