Leopold Aschenbrenner, the former OpenAI researcher who scaled Situational Awareness Fund from $250 million to $13.7 billion in 24 months, has filed a 13F report with the SEC revealing a historic strategic pivot. For the first time in the fund's existence, short exposure has surpassed long exposure, totaling $8 billion across the semiconductor supply chain. This aggressive positioning targets industry giants including NVIDIA, AMD, Broadcom, ASML, and Micron, representing a bet size 40 times the fund's net asset value from 18 months ago. The move signals a definitive shift in the AI investment thesis, moving away from pure compute design toward the physical constraints of power and memory infrastructure.

The core logic driving this reversal is the identification of a new bottleneck: the electron rather than the chip. Data compiled by Woofun AI indicates that while GPU supply remains robust, the infrastructure required to deploy them at scale is critically insufficient. Aschenbrenner notes that major AI labs like Anthropic are seeking partnerships with competitors such as SpaceX for computing power, not due to a lack of silicon, but because of the inability to secure adequate power and cooling. Consequently, the fund has established a $1.9 billion short position in NVIDIA, combining direct put options with exposure through the VanEck Semiconductor ETF (SMH), where NVIDIA holds a 20% weight.

On the long side, the portfolio aggressively targets entities controlling energy access and grid integration. CoreWeave remains a primary holding, valued not merely for its GPU clusters but for its secured permits and capacity within existing power grids. Bloom Energy, a producer of portable gas turbines capable of airlifting power to data centers, saw its position reduced by $1 billion after a rapid rise from $800 million to $2.5 billion, yet the fund retains over $1 billion in exposure.

Furthermore, the fund has expanded positions in Bitcoin mining firms including CleanSpark, Riot Platforms, Applied Digital, and IREN. These companies are viewed as pre-built infrastructure assets, with U.S. miners expected to bring 30 GW of power capacity online this year—a figure comparable to the combined data center power plans of Microsoft, Google, Amazon, and Meta.

Memory infrastructure represents another critical long pillar, specifically targeting SanDisk, which has surged approximately 40,000% in the past year. Despite the crowded nature of this trade, the fund maintains a bullish stance on NAND flash memory, essential for AI model context recall. Market data shows memory prices have risen 300% to 500% over the last nine months, with capacity booked through the end of 2027. Woofun AI analysis suggests this scarcity drives the thesis that capital will flow from the commoditized design layer to the constrained storage and power layers. The fund also holds a complex collar trade structure on Micron, pairing puts with calls to hedge directional uncertainty while capturing premium differentials.

The short thesis extends beyond NVIDIA to include Intel, a former top performer for the fund, and Corning, a fiber optic glass manufacturer. This broad shorting of the semiconductor ecosystem implies a belief that current valuations price in a world where every chipmaker equally benefits from AI demand, a scenario Aschenbrenner disputes. He argues that the design layer is overcrowded, with companies like NVIDIA, Broadcom, and AMD facing potential margin compression as custom chips from hyperscalers emerge.

However, risks remain significant; NVIDIA's CUDA software lock-in and the high secondary market value of older GPUs suggest the moat may be stronger than anticipated. If NVIDIA's earnings guidance on May 20 exceeds $78 billion, the $1.9 billion short position could face substantial losses.

Strategic implications for the broader market suggest a rotation from software-defined compute to physical-world constraints. The fund's approach mirrors a directional pair trade: shorting silicon while longing electricity. This strategy acknowledges that while AI demand persists, the limiting factor has shifted to the ability to manufacture, build, and power the necessary facilities. Retail investors are advised caution, as the S&P 500's recent gains have been heavily concentrated in the Mag 7, creating an overcrowded trade environment. Woofun AI observes that while the probability of an immediate AI bubble burst remains low at 24% according to Polymarket, the volatility implied by this 13F filing warrants a reassessment of portfolio concentration in pure-play semiconductor stocks.