LeopoldTracker has released Leopold’s latest portfolio holdings. First, let’s clarify the context. These two updates from LeopoldTracker are the most recent compilations available, but the publicly disclosed stock weights are based on Q2 13F filings as of June 30. After the July liquidation, Situational Awareness sold the majority of its public equity positions to Citadel. Therefore, it is inaccurate to claim that Sandisk is still at 28% or Micron at 27.5% today. However, when we examine the last complete portfolio before the liquidation alongside subsequent investments made afterward, a clear trend emerges: Leopold is moving upstream along the AI supply chain, specifically targeting bottlenecks that could be suddenly overwhelmed by compute expansion. The most extreme allocation in Q2 was in memory: Sandisk at 28%, Micron at 27.5%—together accounting for over 55% of the entire public equity portfolio. Micron increased its position by more than 277x in a single quarter; Sandisk increased by 119%. Next come Bloom Energy at 9.4%, TSMC at 6.3%, Nebius at 6.1%, followed by CoreWeave, Core Scientific, Applied Digital, IREN, and Riot. This layer is straightforward to understand: HBM/NAND, wafer manufacturing, GPU cloud, data centers, power. But what’s more compelling are his smaller positions—those under 1%—the moonshots. CleanSpark, Bitdeer, Hive—these were originally mining companies. With the rise of AI data centers, the market is now revaluing their existing access to power, land, and infrastructure. Solaris Energy Infrastructure focuses on distributed energy. Babcock & Wilcox manufactures boilers and power generation equipment. ProPetro provides oilfield services. Vishay produces resistors, capacitors, power semiconductors, and other fundamental electronic components. Even T1 Energy, a solar panel manufacturer, has entered the mix. The further down the list you go, the less these resemble traditional “AI companies.” And that’s precisely what makes this so interesting. Leopold’s current logic can be distilled into one sentence: If AI capital expenditures truly continue scaling toward trillions of dollars over the next few years, what will be the next thing to run out? The initial answer was Nvidia GPUs. Then it became HBM. Then data center racks, transformers, power access, and natural gas turbines. Keep going upstream, and you hit boilers, oilfield services, electrical components, photovoltaic manufacturing—and even the equipment required to produce these chips. This is no longer just an AI trade. More accurately, it’s becoming an AI Scarcity Trade. It’s certainly important who sells the most GPUs—but as capital spending expands, the greatest excess profits may increasingly shift toward segments with the stiffest supply curves, slowest expansion rates, and historically under-invested infrastructure. A GPU can be iterated annually. But a power plant, a transmission line, or a semiconductor equipment supply chain cannot be built that quickly. This is why Leopold’s actions after his liquidation are so telling. In July, his highly leveraged AI positions collapsed by 67%, forcing him to sell most of his public holdings and deleverage. Logically, one would expect him to retreat. Yet in August, he invested $400 million into Source Foundry—bringing his total investment to $500 million. Source Foundry manufactures semiconductor fabrication equipment. He has bypassed Nvidia, TSMC, and memory entirely—and moved directly into the layer occupied by ASML. This investment occurred after the liquidation, making it far more revealing than the June 30 13F filing about what Leopold truly believes today. **He may have lost his leverage—but he has not abandoned his original industrial thesis; in fact, he’s doubling down upstream.** Putting the entire picture together yields this progression: GPU → HBM/NAND → GPU Cloud → Data Centers → Power → Power Generation Equipment → Energy → Electrical Components → Semiconductor Manufacturing Equipment. I’m increasingly convinced that one of the most significant shifts in AI investing over the next two to three years will be the abandonment of the question: “Who is the next Nvidia?” Instead, investors will begin asking: If Nvidia and its peers truly sell this many chips, what in the entire industrial ecosystem will be the first to run out?