Been diving into Leopold Aschenbrenner's latest moves and honestly, the fuel logic behind his portfolio shift is kind of genius when you actually trace it.

So this 24-year-old's fund just hit $5.5B (started 2024 with $255M, hit $2B in six months, now this). Most people are still obsessing over him exiting Nvidia completely—sold $300M in puts, dumped Broadcom, TSMC, Micron. Everyone's like "why would he abandon chips?" But his reasoning is pretty straightforward: by late 2025/early 2026, GPU value was already fully priced into the market. The real bottleneck shifted.

Here's where it gets interesting. He looked at the actual constraint facing AI labs right now: not GPUs, but power. The existing grid was literally designed for humans, not for data centers running massive training workloads. So instead of chasing saturated chip plays, he pivoted hard into infrastructure.

His biggest bet? Bloom Energy. 20% of the entire portfolio—roughly $855M. They make oxide fuel cells that convert natural gas directly into usable electricity on-site. No grid dependency needed. You just install it next to your data center. Their backlog hit $20B, revenue grew 34% in 2025, they're projecting 40% growth for 2026. Demand completely outpaces supply. He basically found what could be the Nvidia of the energy sector.

But there's more to the fuel logic. He also loaded up CoreWeave ($300M added, total exposure around $800M now)—they handle the actual deployment of GPUs, power infrastructure, cooling, the whole operational stack. Then he started acquiring Bitcoin mining companies. Sounds random until you realize: these companies already own land, power permits, grid access rights. Obtaining those licenses normally takes months or years. He essentially took a shortcut, stripping out the crypto operations and repurposing the infrastructure for AI data centers. It's like taking over a bar that already has its liquor license instead of waiting years for approval.

He even shorted Infosys. Their entire model is cheap offshore labor arbitrage. With Claude Code and GPT-5 level models now handling complex IT work, that narrative collapses. He's betting against an entire business model.

The through-line? He's moved from software and chips into atoms—manufacturing, real estate, energy, permits. Things AI can't build. Look at what Google, Amazon, and Nvidia just pledged: $650B in capex. Energy is the only resource nobody ever has enough of. That's where the real returns are hiding.

Worth watching how this plays out. His track record on identifying bottlenecks before the market prices them in is pretty solid.