2025 was a great year for me in terms of investment performance: continued bets on neocloud, gas turbines, and HBM through pure supply-demand analysis and physical constraints; a deep understanding of how data centers are actually built that led to several successful shorts, including one stock that dropped nearly 50% within two weeks; a strong short in SaaS after Dario mentioned they wanted to internalize more capabilities.
On paper, the results looked great. The portfolio significantly outperformed the index and likely outperformed more than 90% of funds that year. But in reality, 2025 was also the year I realized how much I had still missed.
The number I couldn't ignore
10x return. That's the number one of my friends' funds did this year. No leverage, concentrated in around 10 names, slightly hedged.
Normally, as a well-trained Wall Street type investor, the instinct is always: what's the risk-adjusted return? What's the hedge? What multiple am I paying? How much downside is protected? That framework worked very well for a long time. If you caught Nvidia in 2023, energy and power in 2024, then HBM and storage in 2025, you already generated incredible returns.
Sometimes sophisticated investing becomes a psychological shield. You hide behind valuation discipline because it sounds intellectually rigorous — while the industry is changing faster than your framework allows.
When the framework itself needs to evolve
Buffett was probably the person who influenced my investment thinking the most. Early Buffett was much closer to a pure Benjamin Graham-style "cigar butt" investor: buy statistically cheap companies, low P/B, low PE, buy $1 for 50 cents.
But eventually he realized something important: cheap businesses are often cheap for a reason. Terrible businesses consume capital. Weak businesses never compound. Then came the famous shift:
"It is better to buy a wonderful business at a fair price than a fair business at a wonderful price."
I think something similar is happening right now in AI infrastructure. The framework is shifting — and investors still using old lenses are going to keep missing it.
Early on the scarcity. Late on the meaning.
For years, my alpha came from identifying where the real bottlenecks were forming before consensus fully recognized them. I remember being involved around power generation projects in Texas when Elon bought 70% of small gas generators and we were struggling to secure even one piece of equipment ourselves. The scarcity was physically visible on the ground long before it appeared in financial models — it took another six months before capital finally rotated into names like GE Vernova and Caterpillar.
The same pattern repeated in HBM. By the time "memory wall" became a mainstream discussion, bandwidth scarcity had already started reshaping customer behavior.
Even though I was early on the scarcity itself, I was still mentally categorizing these companies using old economic frameworks. I sold too early. I missed part of the HBM move. Not because I didn't see the demand — but because I still viewed memory through a traditional lens: eventual oversupply, ASP collapse, margin normalization, commodity cycle.
Price per bit vs. price per bandwidth
Historically, memory was always analyzed through a storage lens: price per bit, commodity cycles, eventual oversupply, margin collapse. The assumption was simple — memory always commoditizes eventually.
But AI workloads changed the customer utility function entirely. Customers no longer view memory through a storage lens. They view it through a compute lens. Even as price per bit goes up, price per bandwidth is falling. Customers are getting more compute value per dollar — and once that happens, they become willing not only to pay more, but to pay upfront.
AI did not just increase demand. It changed the economic identity of infrastructure components.
Identifying bottlenecks is no longer enough
The biggest lesson I learned in 2025 is that identifying bottlenecks is no longer sufficient. The market eventually recognizes every visible scarcity.
The harder question is whether AI changes the economic structure behind that scarcity — and whether investors are still using old frameworks to value a fundamentally new kind of infrastructure.
For years, my edge came from identifying physical bottlenecks before consensus fully recognized them. But AI increasingly forces a second layer of analysis: not just where scarcity exists, but how intelligence workloads reshape the economic meaning of the infrastructure itself.
That was the framework shift AI forced me to confront in 2025.