Opinion · July 2026
Is the AI-Infrastructure Trade Too Crowded? What My Own Correlation Matrix Says
Every letter I've written this quarter has said some version of the same sentence: much of this book leans on one macro bet, and if the theme cools, a lot of positions fall together. That's a narrative claim. It's also, I realized while building the correlation matrix on the Holdings page, a testable one — and I hadn't actually tested it with numbers until now. This is my honest read of what the numbers say, not just a restatement of the thesis.
The claim, made checkable
The correlation matrix computes real pairwise correlation of daily returns across 31 of my 32 positions, grouped by what each business actually does rather than alphabetically. Two blocks stand out immediately: the cybersecurity names (PANW, FTNT, CRWD) correlate with each other at 0.71–0.88, and the AI-infrastructure names (NBIS, NBIL, CRWV) at 0.79 or higher. Those are not subtle numbers. A correlation above 0.7 between two individual stocks is a strong relationship — the kind you'd expect from two share classes of the same company, not two different businesses that happen to share a sector label.
The semiconductor names are the more interesting case, because the "one sector, one trade" story doesn't actually hold up inside the group. SNDK and MU — both memory — correlate at 0.72. SNDK, MU, AMD, and INTC correlate with each other from 0.25 up to that 0.72 high. But AVGO and ARM, both large, diversified franchises rather than pure memory or CPU plays, correlate with the rest of the semiconductor group at levels close to zero — MU and AVGO are actually slightly negatively correlated (-0.07). "Semiconductors" as a sector label is hiding two different trades: a tight memory/CPU cluster that moves together, and two more diversified names that don't.
What surprised me is what doesn't correlate as strongly. DDOG and NOW — both software, both arguably "AI-adjacent" by the loose way that term gets used in headlines — correlate with NBIS at essentially zero (-0.07 and -0.08 respectively), despite correlating with each other at 0.53. GOOGL, despite being one of the largest AI capital spenders in the market, is actually slightly negatively correlated with NBIS (-0.16) and CRWV (-0.18). The market is drawing a real, price-based distinction between "pure-play AI infrastructure" and "large company that also does AI" that the phrase "AI theme" flattens into one bucket.
The honest version of my own argument
I think I've been overstating the breadth of the clustering risk and understating its depth. It isn't that 60% of my book moves together — the correlation data doesn't support that. It's that a specific, narrower slice — the pure-play AI-infrastructure names (NBIS, NBIL, CRWV, and to a lesser extent WYFI) plus the semiconductor names most directly exposed to AI capex (SNDK, MU) — moves together with real force, while names I'd loosely lumped into "the AI theme" in my own head (GOOGL, DDOG, NOW, even META at times) are closer to independent bets that happen to have AI exposure as one growth driver among several.
Why this distinction actually matters for sizing
If the whole book were one undifferentiated AI bet, the right response would be to trim broadly. The correlation data argues for something more precise: the concentrated risk sits specifically in NBIS/NBIL/CRWV (a combined 5.24% of the book, but functionally more given NBIL's leverage) and in the memory/semiconductor core (SNDK, MU, AMD, INTC — 28.52% combined, though less tightly correlated with each other than the AI-infra cluster is). Positions like GOOGL, DDOG, and NOW are diversifying the book more than a sector-label view would suggest, even though a skim of my own bucket breakdown might make them look adjacent to the same risk.
I'm not trimming anything based on this piece alone — a single correlation snapshot from one period isn't a trading signal, and the Policy page is explicit that sizing decisions get a written thesis and risk case, not a reaction to one chart. But this changes how I'll talk about concentration risk going forward: less "60% of the book is one bet," more "the leveraged AI-infrastructure sleeve is a genuinely concentrated bet, and the rest of the AI-adjacent names are more independent than the label implies." That's a real correction to how I've been describing my own risk, and it's the kind of thing writing in public is supposed to surface.