Study · Portfolio Construction · July 2026

The Concentration Premium

By Aydin Ali · July 2026 · ~9 min read · Verified data · AEA Holdings correlation matrix

AEA's own Holdings page reports a Herfindahl-Hirschman Index of 493 — textbook unconcentrated, equivalent to an equally-weighted ~20-position book. That number is true and it is also misleading, because it only measures how large each position is, not how similarly they move. This study uses AEA's own published, real pairwise correlation matrix to ask the sharper question: once you account for co-movement, how many genuinely independent bets is this book actually making?

1. The number that looks reassuring

The Herfindahl-Hirschman Index (HHI) is a standard concentration measure borrowed from antitrust economics: sum the square of every position's weight (in percentage points), scale 0–10,000. AEA's HHI of 493 sits well under the conventional 1,500 threshold for "unconcentrated," and the site is upfront about why: no single position dominates the book, not even SNDK at 7.94%, the largest line item. Restated as an effective position count (10,000 ÷ HHI), the book behaves, by this measure, like roughly 20 equally-weighted positions rather than a true 32-way split.

That's a real, correctly computed number. It's also a number that only ever looks at one input: how big each position is. It has no idea whether those 32 positions move independently of each other or in lockstep — and for a book with 28.50% direct semiconductor exposure sitting inside a 60.93% AI-infrastructure theme, that omission is the whole story.

2. What HHI can't see: correlation

AEA's Holdings page separately publishes a full, real pairwise correlation matrix — Pearson correlation of daily returns, year-to-date, across every position. Below is a sample drawn directly from three rows of that published matrix (SNDK, AMD, and MU against all 29 other positions each, 90 real pairs total), split into pairs within the semiconductor/AI-infrastructure theme versus pairs outside it.

Averaged across this 90-pair sample: positions within the semiconductor/AI-infrastructure theme correlate at 0.334 on average. Positions outside that theme correlate at just 0.104 on average — roughly a third as much. The book-wide sample average across all 90 pairs is 0.173. HHI treats all 32 positions as if they contributed diversification equally. The real, published correlation matrix says a meaningful cluster of them are largely the same bet, worn as different tickers.

3. Turning correlation into an honest position count

Portfolio theory has a standard way to convert average pairwise correlation into an "effective number of independent bets": N ÷ (1 + (N−1) × ρ), where N is the number of positions and ρ (rho) is their average pairwise correlation. At ρ = 0, all N positions are genuinely independent and the formula returns N. As ρ rises toward 1, the formula collapses toward 1 — every position moving together is, in a real risk sense, one position. Try it yourself:

Set N to 32 and ρ to 17% — this study's sampled book-wide average — and the formula returns roughly 5 effective independent bets, not the 19 the HHI-only view implies. Set ρ to the within-theme sample average of 33% instead, holding N at 32, and it falls further still. This is the same underlying book, described two different, both-true ways.

4. A premium, or an accident?

"Concentration premium" is deliberately a double meaning. In one reading, concentrating a book around a small number of correlated, high-conviction ideas is a legitimate strategy — you're accepting correlated risk in exchange for the expected premium of being right about a theme, not accidentally under-diversified. AEA's own Investment Policy Statement argues exactly this: a smaller number of well-understood, high-conviction positions beats a large number of shallow ones, and the book's 1.79 beta versus SPY is a stated, accepted cost of that choice, not a surprise. In the other reading, a portfolio can end up with correlated exposure by accretion — adding names one at a time because each looked good in isolation — without ever deciding, as a portfolio-level choice, to run 5 independent bets instead of 19.

The honest answer for AEA's own book sits closer to the first reading than the second, but only partially. The semiconductor and AI-infrastructure cluster was built deliberately and is disclosed as such throughout this site, including in the July letter's own framing of "one macro bet expressed multiple ways." What this study adds isn't a new fact about the strategy — it's a number for it. "Concentrated on purpose" and "roughly 5 effective independent bets out of 32 positions" are the same portfolio, but only one of those descriptions would make it into a pitch, and the other one is the one worth checking before you scale a position, not after.

5. What I'd actually watch

This study doesn't change AEA's position sizing today — the concentration was chosen deliberately and remains within IPS caps at the single-position level. What it changes is which number I check before adding the next name. A new position that looks small and "diversifying" by weight can still add close to zero real diversification if it correlates at 0.4–0.6 with names already in the book, the same way MU, NBIS, NBIL, and CRWV do with each other above. Going forward, a new position's correlation to the existing book is a more useful gate than its weight alone — HHI would wave it through either way.

Methodology & limitations