The Diversification Illusion: What Two Chip Stocks Reveal About Portfolio Construction

Micron and SK Hynix accounted for roughly 17 percent of the entire return of the MSCI All Country World Index in May. Two companies most retail investors couldn't name moved the entire planet's equity markets almost by themselves. That fact should worry anyone who thinks they have diversified away from the AI trade. Most haven't.

Micron and SK Hynix are not household names. In May, they accounted for roughly 17 percent of the entire return of the MSCI All Country World Index, a benchmark holding thousands of stocks across dozens of countries. Two companies most retail investors couldn't name moved the entire planet's equity markets almost by themselves.

That fact should worry anyone who thinks they have diversified away from the AI trade. Most haven't.

The structure of the problem is not straightforward. It does not appear in portfolio reports, because the usual tools for measuring diversification, such as number of holdings, sector weights, and geographic allocation, measure the wrong thing. They count positions. What they do not measure is whether those positions share a common driver of return. That is the gap between what diversification looks like and what it actually does during a drawdown.

How the Trade Hid in Plain Sight

The obvious AI exposure is chipmakers and hyperscalers. The less obvious exposure is everywhere else. Utilities are now partially AI stocks, because data centres need enormous amounts of power. Real estate is exposed because someone has to own the buildings full of servers. Construction firms, industrial equipment manufacturers, and the companies making commercial air conditioning systems for high-density computing facilities are all riding the same boom, whether their shareholders have noticed or not.

Small caps are not a clean escape either. The Russell 2000 had its best first half since 1991 this year, rising more than 20 percent. Look inside the index and a large share of the strongest performers are semiconductor and chip-equipment suppliers, up between 200 and 400 percent year to date. Investors who thought they had rotated into ordinary American businesses mostly acquired the picks and shovels of the AI trade instead. The category label said small cap value. The return driver was AI capital expenditure.

Even value investing, the supposed refuge from overpriced growth stocks, got pulled in by a mechanism that had nothing to do with individual stock selection. Through most of this year, value indices were loaded with semiconductor names that had rallied hard enough to cross the value threshold defined by their respective index methodologies. In late June, index providers ran their annual rebalance and moved those chip stocks into the growth index, replacing them with Amazon, Apple, and Microsoft. The timing was coincidental and, in hindsight, close to optimal: chip stocks were sold near the top of their run, and value indices picked up mega-cap technology right as it touched its lows for the year. Nobody made an active decision. A calendar did it. The result is that a large share of value-labelled money is now sitting in three of the largest technology companies on earth.

The Arithmetic of Genuine Diversification

The instinct when facing concentration risk is to look for a different set of stocks to own. That instinct misses the actual mechanism at work.

Diversification is not really about owning more line items. It is about owning return streams that do not move together for the same reason. The formal definition is correlational, not numerical: a portfolio of two assets with a correlation of negative one has more genuine diversification than a portfolio of a thousand assets all correlated with the same macroeconomic variable. The number of tickers is noise. The correlation structure is signal.

A portfolio of a hundred stocks that all rise and fall with the same macro driver, in this case AI capital expenditure and the earnings multiples that have been built on top of it, is not diversified in any way that matters during a drawdown. When the driver reverses, every instrument tied to that driver falls together, and the apparent spread of positions provides no protection because there is no structural difference between them. The disaster is not having too few stocks. It is having too many stocks with identical underlying exposures that have been mislabelled as different things.

This distinction between position count and true diversification is not a theoretical subtlety. It is the difference between a portfolio that holds up during stress and one that does not.

A Pattern With a Long Record

History is not short of examples of this structure producing sudden and severe outcomes.

The Nifty Fifty episode of the early 1970s is the most instructive in terms of mechanism. By 1972, a group of roughly fifty large-cap American growth stocks had become so widely held by institutional investors that they were considered permanent holds, not trades. The reasoning was coherent: these were exceptional businesses with durable competitive advantages. The quality of the businesses was not in question. What was in question was the price paid for them, and how thoroughly that quality had been priced into the multiples. When conditions shifted in 1973 and 1974, the same institutional herding that had driven prices up became the mechanism of the decline. The businesses remained exceptional. The returns were deeply negative.

Japan in 1989 demonstrated the same structure at the level of an entire equity market. The Nikkei 225 reached 38,915 on the final trading day of 1989. Nippon Telegraph and Telephone alone, at the peak of its valuation, was worth more than the entire West German equity market. Japanese real estate, Japanese equities, and Japanese banks had become so intertwined through cross-shareholding structures that they functioned as a single correlated position, not as distinct asset classes. The subsequent decline took more than two decades to work through.

The NASDAQ concentration of 1999 and early 2000 is more recent and more directly analogous to the current situation. The ten largest holdings in the NASDAQ 100 accounted for more than 60 percent of the index by early 2000. Investors in broadly diversified index funds believed they owned the broad market. They owned a concentrated bet on internet-era valuations. The index fell more than 80 percent from peak to trough. None of these episodes required investors to have been wrong about the underlying technology or business quality. They required only that the price paid for that quality had incorporated an assumption about the future that could not be sustained indefinitely.

The current AI concentration does not replicate any of these episodes mechanically. The scale, the instruments, and the structural context are different. What is structurally similar is the combination of genuine technological significance, consensus conviction, and the way that conviction has distributed itself invisibly across categories that are nominally separate.

Why This Is a Construction Problem, Not a Stock-Picking One

The appropriate response to identifying hidden concentration is not to find better stocks. It is to identify return sources with a structurally different relationship to the dominant driver.

What actually reduces portfolio risk is exposure to return streams with a different, ideally negative, relationship to the primary risk factor. The language here is important: different relationship, not opposite view. An allocator does not need to believe AI spending will decline in order to benefit from holding something that is not correlated to AI spending. The protection comes from the structure of the return, not from a forecast.

This is the argument for strategies built around systematic futures rather than long-only equity selection. A systematic futures approach that can go short as easily as long, across equity indices, rates, currencies, and commodities, is not making a directional bet on any single narrative continuing. Its return profile comes from price trends and dislocations across multiple markets simultaneously, not from the AI capital expenditure cycle specifically. When equity markets fall sharply, the same trend-following mechanics that captured the rally on the way up are positioned to capture the reversal on the way down, producing what is often described as crisis alpha: returns concentrated in exactly the periods when long-only portfolios are under the greatest pressure.

This is a structurally different relationship to risk than adding another equity fund with a different label attached to it. The question is not whether the fund is named growth, value, small cap, or emerging markets. The question is what happens to its returns on the days that the AI trade reverses hardest. If the answer is that it falls alongside the core equity book, it is not adding diversification. It is adding a label.

The Correlation Test That Most Portfolios Skip

For an allocator, the practical question is not whether the portfolio is overweight technology. Most portfolios already know the answer to that. The harder question is whether the assets labelled as diversifiers actually have a different return driver than the core equity book, or whether they are the same trade with a different name.

That means checking correlation in stress conditions specifically, not average correlation across all market environments. Assets that appear to have low correlation during calm periods frequently converge during sharp drawdowns, because the common driver in a crisis is liquidity, not sector fundamentals. When investors need to sell, they sell what is liquid and what has performed. That process can temporarily correlate assets that are genuinely independent in normal conditions.

The meaningful test is what happened to the diversifier during the worst ten or twenty days for the core equity book over the last three to five years. A commodities allocation that is mostly energy and industrial metals will still move with global growth expectations, which are currently entangled with AI capital expenditure. A market-neutral fund that is long momentum will often be structurally long the same crowded trade it claims to hedge, because momentum strategies naturally accumulate positions in what has been working. The label matters less than the behaviour on the specific days that matter most.

Checking this requires going beyond summary statistics. Monthly correlation figures smooth over the episodes that are most relevant to portfolio construction. The relevant data is daily returns during drawdown periods. That is where genuine diversification reveals itself and where the illusion is most costly.

What This Means for Allocators

None of this requires a forecast. The optimists who argue that AI represents a real and transformative technology are not a fringe view. They are the majority, and their conviction is the reason the boom exists at all. The point is not to dispute the technology or the business quality. The point is about construction.

A portfolio that only performs well if the AI trade continues working is making a directional bet, whether or not anyone has labelled it that way. That is a legitimate choice. The problem arises when the directional bet is invisible inside categories that have been assembled to look diversified. The exposure is real. The diversification label is not.

The practical steps for allocators are specific. First, map the actual return drivers of every position, not the category labels. Second, apply the stress correlation test: what happened during the worst equity drawdowns of the last five years. Third, examine any allocation described as a diversifier and ask whether it has a mechanically different return source, one that does not depend on the same macro driver, or whether the difference is mainly in presentation.

The challenge is that the answer to that third question, for most broadly diversified equity portfolios constructed in the last two years, is that the diversification is largely presentational. The driver is the same. The labels are different. The two chip stocks that moved 17 percent of global equity markets in a single month are not the risk in isolation. The risk is the number of portfolios that are quietly positioned on the same driver, spread across categories that have been assembled to look different from each other, and that will discover their common exposure at the same moment.

The Question Worth Sitting With

Every past concentration episode, from the Nifty Fifty to Japan in 1989 to the dot-com peak, looked exactly like this one feels from the inside: not a mistake, but a consensus. The businesses were real. The technology was real. The prices were what they were. Being right about a good business and being right about the price paid for it are two separate judgements, and history suggests allocators who checked only the first one have consistently found the second more expensive than they expected.

The appropriate response is not to sell the AI trade. It is to check whether the portfolio has any genuine alternative to it, and to do that check with the right tools: return drivers and stress correlations, not position counts and sector labels.

This publication is issued by Black Flower Capital GmbH (Zug, Switzerland) and Black Flower Capital Management SARL (Luxembourg). It is intended for professional and institutional investors only, is provided for information purposes only, and does not constitute investment advice, an offer, or a solicitation to buy or sell any financial instrument. Past performance is not indicative of future results.