Opinion Leaders
When diversification becomes an illusion
-
Jean-Christophe Rochat
Chief Investment Officer, Head of Asset Management
Banque Heritage
It is not the number of holdings that diversifies a portfolio, but the number of truly independent economic drivers.
Diversification is still often measured using inherited categories: equities and bonds, the United States, Europe and Asia, listed and unlisted. Yet these labels say little about what actually generates cash flows, and therefore risk.
Artificial intelligence (AI) offers a striking illustration. The four major hyperscalers are expected to invest around USD 725 billion in AI infrastructure in 2026, 77% more than the roughly USD 410 billion spent in 2025. And this flow reaches far beyond the tech sector. A Taiwanese semiconductor manufacturer, a Dutch equipment supplier, a US power producer, a hyperscaler bond and a private loan financing a data centre belong to five different categories. Economically, however, they largely depend on the same scenario.
This concentration now extends to credit. Hyperscalers have issued more than USD 200 billion of debt since January, roughly twice the total for 2025. Nearly a third of their 2026 capex is expected to be debt-financed. Their share of US dollar investment-grade issuance would thus rise from about 2% over 2022–2024 to nearly 9% in 2026. Morgan Stanley also estimates the private credit opportunity linked to data centres at some USD 800 billion by 2028.
One signal is particularly revealing: some investors are already nearing their issuer exposure limits once they add the debt of dedicated data-centre vehicles to that of the parent companies. This is exactly how a concentration that is invisible in traditional classifications comes to light. The trap is all the harder to detect because our financial models rely heavily on historical correlations. A recent economic link has, by definition, not yet had time to show up in statistical series. Assets that are weakly correlated today may therefore start moving together as soon as the common factor they depend on is called into question.
Diversification must therefore be assessed in terms of future economic dependencies, not just past correlations. From this perspective, “getting out of the United States” is not enough. Replacing Nvidia with Korean or Taiwanese semiconductor stocks diversifies the portfolio’s geography, but not necessarily its risk driver, since all of them remain exposed, to varying degrees, to the same AI investment cycle.
This does not mean that all these exposures are equal. Fundamentals can diverge sharply within the same economic chain: according to MSCI, capex at AI-exposed US companies jumped by nearly 60% in one year, far faster than their sales, while in Taiwan sales are growing at the same pace or even faster. A single risk factor can therefore link several assets without giving them the same vulnerability profile.
AI is merely a case study here. The same reasoning can apply to energy, China, interest rates, the dollar or the US consumer. The real question is therefore no longer “How many asset classes do I hold in my portfolio?” but “How many truly distinct scenarios can contribute to its performance?” Two different assets can carry the same risk. And a portfolio of fifty holdings can be less diversified than a portfolio of twenty if they all rely on the same driver.
Diversification is not measured by the number of positions, but by the degree of independence between their sources of return and risk.