3Q26 Letter--The EM Index is no longer diversifying for allocators
Emerging markets earned their place in institutional portfolios on a simple proposition: a large, heterogeneous set of economies at different stages of development, driven by different things, priced by different investors, and therefore an imperfect substitute for anything else an investor owned. That proposition still describes the asset class. It no longer describes the index. Taiwan and Korea are home to under 2% of the population of the index’s 24 countries, yet they are now over 50% of MSCI Emerging Markets, more than China and India combined. Information technology is 44% of the index, against roughly 22% a decade ago. Three companies in one supply chain — TSMC, Samsung Electronics and SK Hynix — are about 30% of it.
For an active manager, that concentration forces a choice. We own all three of those semiconductor franchises, but we have deliberately limited our exposure to them, and through the first half of 2026 we reduced our holdings in SK Hynix and Samsung Electronics. This piece sets out why we did it, where we do and do not own the artificial-intelligence (AI) capital cycle, and where the rest of the portfolio is invested.
Rapid changes in the MSCI EM Index
Ten years ago the MSCI Emerging Markets Index was recognisably diversified. China was around a quarter of it, Korea and Taiwan around a quarter between them, India roughly 8%, and roughly 40% sat in the nineteen other countries. Financials, at about 26%, were the largest sector. Two large rotations followed. China rose to roughly 40% of the index by late 2020 and fell back to 19.8% by September 2026. And, much faster, Taiwan and Korea went from about a quarter of the index to just over half of it, almost entirely in the last three years. Information technology has gone from 13% of the index in 2008 to 43.8% in September 2026. Energy is now 3.4%. Consumer staples — the classic emerging-market domestic-demand sector — is just 2.6%.
On MSCI’s factsheets at 30 September 2026, the ten largest constituents of MSCI Emerging Markets are 39.0% of the index, against 38.2% for the ten largest in MSCI USA. Seven of the ten are information technology companies in Taiwan and Korea, together 33.6% of the index; the other three, Tencent, Alibaba and China Construction Bank, add 5.5%. At the top, TSMC alone is 15.6%, almost twice Nvidia’s 8.0% weight in MSCI USA. Samsung Electronics is 7.6%, its preference line a further 0.9%, and SK Hynix 5.9%: with TSMC, 30.0%. That combined weight was 13.9% at 30 June 2025, so the three have more than doubled as a proportion of the index in fifteen months. Nor does the sector label capture the full exposure. Our analysis shows that once AI capex-linked stocks with GICS classifications other than information technology are included, more than 50% of MSCI EM is linked to the AI investment cycle.
The equal-weighted index — a fair proxy for the median company — has returned 5.7% in 2026 against 23.4% for the cap-weighted index, and 5.0% a year over ten years against 9.1%. The index return has come from a handful of names rather than broad emerging-market progress. It is also the reason many benchmark-independent managers have trailed the cap-weighted index over this stretch.
For allocators the point is compositional. Taiwanese foundry capacity and Korean high-bandwidth memory (HBM, the stacked memory used in AI accelerators) are not a different bet from Nvidia and the hyperscalers; they are the same bet, taken earlier in the supply chain.
The same dependency runs through the rest of a typical allocator’s book. Taiwan’s foundries and Korea’s memory makers depend on the same AI build-out as US-listed hyperscalers and AI chip designers, as private credit financing data centres and infrastructure, and as venture capital backing AI models and applications (Exhibit 1).
EXHIBIT 1
Source: Variis Partners LLP.
This allocation point is not just a passive problem in EM: most “active” funds are very similar to the benchmark. The peer group’s average stock-level active share (the share of a portfolio that differs from the index) has fallen from 75% to 64% in two and a half years. The Variis strategy closed 3Q26 with an 83.5% active share.
An investment portfolio is meant to be more than one bet
The case for owning a collection of genuinely different bets rests on a simple observation: the risk of a portfolio depends less on how risky each holding is than on how closely the holdings move together. The vocabulary of correlation and covariance as applied to portfolios arrived only in 1952, with Harry Markowitz, but the principle is far older. It is not financial technology. It is ordinary commercial sense, and merchants have practised it for centuries: Venetian and London traders took shares in several voyages rather than staking everything on one ship, and the joint-stock companies of Shakespeare’s day were built on the same idea.
In the opening scene of the Merchant of Venice, Antonio’s friends assume he must be anxious about his ships at sea, and he explains why he is not:
“My ventures are not in one bottom trusted,
Nor to one place; nor is my whole estate
Upon the fortune of this present year.”
The line needed no explanation for a London crowd that included merchants, factors and apprentices; spreading a cargo was simply how trade was done. In Elizabethan English a ship’s “bottom” was its hull, and by extension the ship itself. The same usage survives in “bottomry”, the old maritime loan secured on a ship that the lender forfeited if the vessel was lost at sea — an early form of risk-sharing on a single voyage. Antonio does not believe that each voyage is safe. A ship, sensibly routed, can still meet a storm. He believes that no single storm can sink the estate, because his cargoes are spread across ships, ports and seasons. The play then tests him harshly. By the third act, word arrives that his ventures have failed “from Tripolis, from Mexico and England, from Lisbon, Barbary and India”, and he is brought close to ruin. In the final act, three of his ships come safely into harbour after all. In a crisis everything looks correlated. What matters is whether it really is. Good judgement on each voyage was not enough. The voyages had to be different.
Source: Wikipedia.
Why we limit our AI exposure, and where
We are bullish on the AI build-out—just not bullish enough to make it the only meaningful bet in our strategy or to pay exorbitant multiples on possibly bottleneck-inflated margins.
Compute capacity remains the binding constraint on AI. Existing data centres are insufficient for current demand, let alone future demand — a materially different situation from the late-1990s fibre overbuild, where capacity was laid in anticipation of traffic that arrived years later. We expect a prolonged shortage of advanced semiconductors, and we think AI data-centre demand deserves a higher multiple than traditionally cyclical consumer demand. Our restraint is therefore not a demand call. It rests on three narrower judgements.
First, exceptional returns invite a capital response. We estimate SK Hynix will earn a return on capital employed of roughly 180%, and Samsung a group return of 80–90% in 2026 and 2027, a group figure diluted by the rest of the conglomerate. Returns of that order invite new capacity, customer pushback, renegotiation of long-term agreements and substitution effort. Memory has always been the part of the semiconductor complex where extraordinary economics attract supply fastest.
Second, the market is holding two incompatible views at once. On 2027 consensus earnings, King Slide trades on 31 times, MediaTek on 68 times and Aspeed on 86 times — all AI value chain supplier businesses whose multiples embed years of continued capital-expenditure growth. On the same basis SK Hynix trades on about 4.7 times and the Samsung preference line on about 4 times, which embeds an imminent earnings collapse. If capital expenditure is durable enough to justify nearly one hundred times earnings for some suppliers, memory should not be priced for collapse. If memory earnings are about to fall away, a much more sceptical view is warranted for high multiple suppliers whose economics depend on the same spending. We have resolved the inconsistency by owning the cheap memory bottleneck, in a meaningful but limited size, and avoiding the high multiple suppliers. The expensive names are often not bad businesses. King Slide is a near-monopoly in server rails with roughly 80% gross margins, and we would like to own it; at 31 times 2026 earnings, with a 64% net margin after tax, we would be underwriting the AI capex growth for years.
Third, we apply concentration risk to ourselves. Neither stock was expensive on our numbers when we sold. We sold because unmanaged concentration would have left the strategy hostage to a single technology debate. Over the same period, the index’s information technology weight rose from 31.8% in March 2026 to 43.8% by the end of 3Q26.
A history lesson from 1873
There is no such thing as a sure bet. An investor can make a bullish demand forecast, see it proved right, and still be wiped out. That is the lesson of 1873.
This quarter we have been reading Liaquat Ahamed’s 1873: The First Great Depression and the Making of the Modern World. It tells the story of the first genuinely global financial crisis. The crisis reached New York on 18 September, when Jay Cooke & Company — the well-known banker and the financier of the Northern Pacific Railway — failed. The New York Stock Exchange closed for ten days.
Source: Wikipedia; Liaquat Ahamed, 1873 (book cover).
Railway construction had absorbed as much as a fifth of American capital investment. Some 33,000 miles of track were laid between 1868 and 1873, 7,500 miles in 1872 alone. After the panic, track laid fell to 1,600 miles in 1875, and total mileage barely increased between 1873 and 1878. For the businesses that supplied the build-out — rail mills, ironworks, locomotive and car builders, and the coal mines that fed them — this was not a slowdown but the near-disappearance of their marginal customer, and unemployment was heaviest in exactly those industries: coal, iron and steel, and engineering. The National Bureau of Economic Research dates the contraction from October 1873 to March 1879. At 65 months it is still the longest in its records.
What makes 1873 instructive is that the long-run thesis was right. In 1873 the network was not overbuilt relative to its eventual scale; it was less than a third built (Exhibit 2). Demand for track, rails and locomotives over the following four decades exceeded anything a bull in 1872 could reasonably have modelled.
EXHIBIT 2
Source: US railway mileage series and railway receivership counts; Variis Partners LLP.
Had we been managing Variis in the years before 1873, we would very likely have owned railway supply-chain capital goods businesses. We would have seen demand set to exceed supply for years to come, and we would have been right about the decade and about the half-century. The future is in some sense unknowable; it arrives as a distribution of outcomes, not as a forecast. The lesson we draw is that investing needs conviction and insight combined with prudent diversification.
The bigger point is that we want a portfolio of relatively uncorrelated bets, not one big bet, even when it is a good one. We have as much reason to believe in the AI build-out as a railway investor in 1872 had to believe in the American network, and for a similar reason: compute today, like track then, is far short of its long-run scale. That is precisely why we limit our exposure to AI capital-expenditure names, and why we reduced our holdings in SK Hynix and Samsung as their share prices rose.
Our exposure to the AI build-out sits in what we call AI picks and shovels, across four positions. Samsung Electronics preference shares are the cheapest access point in the complex: the discount to the common shares has run between 23% and 37% through 2026, against a longer-run average nearer 10–15%, and Korea’s Value-Up programme gives that gap a catalyst. Delta Electronics gives us power delivery and thermal management, the physical-infrastructure bottleneck we regard as more structurally scarce than memory bits. SK Hynix is the purest HBM franchise and the one that executed best in the first AI memory wave. TSMC has no substitute at the advanced production node; our restraint there reflects valuation upside and portfolio diversification, not the quality of the business.
Beyond the AI supply chain: where the old emerging-market diversification still lives
Investors originally wanted emerging markets for three reasons, and none of them was semiconductors: an emerging middle class growing its consumption from a low base; markets less efficiently priced than developed ones; and economies driven by different things than in the US. All three still hold, just not in the cap-weighted index. The rest of our portfolio is invested in these original and enduring attractions of EM.
Much of it sits in what we call idiosyncratic emerging markets ideas. Financials are domestic banking franchises compounding in economies where credit penetration is low. Bricks and mortar holdings are physical assets and networks that software cannot disintermediate:. Other idiosyncratic situations are similarly fundamentally uncorrelated with AI capex. We also invest in emerging-market digital platforms, whose drivers are domestic demand. None of these businesses depends on what the hyperscalers spend next year.
Conclusion
The MSCI EM Index trades on under ten times forward earnings, and the emerging and developing world is around 86% of the global population while emerging markets are only about 10% of MSCI ACWI. We built Variis because of that multi-year opportunity. We are arguing that the implementation decision now matters more than the allocation decision. If the objective is diversification, a passive or benchmark-hugging allocation delivers less than people assume. If the objective is AI exposure, many allocators already hold a concentrated version of it in several places. Emerging markets can still offer what allocators came for: a diversifying source of returns. Beyond the AI supply chain are thousands of companies, among them banks, consumer brands, healthcare and infrastructure businesses, driven by domestic growth and local cycles. A high-conviction, high active share portfolio can own the best of them, size each position on its merits and keep the bets different from one another. That is where diversifying alpha can come from.
The main risk to our view is that narrow leadership persists. If the AI capex growth continues sustainably, market concentration will keep rising and the index will continue to outpace benchmark-independent portfolios such as ours. Our view that breadth widens as a bull market matures is a view, not a fact. Our valuation discipline remains: twenty-two names across eleven countries, strict single-position limits, and no energy, materials, utilities or real estate. The four partners have most of their liquid net worth invested in the strategy alongside our investors.
Thank you for your continued interest and support!
Leila, Eko, Rufus and Jamie
Disclaimer
FOR PROFESSIONAL INVESTORS AND ADVISORS ONLY
The contents of this document are communicated by, and are the property of, Variis Partners LLP. The information and opinions contained in this document are subject to updating and verification and may be subject to amendment. No representation, warranty, or undertaking, express or implied, is given as to the accuracy or completeness of the information or opinions contained in this document by Variis Partners LLP or its directors. No liability is accepted by such persons for the accuracy or completeness of any information or opinions. As such, no reliance may be placed for any purpose on the information and opinions contained in this document. The information contained in this document is strictly confidential and is not intended to be advice or an offer or solicitation to invest. The value of investments and any income generated may go down as well as up and is not guaranteed. Past performance is not a reliable indicator of future results.