A game of microns 

Much of the world has spent the past month watching twenty-two people contest a ball across pitches throughout the North American continent. 

We have been watching a different contest, played out in Icheon and Pyeongtaek in South Korea, where the field of play is three-quarters of a millimetre high and the trophy is a share of the fastest-growing large market in the world economy.

Before we describe the game, we should declare our positioning, because it colours everything that follows. We own the three companies at the heart of this story, SK Hynix, Samsung Electronics and Taiwan Semiconductor Manufacturing Company (TSMC), in the FSSA Global Emerging Markets (GEM) Focus strategy. The profitability of the first two has spiked to levels never before sustained through a full cycle of the memory industry – and this has ignited fierce animal spirits among retail investors.

There are now exchange-traded funds (ETFs) which track individual Korean chipmakers and offer a leveraged return. Thousands of retail investors in Korea have completed the one-hour course their regulator requires before such products may be bought.¹ We have seen queues like this at stadium gates.

But memory has a long record of investing its way from shortage into glut – especially when demand is concentrated in a handful of customers, one above all.

If we were able to look back at today from 2027 or 2028, we think there is a good chance that current memory earnings represent a peak. Our weightings in these stocks, which we have been reducing steadily as share prices rose sharply (in contrast to the equally sharp decline in recent days), reflect that apprehension, rather than the market's excitement.

Why, then, devote attention to these businesses at all? Because of something we regard as distinctive in the way we invest. We hold companies for many years, and our advantage comes not from predicting the next quarter but from knowing a business deeply across a whole cycle. That includes the people, the franchise and the structure of the industry it inhabits.

That knowledge does not expire when a share price runs ahead of itself. When the valuation is wrong, the work is still worth doing, provided the business is good, because prices may correct, but understanding compounds. And the contest we are about to describe is exceptional material: a competitive game between three of the most consequential companies on earth, driven by specific engineering problems, whose outcome will shape the economics of computing and Artificial Intelligence (AI) for decades.

We are talking about the contest over high-bandwidth memory (HBM): the stacked memory chips that sit beside the processor in every serious AI system and feed it data. SK Hynix and Samsung Electronics are the finalists, while TSMC owns the stadium.

As spectators (and investors), we believe regardless of whomever turns out to have bought the right ticket, the game itself is worth watching. What follows is offered in that spirit. We explain why we are comfortable holding all three, and why we believe this is one of the rare sporting dramas in which a fan can back more than one side to win.
 

A decade ago, few would have picked SK Hynix as a leader. It was the industry's perennial number two, a respectable but unglamorous producer of commoditised memory, sitting in Samsung's long shadow. What changed was a bet on process. Today, HBM requires memory dies to be thinned to a fraction of the width of a human hair and stacked twelve or more high inside a package no thicker than a credit card.

Around 2019, Hynix developed a proprietary way of packaging and bonding those stacks, which proved to be faster, less energy-intensive and higher-yielding than the industry's standard approach. The result was a rout.

From 2019 through to the third generation of HBM (HBM3E, which fed Nvidia's Blackwell processors), Hynix supplied the majority of the market and earned margins unprecedented for a memory company. Samsung found its bonding process unable to match Hynix's yields and spent two years re-testing and qualifying products while its rival banked the profits.

For students of competition, it was a stark reminder that in manufacturing businesses, better tactics often decide the match. 
 

This year the fixture resumed with a new generation of chips, known as HBM4, and the early running has gone the other way. In February, Samsung became the first company in the world to mass-produce a twelve-layer HBM4, built on its most advanced memory process and a base die from its own foundry, running faster than the specification requires.

Hynix deliberately chose conservative technology for HBM4, prioritising the yields and reliability prized by its largest customer Nvidia.

As we write, England had just scraped a victory, playing Mexico at home to reach the World Cup's last sixteen. Meanwhile, in Korea's memory fabs, sixteen is the new benchmark number of layers both companies are being asked to meet for Nvidia's new Rubin platform. They are attempting it in contrasting ways.

Hynix is pushing its proven bonding process to the outer edge of what anyone believes possible by stacking sixteen dies – each ground to thirty microns, thinner than half a human hair. Samsung is attempting the same product with an entirely new technique called hybrid bonding. It is the acknowledged future of the industry, but its yields today are dreadful; perhaps just 10% of production, where over 60% is needed for commercial viability.

One team, Hynix, is playing a familiar formation at maximum intensity. The other, Samsung, has sent a young prodigy onto the pitch, who might turn out to be “player of the match”, or equally may be sent off with a red card.

We would add one detail that a football supporter will appreciate: the HBM tournament has its own FIFA, and it also changes the rules mid-competition. The industry's standards body, the Joint Electron Device Engineering Council (JEDEC), sets the maximum height for the finished memory package. Twice now, as stacking has approached the limit of the old rules, the crossbar has been raised, from 720 microns to 775, with a further move towards 900 under discussion.²

Rule changes are rarely neutral, and it’s clear who’s in favour: each raised ceiling gives Hynix's incumbent process more runway and defers the day when the industry must switch to the hybrid bonding technique. The new process belongs to nobody yet, but Samsung is pushing it hardest, largely because the exclusivity around Hynix's process leaves it no path to winning the present generation. Meanwhile, Hynix develops the same technology at a deliberate pace as insurance. 
 

Where does TSMC stand in all this? Nearly everywhere, is the simple answer.

The fixtures that matter most are staged at its grounds: when Nvidia buys memory from either company, the stacks are mounted alongside the processor on a thin sheet of silicon that TSMC manufactures and assembles. Hynix has embraced this arrangement under a partnership the two companies describe, a little unsubtly, as "One Team".

Samsung has not; and is building grounds of its own. Its foundry makes the memory stacks, its packaging operation assembles the chips, and its ambition is plainly to host matches rather than forever play away. We admire the attempt, and it is central to why we own Samsung – vertical breadth is a real option on the future shape of the industry.

But for now, the biggest matches are played where the biggest customer says they are played – and that remains at TSMC's venue.

A tournament framing makes good copy, but it misleads in one important aspect: nobody is being knocked out. The HBM market is forecast to grow by more than 50% this year alone, to something approaching USD 55 billion,³ with similar growth expected for years to come.

Both Hynix and Samsung have already sold out their 2026 supply.⁴ When demand outruns the combined capacity of every credible producer, the loser of a given generation forfeits margin, not existence.

Our portfolio weightings reflect this. We hold Hynix for its leadership, its packaging franchise and its intimate customer relationships. We have more measured exposure to Samsung for its front-end strength, its vertical breadth and the option value of a successful early move to the industry's next process. And TSMC is an anchor position for its possession of technical advantages that do not depend on either result. By owning all three, we need not predict the score, only remain confident the sport itself will keep growing.

The World Cup final later this month will be settled in ninety minutes, or a little more. This contest between memory makers will run for years. We intend to renew our season tickets and continue watching with interest.
 

Footnotes

¹ Source: Kwon Woo-seok, Korean investors flock to prelaunch course for single-stock leverage ETFs, ChosunBiz, May 2026.

² Source: JEDEC Reportedly Plans to Relax HBM Height Specs to ~900µm, Potentially Slowing Hybrid Bonding Adoption, TrendForce News, April 2026

³ Source: SK Hynix/Bank of America, January 2026

⁴ Source: Shannon Harwood, HBM Shortage Makes Frontier AI a Luxury Good by 2030, RIKEN Study Finds, Tech Times, July 2026