A $725bn hyperscaler capex ramp and a 20% HBM price rise have made the South Korean chipmaker the second-most valuable company on the KOSPI. The harder question is when supply catches demand.
There is a small joke in semiconductor circles about which part of an AI server is most expensive. The graphics processor used to be the obvious answer. For the past 18 months, increasingly, it has been the memory soldered next to it.
On Monday, the market priced that joke.
SK Hynix, the South Korean memory specialist that supplies the bulk of the world’s high-bandwidth memory for AI accelerators, climbed as much as 12 per cent in Seoul, with shares hitting roughly 1.4 million won, or about $970, in morning trading, according to Reuters.
The rally made SK Hynix the second-most valuable company on the KOSPI, behind only Samsung Electronics, and reflected, the wire said, foreign buying that followed strong earnings and reaffirmed AI infrastructure plans from US hyperscalers the previous week.
It is, by any measure, a remarkable run for a company most consumers have never heard of.
The trigger is straightforward. Big Tech’s combined 2026 capital expenditure is on track to land somewhere between $650bn and $725bn, depending on which analyst’s tally one trusts, an increase of roughly 77 per cent on 2025.
Microsoft has guided to as much as $190bn for the calendar year, with its chief financial officer publicly attributing about $25bn of that to rising memory-chip and component costs.
Meta, in its Q1 update, raised its own range to $125–145bn, citing similar pressures. Amazon’s Andy Jassy has committed roughly $200bn. Google has not been quieter. Practically all of this money flows, in one form or another, towards AI training and inference clusters; a meaningful share of it lands in the bill of materials for high-bandwidth memory, where SK Hynix dominates.
By late 2025, SK Hynix held an estimated 57 per cent of the global HBM market, according to figures cited by analysts at Counterpoint and others. That share is unusually concentrated for a commodity-adjacent business, and it is the structural reason the company’s earnings now look more like those of a software platform than a memory house.
SK Hynix’s first-quarter operating profit, reported on 23 April, was a record. Operating margins on its memory line, by some sell-side estimates, are running above 70 per cent.
Margins of that order do not last forever. They do, however, last as long as supply lags demand.
Why supply is not catching up
HBM is not ordinary DRAM. It is a stacked, 3D-packaged memory built to feed bandwidth-hungry GPUs, and producing it requires a specific set of advanced packaging steps that the industry, including Samsung and Micron, has been slower to scale than buyers would like.
According to TrendForce, both Samsung and SK Hynix have raised HBM3E prices by roughly 20 per cent for 2026, and supply is being booked years in advance by hyperscalers and accelerator vendors.
Samsung’s memory chief publicly warned earlier this year that significant memory shortages were likely to persist through 2027. The chairman of SK Group has gone further, telling investors he expects the wider chip-wafer constraint to last until 2030.
Whether or not those forecasts prove accurate, they explain why long-term supply agreements, in which a hyperscaler effectively reserves output years ahead, are becoming the norm. They also explain why Reuters reports SK Hynix and Samsung increasingly signing such deals with Microsoft and Google directly.
There are, in other words, two kinds of memory in 2026: the kind anyone can buy, and the kind your AI roadmap depends on. SK Hynix is in the second business.
There are, predictably, dissenters. The CAPE ratio on US equities now sits around 38, a level last seen at the height of the dot-com era, and TNW has noted the discomfort that comparison provokes, even as it argues that today’s leading AI-exposed companies are, unlike many in 2000, broadly profitable.
SK Hynix is squarely in that profitable cohort, but it is also unusually leveraged to a single product cycle. If hyperscaler capex moderates, or if competing accelerator architectures reduce HBM intensity per chip, the same operating leverage that has delivered record quarters could work in the opposite direction.
There is also the question of whether the AI build-out will keep producing the kind of returns that justify present capex levels. Meta has been simultaneously announcing record AI investment and cutting roughly 8,000 jobs as it restructures around the spending, a sequence that does not entirely fit the narrative of a healthy organic boom. Investors are, for now, willing to fund both ends of that equation. They have done so before in technology cycles, and they have changed their minds quickly.
The most useful way to read SK Hynix’s 12 per cent move is not as a forecast of where AI ends up, but as a real-time index of how confident the market currently is that AI training clusters will keep being built at this pace. Every dollar of hyperscaler capex announced becomes, eventually, an order book line at a memory supplier. SK Hynix sits in the chokepoint.
If the chokepoint loosens, through Samsung executing on HBM4 at scale, through Micron’s $25bn capacity push, or through architectural shifts that reduce HBM dependency, the dominant share gets shared and the multiples compress. None of those things has happened yet.
Until one does, the market’s read is that the second-largest company on the KOSPI is the one selling shovels to the AI gold rush, and it is not running out of buyers.
The harder, more interesting question is what the chart looks like a year from now. The answer will say less about SK Hynix than about whether the AI build-out has legs, and whether the people writing the cheques in Redmond, Menlo Park, and Seattle are still as certain as they were last week.
