SK Hynix's $26.5B US IPO: A Deep Dive on the Company Behind the AI Memory Wall

On July 10, a Korean chipmaker most US retail investors have never once traded became the largest foreign IPO in American history. SK Hynix priced 177.9 million American depositary shares at $149 and raised $26.5 billion, edging past Alibaba's $25 billion 2014 record (Bloomberg). The headline is the IPO. The actual story is that the single tightest bottleneck in the entire AI buildout — high-bandwidth memory — now has a liquid, US-listed proxy, and its economics are more extreme than most of the AI trade appreciates.

The deal, and why the frenzy

The offering was more than seven times oversubscribed and priced at a 2.7% premium to SK Hynix's three-day average in Seoul, an unusual show of demand for a secondary listing of an already-public company (TechCrunch). The stock (Nasdaq: SKHY) opened roughly 14% above its offer price, and chairman Chey Tae-won told CNBC in his debut interview that "demand is enormous" (CNBC).

What matters for the thesis is where the money goes. The proceeds are earmarked for a new South Korean fab already under construction, a new domestic packaging facility, and EUV lithography scanners for next-generation chips (TechCrunch). This is a capacity-expansion raise, not an insider cash-out — a company doubling down on capex to feed AI memory demand it evidently expects to keep running hot.

The margins are the tell

Memory has historically been one of the ugliest businesses in tech: commoditized, viciously cyclical, thin margins at the bottom of every cycle. SK Hynix's most recent quarter looks nothing like that. In Q1 2026 the company reported 52.58 trillion won (~$35.5 billion) in revenue, up 198% year-over-year, its first quarter ever above 50 trillion won. Operating profit hit 37.6 trillion won for a 72% operating margin, with net margin at 77% (SK Hynix, CNBC).

A 72% operating margin at a DRAM maker is not a normal data point — it is the signature of a genuine supercycle, where demand outruns supply and pricing power flips from the buyer to the seller. The question every analyst underwriting this stock has to answer is whether that margin is a peak to be faded or a new plateau. The answer runs entirely through one product line.

HBM: the moat and the concentration, in one number

High-bandwidth memory — the stacked DRAM that sits next to every AI accelerator — is now more than half of SK Hynix's total revenue, up from roughly 25-30% in 2024. In effect, a commodity memory maker has quietly restructured itself into an AI-accelerator components company. SK Hynix got there first: it held an estimated 90% of the HBM market in 2023 and still leads with roughly 56% share in Q1 2026, plus a mid-50s% slice of Nvidia's next-generation HBM4 allocation (HBM market data).

That leadership is also the risk, because it rests on one customer. Per its US prospectus, Nvidia accounts for roughly 15% of SK Hynix's consolidated revenue, and industry estimates put Nvidia at north of 50% of HBM revenue specifically (SEC Form 424B4). Buying SK Hynix is, to a first approximation, buying Nvidia's demand curve with a memory-cycle beta bolted on top.

And the 56% moat is narrowing. Micron has climbed from about 2% HBM share in 2023 to roughly 19-20%, and Samsung has been requalifying its HBM3E across Nvidia and AMD platforms (HBM market data). SK Hynix remains the leader, but the days of near-monopoly are over — this is a three-way war now, and share erosion at the top is the base case, not the tail risk.

The policy wildcard: where do the next fabs go?

There is one more overhang the stock price doesn't cleanly capture. US Commerce Secretary Howard Lutnick is reportedly in talks with both SK Hynix and Samsung about building new factories on American soil, with the stated aim of preventing South Korea from continuing to dominate a strategic technology (TechCrunch). Yet the IPO proceeds are pointed at Korean fabs, Korean packaging, and EUV tooling — not US capacity. That gap, between where SK Hynix wants to deploy capital and where Washington wants it deployed, is exactly the kind of US-Korea industrial-policy negotiation that can move a semiconductor stock independent of its earnings.

What to watch

The US listing hands investors a rare pure-play on the AI memory wall. But pure-play cuts both ways: in a single ticker you are underwriting Nvidia's order book, the memory cycle turning at some point, an eroding HBM share, and a cross-border fab negotiation. Three things worth tracking from here: the HBM4 ramp and its yields (the next battleground for share); whether SK Hynix can durably diversify its customer base — a second buyer already crossed the 10% revenue threshold at roughly $4.4 billion, the first non-Nvidia customer to do so (SEC Form 424B4); and any concrete commitment on a US fab.

The market has decided SK Hynix is the cleanest way to own the memory side of AI. The financials support the enthusiasm. The concentration and policy risks are the reason to read the prospectus before you read the ticker.

This analysis was produced with dwork.ai's research agent.