If You Missed NVIDIA, Is This Way Bigger? SK hynix's HBM Crown and the ADR Premium Equation

Had you put ten thousand dollars into a freshly listed industrial stock three years ago, it would be forty thousand today; the same amount placed in a 2020 direct listing of a data platform would be close to two hundred thousand. Ticker Symbol: YOU host Alex, an MIT-trained electrical engineer and AI researcher for eight years who says he spotted Nvidia, Micron and TSMC early, turns that lens on SK hynix, the memory maker every AI data center chases, and its new NASDAQ ADR. The promise is clear, the question is clearer — is this memory leader worth buying at a premium?

From November 2022 to the Trillion-Dollar Bottleneck

When OpenAI released its chat assistant in November 2022 in San Francisco, it hit one hundred million users in two months, four times faster than TikTok and more than fifteen times faster than Instagram. The wave that followed became the biggest tech investment cycle since the internet, lifting Nvidia roughly twelvefold toward a five trillion valuation as its accelerators became the engine of generative AI. Yet those accelerators became so fast and power-efficient that the rest of the data center fell behind: fans could not cool enough, networks could not move fast enough, memory could not feed fast enough, and expensive GPUs sat idle burning cash instead of making tokens. The contest shifted from who had the fastest chip to who had the infrastructure to feed it.

Hyperscaler Spend and the HBM Trio

That is why Amazon, Microsoft, Google and Meta are expected to spend well above a trillion dollars on AI infrastructure in 2025 and 2026 combined, on chips, cooling, networks and memory. At the system level an investor could buy the leader in GPUs, cooling and networking, but in high-bandwidth memory the US menu was thin: of the three major HBM players — SK hynix, Samsung and Micron — only Micron traded in the United States, which is why Alex had favored it before. Today SK hynix alone holds about half the HBM market, as much as Samsung and Micron combined, a position that was out of reach for most US portfolios until this summer.

The door opened on July 10. SK hynix began trading on NASDAQ as SKHY, an American depositary receipt that represents one tenth of a Korean common share, modeled on how TSMC and Alibaba list in New York. The offering raised $26.5 billion, described as the largest US IPO by a foreign company, with proceeds earmarked for new fabs and EUV tools. The catch is price. With 729 million shares outstanding, the Korean listing implies about $920 billion of value, roughly $1,260 per share or $126 per ADR, yet the ADR changes hands around $175 on NASDAQ. Same firm, same share, roughly a 40% premium on the US line, $920 billion in Seoul math versus $1.3 trillion in NASDAQ math. Alibaba trades near parity between Hong Kong and New York because conversion is open; TSMC trades at about a 12% premium due to limited ADR supply. SK hynix is tighter still, with only 2.5% of the company convertible into ADRs at a time, which structurally widens the gap. If that pool expands, the premium and the ADR price should compress together.

Inside the Business: 73% DRAM, 27% NAND

A common mistake is to read SK hynix as a pure HBM play; it is a full-line memory maker. Last quarter about 73% of revenue came from DRAM and 27% from NAND, with non-memory and foundry below 1%. HBM lives inside DRAM as a stacked, accelerator-adjacent memory that feeds the chip so it does not stall. That slice carries the growth and margin story, but DRAM also includes conventional server, phone and graphics memory that prices like a commodity with normal cycles. On the NAND side the headline is enterprise solid-state drives for data centers. The Solidigm unit, bought from Intel for $9 billion, offers a flagship drive at 122 terabytes and makes SK hynix the second-largest enterprise SSD supplier after Samsung. Because every trained model, every dataset and every checkpoint sits on NAND, the enterprise SSD market more than doubled in a single quarter and SK hynix's NAND revenue more than quadrupled year over year on higher price and volume. It is a second AI hardware business in plain sight.

The sum was about 79 trillion won, roughly $55 billion of quarterly revenue, up 257% year over year. In scale, the NAND operation alone is about 60% larger than all of SanDisk, and the DRAM operation is about 25% larger than Micron. In HBM the firm remains the leader but its share slipped from 64% a year ago to about 50% as Samsung, first to mass-produce HBM4, jumped from 21% to 33% in a single quarter. Even so, SK hynix ranks among the top two globally in combined DRAM and NAND, so owning the stock is owning the memory layer of the AI buildout.

From Cyclical Commodity to Software-Like Margin

Memory was once the definition of a cyclical, low-margin business where the industry, not the maker, set price. When demand outran supply, price and margin spiked; when supply outran demand, both collapsed, and makers lived on forecasting skill. SK hynix booked about a $5.9 billion operating loss in the last major downturn in 2023. The equation has flipped because demand now grows far faster than supply. Last quarter the company reported about 79 trillion won of revenue and about 65 trillion won of operating profit, roughly $55 billion and $42 billion, for a 76% operating margin, the highest in its history. That is a profile you expect from a software startup, not a forty-three-year-old hardware maker. The driver is contracting. While conventional DRAM still floats with the commodity cycle, HBM is built to spec and sold on pre-negotiated multi-year agreements where volume and price are settled before manufacturing. SK hynix locked in terms with roughly ten major customers for five years or more, all with upfront deposits. As contracted revenue rises, sales and margins become more predictable, which holders like, but the same structure caps upside: when memory prices jumped 30 to 60% last quarter, the firm, more contracted than Samsung or Micron, captured less upside and missed operating-profit expectations, sending the shares down close to 10% despite record results. Protection on the way down costs participation on the way up.

Why supply stays tight is explained with a fuel-line analogy. A modern accelerator spends a meaningful share of time waiting for data; if the GPU is the engine, HBM is the fuel line. HBM needs about three times the wafers of standard memory, yields are lower, and the specialty fabs take four to five years to build, so you cannot patch capacity like software. SK hynix is fully integrated, doing its own wafer fabrication, packaging and test, stacking dies into HBM cubes before TSMC places them next to Nvidia GPUs. It even broke ground on its first US packaging plant, with production scheduled for 2029. On the demand side, Nvidia lifted its procurement pledges to about $279 billion from roughly $119 billion within a single quarter to lock in memory. Hyperscaler commentary from both SK hynix and Micron points to HBM, DRAM and NAND staying tight for years because the designs, factories and processes must be aligned years in advance.

Roadmap and Co-Writing with Nvidia

On technology, SK hynix shipped HBM4 in volume last quarter and sampled the next iteration, HBM4E, in June. Each stack holds 48 gigabytes across 12 layers, running up to 16 gigabits per second per pin for about 4 terabytes per second per stack, enough in the video's shorthand to move roughly three hundred full-length 4K movies to the accelerator every second. On top of raw bandwidth, the new generation is described as more than 20% more power efficient and about 17% more heat resistant, so it runs cleaner and cooler. Because a memory generation must keep pace with GPU and network speeds, the roadmap, the fab and the process have to be choreographed years ahead, which is why supply constraints do not vanish overnight.

Strategy now choreographs that future with Nvidia. On June 7 the two announced a multi-year co-development pact spanning four Nvidia platforms — Vera Rubin AI supercomputers, Vera CPUs, RTX Spark PCs and Jetson Thor for robotics — followed six weeks later by a letter of intent covering memory supply and AI factories valued north of $500 billion. The first Vera Rubin AI factory powered by SK hynix HBM4 is slated for 2027, and while Nvidia has qualified all three memory houses for Vera Rubin, SK hynix is expected to supply about two thirds of the HBM4 this year. The Computex 2026 anecdote where Jensen Huang wrote please make more on an HBM4E wafer at the SK hynix booth and signed it is told as a symbol: this is not a vendor filling orders but a co-development where Nvidia designs around what next-generation memory can deliver and SK hynix designs memory to Nvidia's roadmap years before silicon ships.

The Valuation Terrace: Micron and SanDisk Side by Side

For valuation, Alex builds a terrace and shows SK hynix twice — the Korean price and the NASDAQ ADR price as separate rows — alongside Micron and SanDisk, apples to apples as far as fiscal calendars and currency allow. On size, Micron is around a trillion, SanDisk about a quarter of that, SK hynix about $920 billion in Seoul and $1.3 trillion as an ADR. On growth, SanDisk leads at 372% year over year, Micron at 346% and SK hynix at 257%; the lower rate is from the largest base, so the absolute gain remains huge. On profitability, the trio posts software-like operating margins: roughly 80% for Micron, 78% for SanDisk and 76% for SK hynix. On earnings, the distortion is Kioxia. SK hynix invested about $3.5 billion in 2018 in the Toshiba-spun NAND maker; Kioxia's spring surge forced a $45 billion paper gain that inflates net income, so Alex strips it out. Adjusted, SK hynix trades around 12 times earnings in Seoul and around 16 times as an ADR, still below Micron and SanDisk above 20. On forward earnings, SanDisk sits near 7.3, Micron near 6.8, SK hynix near 4.2 in Seoul and 5.8 as an ADR, about a 15% discount to Micron even after the premium. The company itself backs the cheapness claim with a 40 trillion won, about $29 billion, buyback, explicitly calling the Korean shares undervalued in an SEC filing — not the ADR. For access, Interactive Brokers has offered the Seoul line since May; otherwise the ADR remains comparatively cheap versus peers, with the clear caveat that if more ADRs are created the premium can close and the ADR price can fall. Risks stay where the video left them: Samsung's HBM4 catch-up, the upside cap of long contracts and the reliance on mid-70s margins.