AI · Semiconductors · Emerging Tech

SK Hynix Just Listed on Nasdaq — Here Is the Chip That Powers the World’s AI Systems and Why You Have Never Heard of It

SK Hynix raised $26.5 billion on Nasdaq last Thursday in the largest US listing ever by a foreign company. Regular trading under ticker SKHY started today. The company makes the memory chip that the majority of the world’s AI infrastructure depends on — and most people have no idea it exists.

What Just Happened

On July 10, 2026, SK Hynix — a South Korean semiconductor company most people outside the tech industry have never heard of — completed the largest US stock market debut by a foreign company in history. It sold 177.9 million American Depositary Receipts at $149 each, raising $26.5 billion. That surpasses Alibaba’s roughly $25 billion New York debut in 2014, which held the record for over a decade. The stock opened at $170 on day one and closed up 13% at $168.01. Regular trading under the permanent ticker SKHY began this morning, July 13.

Today the US-listed SKHY ADRs fell around 7–10% on Nasdaq — Seoul-listed Korean shares dropped more sharply at 15.4% — after a South Korean brokerage report suggested the company’s operating profit for the current quarter may miss estimates. That kind of early correction after a heavily anticipated listing is not unusual, and it does not change the underlying story of what SK Hynix actually is or why it matters. So let’s back up and explain that properly, because the financial coverage has mostly skipped it.

SK Hynix’s rise to a $1 trillion valuation — context on the company’s role in the AI memory supply chain. Published May 2026. Independent production, not affiliated with Wangdoo.

What SK Hynix Actually Makes

SK Hynix manufactures three main categories of semiconductor products: DRAM, NAND flash, and High-Bandwidth Memory. The first two you probably know by another name — the RAM in your laptop and the storage in your phone are variants of these. The third is the one that matters for AI, and it is the reason this company went from being largely unknown outside South Korea to a $1 trillion market capitalisation in under three years.

High-Bandwidth Memory, or HBM, is a type of computer memory built specifically for workloads that need to move enormous amounts of data between memory and a processor extremely quickly. Standard laptop RAM sends data along a 64-bit channel. HBM3E — the current generation in production — uses a 1,024-bit channel. SK Hynix’s HBM3E delivers over 1.1 terabytes per second of memory bandwidth per stack. HBM4, now entering production for Nvidia’s next-generation Rubin platform, doubles that to 2,048 bits, delivering up to 2 terabytes per second of memory bandwidth per stack. To put that in terms that land: standard DDR5 memory manages roughly 50-60 gigabytes per second. HBM4 is doing 2,000. That is not a marginal improvement — it is an architectural difference in kind.

AI training and inference models require constant, high-speed transfers of data between memory and compute. The GPU does not store the data it is processing — it pulls it from memory, computes, writes it back, and pulls more. If the memory cannot keep up, the GPU sits idle waiting. HBM exists specifically to prevent that bottleneck. Without it, the GPU utilisation rates that make large language model training economically viable would collapse.

Why It Cannot Be Replaced Once Built

This is the technical detail that most coverage of the SKHY listing has glossed over, and it is the detail that explains a great deal about SK Hynix’s market position.

HBM is not a chip you insert into a slot. It is not a module you can swap like the RAM sticks in a desktop computer. HBM is physically bonded to the GPU or AI accelerator at the time of manufacture using a process called CoWoS — Chip on Wafer on Substrate — which is a form of advanced packaging developed by TSMC. In CoWoS, the HBM stacks and the GPU die are placed side by side on a silicon interposer and bonded together into a single package. The process involves applying a moulded underfill compound across the assembled package under heat — SK Hynix’s proprietary version of this process is called Mass Reflow Molded Underfill, or MR-MUF, which reduces thermal resistance across the stack by approximately 17% compared to standard methods.

Once assembled, this package cannot be disassembled without destroying both components. The HBM is not a component inside the chip — it is co-equal with the GPU die, bonded at the silicon level. SK Group Chairman Chey Tae-won was explicit about this when he rang the Nasdaq opening bell on July 10: if SK Hynix’s HBM were substituted in a deployed system, the AI accelerator may not function correctly. That is a statement of manufacturing physics, not marketing.

How SK Hynix Came to Control This Market

SK Hynix produced the world’s first HBM chip in 2013, when the architecture was still an emerging standard co-developed with AMD and ratified by JEDEC that same year following a joint proposal by AMD and SK Hynix in 2010. The first GPU to use HBM was AMD’s Fiji in 2015. Nvidia adopted HBM2 in its Tesla P100 data centre GPU in 2016, and from that point on the company became the primary customer driving HBM demand.

SK Hynix’s lead over Samsung in HBM is not primarily a fabrication advantage — both companies can manufacture advanced DRAM at comparable process nodes. The lead comes from a combination of packaging process maturity and customer qualification history. Nvidia’s HBM validation process is lengthy and demanding: it tests signal integrity, thermal reliability, and manufacturing yield at scale across millions of cycles. SK Hynix has been through that process more times and for longer than its competitors, which means its products qualified first for each new generation and its engineers understand Nvidia’s specific requirements at a level of depth that takes years to accumulate.

Samsung fell behind in HBM3E qualification — the generation currently powering Nvidia Blackwell GPUs — due to thermal management issues with its packaging process. It has since caught up, and both companies are now qualified for HBM4. But the ground lost during the HBM3E generation explains why SK Hynix’s revenue market share reached 56.4% while Samsung’s sits at around 21%.

The Economics — Why the Margins Are So High

SK Hynix posted an operating margin of approximately 71% in Q1 2026. For context, Apple — routinely cited as one of the most profitable hardware companies in the world — operates at roughly 30% margins. SK Hynix’s margins are exceptional even within semiconductors, and they come from two compounding constraints.

First: HBM requires approximately three times the wafer capacity of standard DDR5 DRAM to manufacture, because stacking DRAM dies, forming through-silicon vias, and assembling the CoWoS package are each complex additional process steps. That conversion ratio limits total supply. When SK Hynix devotes production capacity to HBM, that capacity is unavailable for conventional memory — and its entire 2026 HBM allocation is already committed. Nvidia alone is estimated to consume roughly 60% of SK Hynix’s HBM output.

Second: because HBM is physically bonded into deployed accelerators, there is no spot market for the product in operational systems. Customers cannot switch suppliers mid-cycle. They sign long-term agreements and pay premium prices because there is no alternative for the systems they are building. The result is a pricing dynamic that looks more like a sole-source industrial contract than a commodity chip market.

Where the $26.5 Billion Goes

The entire raise is committed to manufacturing expansion rather than general corporate purposes. SK Hynix has identified three primary uses. The largest share funds Phase 1 of the Yongin Semiconductor Cluster in South Korea — a fabrication campus whose total projected cost runs to approximately $390 billion across multiple phases and decades. A second portion funds a new advanced packaging plant in Cheongju. The third funds EUV lithography equipment from ASML — the Dutch company that manufactures the only machines capable of patterning the most advanced chip layers, and which has a waiting list measured in years.

SK Hynix is also building its first US production facility — a $4 billion advanced packaging plant in West Lafayette, Indiana, scheduled for completion in 2028. The facility qualifies for up to $458 million in CHIPS Act funding from the US government. An expansion of its Solidigm enterprise storage business near Sacramento, California, is also underway. The Indiana plant is significant because advanced packaging — the CoWoS bonding process — is currently concentrated almost entirely in TSMC’s facilities in Taiwan, which creates a supply chain concentration risk that both the US government and SK Hynix’s customers want reduced.

HBM4 — What Comes Next

The current generation in volume production is HBM3E, which delivers over 1.1 terabytes per second of bandwidth per stack across a 1,024-bit interface at up to 9.8 gigabits per second per pin. HBM3E is the memory in Nvidia’s current Blackwell GPU generation.

HBM4, standardised by JEDEC in April 2025, doubles the interface width to 2,048 bits and targets 2 terabytes per second per stack. SK Hynix has been delivering HBM4 samples to Nvidia since late 2025 and is targeting mass production in Q3 2026 for Nvidia’s Rubin platform. The 12-layer HBM4 stack carries 48 gigabytes of capacity per stack — compared to 36 GB in HBM3E — and uses a logic base die manufactured on TSMC’s 12nm process, which means the memory stack itself now contains programmable logic that can be customised for specific AI workloads. That is a meaningful architectural change, not just a speed upgrade.

Samsung has also qualified for HBM4 supply to Nvidia and is shipping early samples. Its approach differs from SK Hynix’s: Samsung uses its own 4nm foundry process for the logic base die and handles the entire stack — logic, memory, and 3D packaging — internally, making it the only HBM4 supplier with a fully vertically integrated process. Whether that proves an advantage or a complexity in yield management will become clearer as volume production ramps.

My Take — Mr Wangdoo

I have covered a lot of semiconductor stories and most of them require you to care about spec numbers in ways that only matter if you are actually buying the hardware. This one is different, because HBM is not a product category — it is a structural bottleneck in the global AI supply chain, and understanding that changes how you read practically every other AI story.

When you read that a hyperscaler is ordering another 200,000 Nvidia GPUs, you are really reading a story about HBM supply. When you read that an AI startup cannot get enough compute, the compute shortage is in significant part a memory shortage. When Nvidia announces a new GPU generation, the question of whether it ships on schedule is partly a question of whether SK Hynix or Samsung can produce sufficient qualified HBM4 at the required yield rates. The chip inside the chip is the constraint that governs the pace of everything else.

The Nasdaq listing does not change any of that. What it changes is access. US retail and institutional investors can now hold SK Hynix as directly as they hold Micron or Nvidia, without navigating the Korean exchange infrastructure. That accessibility — what analysts call closing the Korea Discount — is the thesis behind the $149 IPO price. Whether the 13% first-day gain and subsequent drop this morning represent the market finding a fair price or an IPO premium correcting is a question I will leave to people with financial licences. What I will say is that the underlying business — a company whose product is physically bonded into deployed AI infrastructure and cannot be substituted — occupies a structural position in the semiconductor industry that very few companies hold. Most chip companies sell into a market. SK Hynix, for the accelerators already in the field, effectively is the market.

Frequently Asked Questions