韩国相亲战服之王，SK海力士如何逆袭三星？

Original Author: TechFlow (Shen Chao) 

In Seoul's dating agencies, a peculiar new trend has recently emerged.

Some men attending arranged meetings deliberately tuck their business cards deep inside the inner pocket of their suit jackets. Only after confirming that their date has "good character" do they carefully pull out the card bearing their company name, which reads four English letters: SK Hynix.

Source: Korean Variety Show

Kang Eun-sun, a senior executive at the Korean matchmaking agency Gayeon, publicly told the media that since the start of the semiconductor super cycle, employees from Samsung Electronics and SK Hynix have seen their popularity soar. "The market clearly prefers engineers with significantly higher actual incomes over lawyers whose earnings have declined." A joke even circulating on social media reads: "When Hynix employees go on blind dates, they humbly say they work at Samsung. Only when they meet someone with good character do they confess they actually work at Hynix."

What turned a company uniform into a "dating armor" is a set of numbers that make workers worldwide green with envy.

In 2025, SK Hynix recorded an operating profit of 47.2 trillion KRW. According to a new agreement reached with the union in September last year, 10% of operating profit goes into the employee bonus pool. Based on 35,000 employees, each person could receive approximately 140 million KRW, equivalent to about 650,000 RMB.

In the first quarter of this year, SK Hynix's operating profit surged over 400% year-on-year to 37.6 trillion KRW. According to forecasts by analysts from various countries, this year's operating profit is expected to fall between 210 trillion and 250 trillion KRW. Based on this estimate, this year's average per-person bonus could range from 2.9 million to 3.3 million RMB.

International investment bank Macquarie Securities further predicts that if operating profit reaches 447 trillion KRW in 2027, the average per-person bonus could be as high as 1.29 billion KRW, approximately 6.1 million RMB.

More noteworthy than the "6.1 million RMB per person bonus" is this story: This company has long been the second fiddle in the Korean semiconductor industry, the little brother standing behind Samsung.

What did it do right to cause Samsung, whose screens and chips even Apple relies on, to fall from its position as the global memory leader, watching it happen helplessly?

2008: Recovery from the Brink of Bankruptcy

Rewind to 2008, and no one would have placed the title "future dominator" on Hynix.

Its predecessor was Hyundai Electronics. When the internet bubble burst in 2001, DRAM prices collapsed. The company was saddled with a massive debt of $14 billion. Taken over by its creditors, it entered a five-year "workout program," a state akin to "rehabilitation trusteeship" in the Korean context. For five full years, its factories, R&D budgets, and staffing were all tightly constrained.

By 2007, Hynix finally emerged from this "trusteeship" state, but it was weakened and barely surviving.

It was at this moment, on the other side of the Pacific, that a company called AMD came calling.

AMD's situation was also grim. It was the second player in the GPU market, struggling to breathe under NVIDIA's dominance in gaming graphics. Its researcher, Bryan Black, was working on a peculiar concept called "High Bandwidth Memory" (HBM). This involved stacking multiple DRAM chips vertically, like building a skyscraper, and connecting them using a technology called TSV (Through-Silicon Via).

Why do this? Because AMD saw a problem that few others had noticed: CPU/GPU computing speeds were getting faster, but the data transfer rate of memory couldn't keep up. The compute unit often finished a calculation and then just idled, waiting for the memory to send the next batch of data. The industry calls this bottleneck the "Memory Wall."

To use an imprecise analogy: Imagine a super chef who can prepare 10 dishes per second, but the food runner can only bring in 2 ingredients per second. The result is that the chef spends 80% of their time waiting for ingredients. No matter how fast the computing chip is, if data can't get in, it's just spinning its wheels.

AMD's idea was straightforward: Instead of widening the transport channel horizontally (the traditional DDR approach), why not make the memory "taller" by stacking vertically on the chip, allowing data to be sent in over a shorter distance using a wider bus? This vertically stacked "mini-skyscraper" is what HBM is.

The solution sounded great. The problem was that in 2008, there was no AI demand, no large model training, no so-called "computing power revolution." At the time, the only visible application for HBM was high-end gaming graphics cards. The market was small, the manufacturing process was extremely difficult, and the cost per chip was far higher than ordinary DRAM.

AMD looked around, but no one wanted to take it on. Samsung didn't want to; it was fully committed to the HMC (Hybrid Memory Cube) technical route, another vertical stacking solution it was developing with Micron. Micron wasn't interested either; it was following Samsung's lead.

The only one willing to take it on was Hynix, which had just recovered from the brink of bankruptcy and couldn't afford to be picky about any major orders.

In 2009, Hynix formally initiated HBM R&D. It wasn't until 2013 that the first HBM chip was born at Hynix's factory in Icheon, South Korea.

Back then, who could have imagined that 15 years later, this chip would become something all AI giants would be scrambling to buy, with production capacity "sold out through 2030"?

No one could have imagined it, not even Hynix itself.

2012: A Chairman's Gamble

In 2012, another key figure entered the story.

SK Group Chairman Chey Tae-won led a consortium formed through SK Telecom to acquire a 21.05% stake in Hynix from its creditors for approximately 3.4 trillion KRW (about $3 billion). This semiconductor company was henceforth known as SK Hynix.

What kind of person is Chey Tae-won? The Korean book Super Momentum describes the acquisition aftermath: In a short period, he held one-on-one meetings with 100 of Hynix's top executives. His first actions were not layoffs or cost-cutting. Instead, he integrated SK Group's management system with Hynix's technical capabilities and then reopened the funding taps for the FAB (wafer fabrication) investments and process improvement that had been paused.

The key here was timing.

From 2012 to 2014, the entire DRAM industry was still reeling from the memory downturn cycle of 2011-2012. All rational financial models were telling management, "This is an industry low point, be conservative." But Chey Tae-won made a contrarian decision: expand investment.

Even more critical was HBM.

Over the 11 years from 2011 to 2022, SK Hynix invested a cumulative total of approximately 860 billion KRW in HBM-related R&D and about 1.5 trillion KRW in facilities and equipment. A significant portion of these investments occurred during years when the market was sluggish and HBM's commercial prospects were completely invisible.

What happened during this period?

HBM2 failed to meet performance expectations and had to be scrapped and redone, resulting in a revised version called "HBM2 Gen2." The HBM team became a "department no one wanted to join" within the company; key members were rotated out, and morale was low. AMD's flagship graphics card, the R9 Fury X, used first-generation HBM in 2015, but its market reception was lukewarm because it was too expensive, and few bought it.

What made Hynix even more anxious was this: Around 2016-2017, Broadcom approached Samsung, wanting Samsung to supply HBM2 for Google's second-generation TPU. If Samsung could meet 100% of the demand, Broadcom promised exclusivity. This should have been a crucial moment for HBM's first entry into the data center.

But Samsung dropped the ball.

A report from Korea's JoongAng Ilbo reconstructed the chaos of that time: The Google TPU project involved three companies: Broadcom (design), Samsung (memory), and TSMC (foundry). Samsung's HBM experienced memory issues, and engineers reported that TSMC refused to let them enter the factory site for inspection. The three companies shifted blame, the problem remained unresolved, causing delays of up to six months. "This kind of stalemate was common between 2016 and 2017," an informed executive recalled later.

Google gradually leaned towards collaborating with SK Hynix. The biggest beneficiary of HBM's first real landing in the data center was not Samsung.

But while all this was happening, the world had no idea that HBM would become the most critical bottleneck of the AI era.

Chey Tae-won later said only this in an interview for Super Momentum: "We were standing at a crossroads."

He didn't elaborate on why he persisted. But in hindsight, the logic might be this: Hynix's fate had already skirted the edge of bankruptcy in 2001. It didn't have Samsung's diversified businesses to spread risk. Memory was the only thing it could do. Either it became the world's number one in this field, or it would live forever in Samsung's shadow.

So "betting on HBM" wasn't a choice for Hynix; it had no choice.

2022: Jensen Huang Passes the Match

In June 2022, SK Hynix began mass production of HBM3. That year, the first batch of HBM3 chips was installed on a GPU called the H100, from a company called NVIDIA. At that time, NVIDIA was valued at around $300 billion and was considered "important but not world-changing" in both the gaming GPU and data center markets.

In November, OpenAI released ChatGPT.

What happened next is well-known. The demand curve for AI computing power transformed from a gentle slope into an almost vertical rocket trajectory. Every GPU used to train large models needed HBM as its closest "data porter."

At that moment, the 14-year gamble Hynix had been making was finally revealed.

By the second quarter of 2025, Hynix had captured 62% of the global HBM market share. Samsung had fallen to 17%, trailing even the latecomer Micron (21%).

For the full year 2025, Hynix's operating profit was 47.2 trillion KRW, compared to Samsung Electronics' 43.6 trillion KRW. This was the first time in Hynix's history that its annual profit surpassed Samsung's.

During the 2025 Computex exhibition in Taipei in August, NVIDIA CEO Jensen Huang visited SK Hynix's booth and left a handwritten message on the display board: "JHH LOVES SK HYNIX!" This photo was subsequently circulated widely by Korean media. In engineering culture, there is no more direct form of official endorsement.

And SK Hynix's engineers, among themselves, have given HBM a new full name. They say HBM actually stands for "Hynix Best Memory".

Where Did Samsung Actually Lose?

So the question is, where did Samsung, the company that once crushed Japanese DRAM manufacturers and cornered Micron, actually lose?

Samsung missed the early HBM layout, made the wrong choice of technology roadmap (NCF packaging vs. Hynix's MR-MUF packaging), and its HBM3E kept failing NVIDIA's certification tests...

These are all facts, but they are not the root cause. The root cause is something more awkward and ironic: Samsung was too successful, so it couldn't afford to lose or gamble intensely.

Compare the two companies at that point in 2008. Hynix had just crawled out of bankruptcy trusteeship. The entire company had only one business line—memory. It had no diversified cash cow to rely on.

It needed to bet on HBM, not because its vision was exceptionally sharp, but because it had no other choice. Any opportunity to shed the label of "the little brother behind Samsung" was one it had to seize.

And Samsung?

In 2008, Samsung was on the eve of its peak. Its mobile phone business was about to take off with the Galaxy series. Its semiconductor business was number one in the world for both DRAM and NAND. Its display business was about to land the massive OLED order for Apple's iPhone. Its cash flow was incredibly abundant, its business empire was vast, and the stakeholders it needed to balance were extremely complex.

For a company like this, what was HBM in 2008? It was a high-risk gamble with a tiny market, an extremely long return cycle, and one that conflicted with another technical route (HMC) it was actively promoting. No rational finance committee would approve going all-in on this venture.

This is the classic "Innovator's Dilemma": Successful large companies are always constrained by their own success. The markets they have already won are too large, too important, and require too much protection for them to be able to bet their entire future on an unpromising new direction, unlike a cornered challenger.

The deeper irony is that Samsung didn't fail to see HBM. It started investing in HBM-related research as early as 2011 and was even the first to mass-produce HBM2 in 2016. But every time, Samsung wasn't all-in. Its energy was simultaneously spread across over a dozen fronts like HMC, GDDR, LPDDR, and enterprise SSDs. While Hynix's HBM team was "marginalized but kept pushing," Samsung's HBM team was also "marginalized," but no one there was fighting to keep it alive.

By 2024-2025, Samsung finally realized it had to go all-in on HBM, but it was too late. The gap in technology and process had already formed. The moat in customer relationships had been jointly built by NVIDIA and Hynix.

Jun Young-hyun, Vice Chairman of Samsung Electronics' Semiconductor Business, said in his 2026 New Year's address: "Our customers tell us, Samsung is back."

The word "back" itself was a form of self-admission.

Two Questions

What does the story of SK Hynix mean? At least two questions are worth considering.

First, why do similar stories happen more easily in South Korea, rather than elsewhere?

SK Hynix's success didn't come from nowhere. It rests on a unique industrial foundation. The Korean chaebol (conglomerate) system, though criticized for decades, objectively allows a company, driven by a single decision-maker's will, to bet on a gamble with a 20-year return