Every time someone fires up ChatGPT, trains a new AI model, or uses a generative AI tool at work, something happens in the background that almost nobody talks about. Data moves. Fast. Enormous amounts of it, flowing between processors at speeds that were basically science fiction a decade ago. And the chips that make that possible? Odds are, they were made in South Korea.
I've been following the Korean semiconductor story for a while now, and honestly, the numbers from the last twelve months have been jaw-dropping even for me. SK Hynix beat Samsung in annual operating profit for the first time ever in 2025. Samsung shipped its first HBM4 chip in February 2026 — the most advanced AI memory chip on the planet — while SK Hynix was simultaneously building a $12.9 billion packaging plant in Cheongju. And all of this is happening while the world's biggest tech companies can barely get enough of these chips to meet demand.
If you want to understand where artificial intelligence is actually going, you need to understand South Korea's memory chip industry. This is that guide.
📋 Table of Contents
- What Is HBM and Why Does AI Need It So Badly?
- SK Hynix: How an Underdog Became the AI Industry's Most Wanted Supplier
- Samsung's Comeback: HBM4 and the Fight for Rubin
- The Numbers Don't Lie: HBM Market Data That Will Surprise You
- Why South Korea? The Structural Reasons Behind the Dominance
- The HBM4 Race: What Happens Next
- What This Means for Korea's Future (and Mine)
- FAQ: South Korea AI Chips — Your Questions Answered
What Is HBM and Why Does AI Need It So Badly?
Let me put this simply before we get into the data. A standard memory chip stores information in one flat layer. HBM — High Bandwidth Memory — stacks multiple layers of those chips on top of each other, like a tiny skyscraper, then connects them through microscopic vertical tunnels called Through-Silicon Vias (TSVs). The result is a chip that can move data radically faster while consuming less power.
Why does AI care? Because training a large language model like GPT-3 requires over 350 GB/s of memory bandwidth. Standard DDR5 DRAM just can't supply that. HBM can. It's not a preference — it's a physical requirement for modern AI at scale.
HBM2e memory costs about 5x more per GB than standard DDR5, which tells you something about both the technical complexity and the premium customers are willing to pay. NVIDIA doesn't stick HBM into its H100, H200, or upcoming Rubin accelerators because it's trendy. It's because nothing else works.
✍️ Personal Take #1
The first time I really understood the HBM story was when I read about hyperscalers securing only 70% of their server DRAM orders even after accepting 50% price hikes. Think about that. Amazon, Google, Microsoft — companies that negotiate billion-dollar deals for breakfast — couldn't get the chips they wanted and paid a massive premium anyway. That's not a supply chain hiccup. That's a structural shortage created by one of the most complex manufacturing processes on earth. And two Korean companies are essentially the gatekeepers.
SK Hynix: How an Underdog Became the AI Industry's Most Wanted Supplier
For most of its history, SK Hynix played second fiddle to Samsung. In memory chips, Samsung was the gold standard — bigger, better-funded, with manufacturing scale that seemed untouchable. Then AI happened, and SK Hynix rewrote the story completely.
In Q2 2025, SK Hynix did something that had never happened before: it surpassed Samsung as the world's largest memory chip supplier by revenue, posting $15.1 billion in quarterly sales. The same quarter, it held a 62% share of the global HBM market. For context, Samsung was at 17% and US-based Micron at 21%.
The reason comes down to a single relationship: NVIDIA. SK Hynix locked in as NVIDIA's primary HBM supplier years before the AI boom, building the kind of trust-in-yields that only comes from consistently hitting production targets at scale. By Q3 2025, SK Hynix's quarterly operating profit had surpassed 10 trillion won ($7+ billion) for the first time ever, with revenue up 39% year-on-year and operating profit surging 62%. Goldman Sachs called it explicitly: SK Hynix is "clearly an outstanding AI winner in Asia."
The 2026 supply situation is even more striking. SK Hynix's entire 2026 HBM supply was already sold out before the year even began — a detail that should tell you everything about how tight this market is. The company is now investing 19 trillion won ($12.9 billion) in a new advanced packaging plant in Cheongju, with completion targeted for end-2027.
For NVIDIA's next-generation Vera Rubin AI accelerators, industry sources say SK Hynix is expected to supply roughly two-thirds of all HBM4 demand — well above the ~50% originally expected. That's not just a market share number. That's a sign of just how deep the trust runs between these two companies.
✍️ Personal Take #2
What gets me about the SK Hynix story is that this wasn't luck. In conversations with people who follow the Korean semiconductor industry closely, the consensus is that SK Hynix made a strategic bet on HBM years before anyone knew AI would explode the way it did. They invested in a niche, technically brutal product when the market for it was still small. The analogy I keep coming back to: it's like a Korean restaurant in a neighborhood before it becomes trendy, building a loyal clientele while everyone else is still selling the same old thing. When the neighborhood blew up, they were already the institution. Except the neighborhood here is global AI infrastructure worth trillions.
Samsung's Comeback: HBM4 and the Fight for Rubin
Samsung's HBM story in 2024–2025 was uncomfortable to watch if you're rooting for them. The company fell behind SK Hynix in HBM generations, struggled to pass NVIDIA's quality tests for HBM3E, and saw its HBM market share collapse to as low as 17% in Q2 2025. For a company that had been the undisputed king of memory for decades, it stung.
But Samsung doesn't quit, and 2026 has been a different chapter. In February 2026, Samsung became the first company in the world to begin commercial shipments of HBM4 — the sixth generation of HBM technology. These aren't average chips either. Samsung's HBM4 delivers data transfer speeds of 11.7 Gbps, a 37% improvement over the JEDEC 8 Gbps standard and a 22% gain over its own HBM3E. The company achieved this by combining a 1c (10nm-class sixth generation) DRAM process with an in-house 4nm logic die — a technical combination nobody else has attempted at this scale.
Samsung also secured qualification with both NVIDIA and AMD for the Rubin and Instinct MI400 platforms. In Q4 2025, the company posted a record operating profit of 20.07 trillion won ($13.9 billion), driven by premium semiconductor products including HBM. And in March 2026, Samsung and AMD signed an MOU to expand strategic collaboration on HBM4 supply — a partnership that signals Samsung is fighting on multiple fronts, not just waiting for NVIDIA to call.
According to TrendForce projections, Samsung's HBM bit output share is expected to climb from 20% in 2025 to 28% in 2026 as its HBM4 production ramp accelerates. That's still behind SK Hynix, but it's a meaningful recovery from the lows of early 2025.
Unlike SK Hynix, which is almost entirely focused on memory, Samsung's memory segment alone generated operating profits of about 24.9 trillion won in 2025. The integrated ecosystem — designing, fabricating, and packaging within one organization — gives Samsung leverage that a pure-play memory company simply can't replicate.
✍️ Personal Take #3
There's something distinctly Korean about Samsung's response to the HBM crisis. In Korean culture, you don't just accept falling behind — you outwork the problem. Samsung's engineers reportedly ran parallel qualification tracks at 10 Gbps and 11 Gbps data rates simultaneously, which is the kind of all-in engineering commitment that most companies would never approve because of the resource cost. They were behind, they knew it, and they went harder rather than cutting losses. Whether you love or hate big chaebol, that particular quality — responding to adversity with total commitment — is genuinely impressive.
The Numbers Don't Lie: HBM Market Data That Will Surprise You
Numbers tell a cleaner story than any narrative, so here's what the data actually looks like for HBM and South Korea's position in it.
The HBM market grew 178% year-over-year in Q2 2025. Not 78%. 178%. That's not growth — that's a structural transformation of an entire market segment. Bank of America has characterized 2026 as a "supercycle similar to the boom of the 1990s," forecasting global DRAM revenue to surge 51% year-on-year and naming SK Hynix the industry's "Top Pick."
The BofA 2026 HBM market forecast stands at $54.6 billion — a 58% increase from 2025. Goldman Sachs separately forecasts that HBM demand for custom ASIC-based AI chips (think Google's TPUs and Amazon's Trainium) will skyrocket 82%, accounting for roughly one-third of the overall market. Looking further out, Micron projects the HBM total addressable market will reach approximately $100 billion by 2028 — a milestone analysts originally projected for 2030, now arriving two years early.
South Korea controls the center of this market. SK Hynix and Samsung together account for over 90% of global HBM supply. Asia-Pacific dominated the HBM market with a 41% share in 2025 and is growing at a 26.66% CAGR through 2031. Government subsidies in South Korea and Japan reduce fab costs by 20–40%, a structural advantage competitors in the US and Europe can't easily close.
Unlike standard consumer semiconductors where demand fluctuates with smartphone cycles, the HBM market now looks structurally different. SK Hynix management has stated explicitly that HBM supply is expected to remain tight even in 2027. The CEO of SK Hynix has publicly forecast annual HBM demand growth of roughly 60% in the medium-to-long term. These aren't hopeful projections from a company talking up its stock — they're supply plans already locked in with the world's biggest AI hardware customers.
Why South Korea? The Structural Reasons Behind the Dominance
This is the question I get asked most often when I explain the HBM story to people who aren't deep in tech: why South Korea? Why not Taiwan, Japan, or the US?
The short answer is decades of compound investment with no off-ramps. South Korea made a national bet on memory semiconductors in the 1980s and never stopped doubling down. Unlike logic chips — where design firms like Qualcomm and Apple can outsource fabrication to TSMC — memory manufacturing is uniquely dependent on vertical integration, process precision, and manufacturing culture that takes generations to build.
Samsung and SK Hynix have been competing intensely against each other for 40+ years inside the same domestic market. That sustained rivalry produced engineering talent, process innovations, and manufacturing disciplines that no other country has replicated. Compared to US chipmakers who historically outsourced commodity memory, Korean fabs kept every part of the process in-house. That decision looks genius in retrospect.
The South Korean government's role is also structural, not just symbolic. South Korea and Japan together have pledged more than $100 billion in semiconductor incentives to reduce capital intensity for new fabs by up to 40%. A government-scheduled expanded fabrication cluster in South Korea is set to open in 2027. These aren't press releases — they're funded commitments tied to specific construction timelines.
There's also a geographic concentration effect. Samsung's Pyeongtaek campus and SK Hynix's Icheon and Cheongju facilities sit within a few hours of each other. The talent ecosystem — engineers, equipment specialists, materials scientists — is intensely concentrated in a small area, creating network effects in innovation that a distributed industry can't easily reproduce.
✍️ Personal Take #4
I've visited the Pyeongtaek region a couple of times, and there's something quietly surreal about the area around Samsung's mega-fab. It doesn't look like the center of global AI infrastructure. Rice fields, small towns, then suddenly these enormous, immaculate white buildings stretching for kilometers. The cleanrooms inside are more sophisticated than most hospital operating suites. And yet this is where the chips that run the world's most powerful AI systems are literally born. Korea has this pattern where world-class industrial capacity hides inside surprisingly unassuming geography — K-drama fans who visit quiet neighborhood streets in Seoul feel the same way. The quiet exterior and the extraordinary interior performance.
The HBM4 Race: What Happens Next
The competition between Samsung and SK Hynix is entering its most technically demanding chapter yet with HBM4, and the outcome matters far beyond Korea.
Samsung began mass-production HBM4 shipments in February 2026, achieving 11.7 Gbps data rates with its 1c DRAM + 4nm logic die combination. SK Hynix, meanwhile, showcased at CES 2026 the industry's first 16-layer HBM4 stack at 48GB capacity, targeting mass production of this variant by Q3 2026. Samsung is leading immediate volume shipments while SK Hynix bets on density as the next frontier — two different strategies, both aimed at the same customer: NVIDIA's Vera Rubin platform.
By 2026, over 40% of advanced processors are expected to incorporate chiplet-based architectures, many leveraging HBM stacks to overcome data bottlenecks. HBM3E is projected to account for approximately two-thirds of total HBM shipments in 2026 while HBM4 gradually captures share. Looking at 2028, some forecasts suggest the HBM market size will surpass the entire DRAM market of 2024 — meaning HBM alone will be bigger than all memory combined just four years ago.
Both companies are also beginning to look beyond pure memory. Samsung's foundry division winning orders from Tesla signals a move into AI accelerator logic. SK Hynix's long-term collaboration agreement with Applied Materials (signed March 2026) focuses on next-generation DRAM and HBM development at their EPIC Center. The game is expanding from memory supply to deep involvement in the full AI chip ecosystem.
What This Means for Korea's Future (and Mine)
I find it genuinely moving, in an odd way, how much of the AI revolution runs through Korean hands. Every major announcement in AI infrastructure — NVIDIA's Blackwell, Rubin, Google's TPU expansions, Amazon's Trainium buildouts — every single one depends on chips that came out of facilities in Icheon, Pyeongtaek, and Cheongju. Korea doesn't always get the credit it deserves in the global AI narrative, which tends to focus on the US software companies and the Chinese competition. But dig into the supply chain and the Korean fingerprints are everywhere.
What I'm watching most closely in the next 12 months: whether Samsung's HBM4 speed advantage translates to stable mass-production yield. Samsung has a history of making technical breakthroughs and then struggling to scale them at the quality levels NVIDIA demands. If Samsung can fix the yield story while maintaining 11.7 Gbps performance, the HBM market gets genuinely competitive in a way it hasn't been since 2023. If SK Hynix locks in the Vera Rubin allocation at two-thirds and executes its 16-layer HBM4 ramp by Q3 2026, the gap widens again.
Either way, Korea wins. That's the real story.
❓ FAQ: South Korea's AI Chips — Everything You Actually Want to Know
What is HBM and why is it critical for AI?
HBM (High Bandwidth Memory) is a 3D-stacked memory technology that connects multiple DRAM layers vertically through microscopic tunnels called Through-Silicon Vias. Training a large AI model like GPT-3 requires over 350 GB/s of memory bandwidth — a demand standard DDR5 DRAM cannot meet. HBM solves this bottleneck and has become a non-negotiable component in every major AI accelerator from NVIDIA, Google, and AMD.
Who leads the global HBM market in 2026?
SK Hynix leads with approximately 57–62% market share (varying by quarter), followed by Samsung at 22% and Micron at 21%, according to Counterpoint Research. SK Hynix is expected to supply roughly two-thirds of NVIDIA's HBM4 for the Vera Rubin platform in 2026. Samsung began first-in-world commercial HBM4 shipments in February 2026 and is recovering market share.
How big is the HBM market and how fast is it growing?
The global HBM market is forecast by Bank of America to reach $54.6 billion in 2026, a 58% year-on-year increase. Goldman Sachs forecasts HBM demand for ASIC-based AI chips will grow 82% in 2026. Looking further out, Micron projects the market will reach approximately $100 billion by 2028 — two years ahead of previous forecasts. SK Hynix forecasts annual HBM demand growth of ~60% in the medium-to-long term.
Why does South Korea dominate HBM production?
South Korea's dominance comes from four decades of compounded investment in memory manufacturing, intense domestic competition between Samsung and SK Hynix, significant government subsidies that reduce fab costs by 20–40%, and a concentrated talent ecosystem around major fab clusters in Icheon, Pyeongtaek, and Cheongju. SK Hynix and Samsung together supply over 90% of global HBM output.
What happened when SK Hynix beat Samsung in 2025?
In 2025, SK Hynix beat Samsung in annual operating profit for the first time in the companies' history, driven by its HBM dominance. In Q2 2025, SK Hynix also briefly became the world's largest memory chip supplier by revenue at $15.1 billion, with a 62% HBM share. Samsung's memory segment still generated 24.9 trillion won in 2025, and Samsung reclaimed the top spot in overall DRAM revenue in Q4 2025.
What is HBM4 and who makes it?
HBM4 is the sixth generation of High Bandwidth Memory, designed for next-generation AI accelerators like NVIDIA's Vera Rubin platform. Samsung began commercial HBM4 shipments in February 2026, achieving 11.7 Gbps data transfer speeds using a 1c DRAM process with a 4nm logic die — 37% above JEDEC's 8 Gbps standard. SK Hynix is targeting mass production of its 16-layer 48GB HBM4 variant by Q3 2026. HBM4 is expected to become the dominant AI memory chip by 2027.
How does South Korea's semiconductor industry affect global AI development?
With SK Hynix and Samsung controlling 90%+ of HBM supply, and HBM being a non-substitutable component in major AI accelerators, South Korea effectively sits at a critical chokepoint in global AI infrastructure. Supply shortages in Korean fabs directly delay AI data center buildouts worldwide. As of 2025, many major hyperscalers could only secure 70% of their server DRAM orders even after accepting 50% price premiums.
🔑 3 Key Takeaways
- SK Hynix redefined the competitive order in 2025, beating Samsung in annual operating profit for the first time in history and holding 57–62% HBM market share — with its 2026 HBM supply already sold out before the year started.
- HBM4 is the new battleground: Samsung began world-first commercial HBM4 shipments in February 2026 at 11.7 Gbps, while SK Hynix targets a denser 16-layer 48GB variant by Q3 2026 — both racing for NVIDIA's Vera Rubin platform allocation.
- The market is only getting bigger: BofA forecasts $54.6 billion HBM market in 2026 (+58%), with some projections reaching $100 billion by 2028 — and South Korea controls 90%+ of supply.
The Quiet Superpower in Every AI Chip
AI gets talked about in terms of models, algorithms, and data. But strip away the software and what you find underneath is a physical manufacturing challenge of extraordinary difficulty — one that South Korea has spent four decades preparing to meet.
Samsung and SK Hynix didn't stumble into this position. They built it through engineering culture, national investment, and a competitive rivalry that pushed both companies to innovate faster than any single monopolist would have. The AI boom didn't create Korea's semiconductor leadership. It revealed it.
The HBM race in 2026 is genuinely fascinating to follow — real technical competition between two Korean giants with genuinely different strategies, playing out in the context of global AI infrastructure investment measured in hundreds of billions of dollars. I'll keep updating this as the story develops.
What do you think — in the HBM4 race, are you putting your money on SK Hynix's density strategy or Samsung's speed-first approach? Drop a comment below 👇
🔗 Explore More on All About K-Culture
- How Samsung Became a Global Tech Empire — The Full Story
- Made in Korea: Tech Dominance in Semiconductors, EVs & Defense 2026
- How Hyundai & Kia Conquered the Global Auto Industry
- Why Naver and Kakao Dominate Korea's Internet Ecosystem
Comments
Post a Comment