There's a story I keep thinking about from a few years back — before ChatGPT, before the AI gold rush, before Jensen Huang became the most photographed CEO in Seoul.
I'd heard someone mention HBM memory at a tech meetup. High Bandwidth Memory. Stacked chips, crazy fast data transfer, made by Samsung and SK Hynix. I nodded like I understood, then quietly filed it under "probably niche stuff that'll never go mainstream." It sounded almost too specialized to matter outside of supercomputers and research labs.
I was spectacularly wrong.
Today, HBM is the single most important component in the global AI supply chain. Every ChatGPT query, every Gemini image generation, every AI video tool you've ever used — none of it happens without Korean memory chips. And right now, the entire world doesn't have enough of them.
Jensen Huang flew to Seoul in June 2026 for five days. He ate galbi. He drank soju. He signed a wafer at SK Hynix's Computex booth with the words: "Please Make More."
That sentence tells you everything.
📋 TABLE OF CONTENTS
What Is HBM Memory, Actually? {#what-is-hbm}
Let's be honest — "High Bandwidth Memory" sounds like something you'd find buried deep in a technical whitepaper, not in a lifestyle blog about Korean culture. But stick with me, because this is genuinely fascinating once you get past the jargon.
Regular computer memory — the DDR5 RAM inside your laptop — works like a two-lane highway. Data moves quickly, but the road is narrow. Standard DDR5 reaches around 4.8 to 6.4 gigabytes per second of transfer speed, which is fine for everyday computing.
HBM works on an entirely different principle. Instead of spreading memory chips flat across a circuit board, HBM stacks multiple DRAM dies vertically — like floors of a building — and connects them with thousands of tiny copper columns called through-silicon vias, or TSVs. Think of it as a high-rise building of chips, with elevators connecting the floors. That wide, stacked architecture creates an enormous data pathway. A single HBM stack delivers between 3 and 5 terabytes per second of bandwidth, compared to just 50 to 100 gigabytes per second for standard DDR5. That's not a slight improvement. That's roughly 50 times faster.
The reason AI needs this? Frontier AI models like GPT-4 or Gemini don't just process information — they process staggering amounts of it simultaneously. Every time an AI model runs a query, it needs to shuttle billions of numbers back and forth between the processor and memory at incredible speed. Slow memory means the most powerful GPU in the world just... sits there waiting. Without sufficient HBM, even the most powerful GPU idles, waiting for data.
This is why HBM went from "interesting niche component" to "geopolitical asset" in about three years.
Why NVIDIA Can't Build AI Without It {#why-nvidia}
NVIDIA's AI accelerator chips — the H100, the B200, and now the upcoming Vera Rubin — are essentially useless without HBM sitting right next to them on the package. Every major AI chip, from NVIDIA's H100 to AMD's MI300X to Google's TPUs, depends on HBM.
Here's what the numbers look like in practice. NVIDIA's next-generation Vera Rubin AI accelerator uses eight HBM4 stacks per processor. Now multiply that by the millions of GPUs being deployed across data centers globally, and you start to understand the scale of the demand problem.
NVIDIA CEO Jensen Huang confirmed at the GTC Taipei 2026 keynote on June 1 that the Vera Rubin platform is now in full production, naming Samsung Electronics, SK Hynix, and Micron as its HBM4 memory suppliers. All three companies are in a full sprint to meet demand that is growing faster than anyone projected.
How fast? The supply-demand imbalance for HBM is estimated at 50 to 60% for 2026 — meaning demand is roughly double what suppliers can currently produce. Every major supplier has already sold out its entire 2026 allocation. SK Hynix confirmed it has already sold out its planned supply of memory chips for 2026.
When Jensen Huang says "please make more," he's not being polite. He's describing a genuine bottleneck that is limiting the pace of global AI development.
Worth Noting: Here's something that caught me off guard when I first dug into this: the GPU shortage everyone panicked about in 2023 and 2024? That was just the opening act. In 2026, the bottleneck has shifted from silicon to memory — specifically HBM. NVIDIA can theoretically build more GPU chips. What they can't easily do is conjure more HBM out of thin air, because the manufacturing process is extraordinarily complex and capacity takes years to build. The crisis moved upstream, and most people haven't noticed yet.
How Korea Came to Own 90% of Global HBM Supply {#korea-dominance}
This isn't an accident of geography. It's the result of decades of calculated investment in semiconductor manufacturing that most of the world dismissed as too expensive, too risky, and too far ahead of its time.
Samsung and SK Hynix together control 90% of global HBM production. The third player, Micron of the United States, covers most of the remaining 10%. There are no other meaningful competitors. This means two Korean companies hold the keys to the most critical component in the global AI infrastructure race.
SK Hynix, based in Icheon, Gyeonggi-do, was actually the first company to mass-produce HBM chips commercially. They've maintained a lead in each successive generation — HBM2, HBM3, HBM3E — and are currently first to market with HBM4 samples. Samsung, with its massive fab capacity in Pyeongtaek, has been closing the gap fast. Samsung shipped HBM4 production units for the first time in the industry in February 2026, and has since shipped HBM4E seventh-generation samples, while also unveiling an HBM5 prototype at Computex 2026 in Taipei.
The manufacturing complexity is part of why no new competitor can simply enter this market. HBM is not simply regular DRAM packaged differently — it requires advanced 3D stacking technology, through-silicon vias, and extremely precise manufacturing processes that make it more expensive and more time-consuming to produce than conventional memory. Building a competitive HBM fab from scratch takes five to seven years minimum, which is why the Korean duopoly isn't going anywhere soon.
Korea's semiconductor investment confirms this strategic position. Seoul has committed roughly $470 billion to expand its semiconductor industry, aiming to strengthen its position in global supply chains.
Jensen Huang in Seoul: What the Visit Really Meant {#jensen-visit}
I want to talk about this visit specifically, because it happened literally this week and it tells you more about Korea's AI position than any earnings report could.
NVIDIA CEO Jensen Huang made a four-day trip to Seoul, South Korea, securing critical supply chain agreements spanning advanced memory chips, gigawatt-scale data centers, and autonomous mobility — including deals with SK Hynix, SK Telecom, Naver, Doosan, LG Group, and Hyundai Motor.
Huang dined on grilled pork belly and local soju with the country's top corporate bosses, threw a baseball pitch, and met with a well-known gamer. This wasn't a quick supplier check-in. This was a full immersion — the kind of visit a CEO makes when a country has become genuinely indispensable to their business.
On the SK Hynix partnership, Huang stated that SK Hynix has been and will continue to be NVIDIA's largest memory partner — a relationship generating billions of dollars annually, with that figure expected to grow substantially.
At the Korea AI Ecosystem Reception held at the Shilla Seoul hotel, Huang also noted that SK Hynix had announced plans to double its memory production capacity by 2030 — and even that wasn't enough. Huang was effectively asking for more.
The visit ended with Huang declaring that Korea is an "irreplaceable" link in the global AI supply chain — with strengths extending beyond HBM into robotics, automotive, and platforms.
Been There (Sort of): I remember hearing the name "SK Hynix" years ago in a completely different context — memory sticks for computers, back when RAM was just RAM. Nobody was talking about them as a strategic AI asset. The same company that made the chips in your old laptop is now the company that NVIDIA's CEO flies across the Pacific to personally ask for more product. That trajectory is genuinely astonishing, and I think Koreans don't always appreciate how significant that shift is on a global scale.
The HBM4 Race: Samsung vs SK Hynix vs Micron {#hbm4-race}
Right now, HBM4 is where the competition is fiercest — and it's basically a two-Korean-company race with Micron trying to keep up.
NVIDIA has already signaled interest in 16-layer HBM for delivery as early as late 2026, forcing Samsung, SK Hynix, and Micron into a race to meet the requirements of next-generation AI accelerators. The engineering challenges involved are enormous — 16-layer HBM hasn't been commercialized previously, with major technological hurdles including DRAM stacking and wafer thickness at extreme scale.
SK Hynix has the early lead. SK Hynix's 2026 capacity is fully booked, with revenue hitting 24.45 trillion won ($17.13 billion) in one quarter — a 39% year-on-year increase, with quarterly profit exceeding 10 trillion won for the first time.
Samsung is closing the gap aggressively. Industry forecasts suggest Samsung's HBM4 market share could expand to around 35%, exceeding the initial estimate of 25%.
Meanwhile, Samsung and SK Hynix signed a letter of intent with OpenAI for the supply of 900,000 DRAM wafers per month to support the Stargate project. OpenAI's Stargate data centers — the infrastructure powering the next generation of AI — will run on Korean memory.
The numbers on pricing tell a story too. Server DRAM prices have surged 60 to 70% for 2026 compared to the previous year. AI's demand for memory has genuinely restructured global pricing.
The Part Nobody Warns You About: Your RAM Prices {#consumer-impact}
Okay, this is the part that affects you directly — and the part that makes me genuinely annoyed every time I check prices on consumer memory.
Here's what's happening: manufacturing HBM is so complex, and so profitable, that Samsung and SK Hynix have been diverting wafer capacity away from regular consumer DRAM to produce it. DRAM prices in early 2026 have experienced compounded increases, some exceeding 200% since early 2025, because HBM demand from the AI sector is crowding out commodity DRAM capacity — with a 3-to-1 conversion ratio between HBM and DDR5 wafer capacity meaning every HBM ramp directly compresses general-purpose memory supply.
Gaming GPU production faces 40% cuts as a direct result of HBM's manufacturing demands on overall wafer capacity. If you've been wondering why upgrading your PC has gotten so expensive lately, AI infrastructure is literally part of the answer.
Real Talk: I buy RAM and storage through Coupang fairly regularly. Over the past year I've watched prices creep up in ways that felt random but weren't — and this is why. Every time someone runs a ChatGPT query, every time an AI video gets generated, every time a data center adds another server rack, the demand chain pulls on the same semiconductor capacity that makes your laptop RAM. The AI boom has a direct cost that most consumers are paying without realizing it. When I first connected those dots, it bothered me more than I expected. These are Korean companies building globally essential technology — and Korean consumers are among the ones absorbing the price increase on the consumer side.
FAQ {#faq}
What is HBM memory and why does it matter for AI? HBM (High Bandwidth Memory) is a specialized type of DRAM that stacks multiple memory dies vertically, delivering data transfer speeds of 3 to 5 terabytes per second — roughly 50 times faster than standard laptop RAM. AI chips like NVIDIA's H100 and Vera Rubin require HBM because they process billions of calculations simultaneously and need memory that can feed data fast enough to keep up.
Which Korean companies make HBM? Samsung Electronics and SK Hynix are the two dominant producers, together controlling approximately 90% of global HBM supply. SK Hynix has led in HBM technology across most generations; Samsung is closing the gap with HBM4 and HBM5 development in 2026.
Why did Jensen Huang visit South Korea in June 2026? Huang made a five-day visit to Seoul to solidify supply chain agreements with Samsung, SK Hynix, and other Korean conglomerates. He secured a multi-year HBM supply deal with SK Hynix and discussed HBM4 and foundry cooperation with Samsung. He publicly called Korea an "irreplaceable" part of NVIDIA's AI infrastructure.
Is HBM the same as regular computer RAM? No. Regular DDR5 RAM and HBM are both types of DRAM, but they serve different purposes. Regular RAM is mass-produced for consumer devices at lower cost. HBM uses far more advanced manufacturing — including 3D stacking and through-silicon vias — making it significantly more expensive but orders of magnitude faster. It's manufactured only by SK Hynix, Samsung, and Micron.
Why are consumer RAM prices rising in 2026? Because producing HBM requires diverting semiconductor wafer capacity away from regular DRAM production. Standard DRAM prices rose over 60% in 2025, with further increases in 2026, directly linked to AI demand for HBM crowding out conventional memory production.
What is HBM4 and when is it available? HBM4 is the sixth generation of High Bandwidth Memory, offering significantly higher bandwidth and capacity than its predecessor HBM3E. Samsung began shipping HBM4 production units in February 2026 — an industry first. SK Hynix has HBM4 mass production underway. NVIDIA's Vera Rubin platform, confirmed in full production as of June 2026, uses HBM4 as its primary memory.
Korea Isn't Just Making Chips — It's Making the Future
Here's the thing that gets lost when this story gets told through a pure finance lens: what Samsung and SK Hynix have built isn't just industrial capacity. It's decades of accumulated knowledge that can't be replicated quickly. The engineers, the process know-how, the yield optimization, the tooling relationships — all of it took thirty-plus years to develop.
When Jensen Huang signs a wafer and writes "Please Make More," he's acknowledging that the most advanced AI company in the world is dependent on Korean engineering for its most critical bottleneck. That's not a supply chain detail. That's a geopolitical fact.
The country that brought you K-pop, kimchi, and K-drama is also quietly powering every AI assistant, every generative model, and every data center that the global tech industry is racing to build. Most people who use AI every day have no idea.
Now you do.
Explore More on K-Culture Insider:
- How Samsung Became a Global Tech Empire
- Every Time You Use ChatGPT or Gemini, This Happens
- Why Korea's Robot Technology Is World-Class in 2026
- The Science Behind Korea's Spicy Food Obsession
#KoreanTech #HBMMemory #NVIDIAKorea #SKHynix #SamsungSemiconductor #KoreanSemiconductor #AIChips #JensenHuang #KoreaAI #HBM4 #VeraRubin #OnlyInKorea #KCultureInsider #KoreaTech2026 #AIInfrastructure
Comments
Post a Comment