11 min read
What is HBM memory SK Hynix makes for AI? A tower of 8 to 12 DRAM chips placed right beside an AI processor and wired with thousands of tiny vertical connections, so data moves far faster than ordinary memory allows. AI accelerators such as Nvidia’s data-center GPUs can’t hit their speed without it, and SK Hynix, the Korean chipmaker, has been the biggest supplier since the AI boom began.
This guide explains how high-bandwidth memory works, why it costs more and pays more than regular memory, how SK Hynix stacks up against Samsung Electronics and Micron, and what all of it means if you follow Korean chip stocks from abroad.
Key takeaways
- HBM stacks DRAM chips vertically and links them with through-silicon vias (TSVs), giving each stack a data path many times wider than a single 64-bit desktop memory module.
- AI chips need HBM because training and running large models is limited by how fast data reaches the processor, not only by raw compute.
- Three companies make HBM at scale: SK Hynix and Samsung Electronics in Korea, and Micron in the US.
- SK Hynix shares stood at 1,199,000 won on Nov 10, 2026 (Yahoo Finance), up 5.55% on the day but down 25.3% over one month.
What is HBM memory: SK Hynix’s stacked-DRAM bet explained
HBM stands for high-bandwidth memory. It is still DRAM, the same short-term working memory that sits in your laptop, but it’s built and packaged very differently.
Regular DRAM chips lie flat on a circuit board, some distance from the processor, and talk to it over a fairly narrow connection. HBM stacks the chips on top of each other like floors in a building, then places that stack millimeters from the processor inside the same package. The result is a much wider, much shorter road for data.
SK Hynix helped invent the format. It developed the first HBM with AMD, and the industry standards body JEDEC adopted HBM as a standard in 2013. AMD’s Radeon R9 Fury graphics cards in 2015 were among the first products to use it. For years HBM was a niche, expensive part for high-end graphics and supercomputers. Then large AI models arrived and turned it into one of the most sought-after components in tech.
A quick bit of context for readers outside Korea: SK Hynix belongs to SK Group, one of Korea’s chaebol (large family-controlled business groups). It was once Hyundai Electronics and joined SK in 2012. It trades on the KOSPI, Korea’s main stock index, under the code 000660.
How high-bandwidth memory works, step by step
You don’t need an engineering degree to follow the build. An HBM stack comes together in five basic steps.
- Make thin DRAM dies. Memory chips (dies) are made on silicon wafers, then ground down very thin so many can be stacked without the tower getting too tall.
- Drill vertical wires. Thousands of microscopic holes are etched through each die and filled with metal. These are through-silicon vias, or TSVs. They carry signals straight up and down the stack instead of around the edges.
- Stack and bond. The dies are stacked, commonly 8 or 12 high in current products, and bonded together. How you fill the gaps between layers matters for heat and yield, and it is one place where suppliers differ.
- Add a base die. At the bottom sits a logic (base) die that manages the stack and talks to the processor.
- Package beside the processor. The finished stack goes onto a silicon interposer, a thin silicon slab with dense wiring, right next to the GPU or AI chip. This side-by-side setup is called 2.5D packaging. TSMC’s CoWoS process is the best-known example, and many AI GPUs carry several HBM stacks around one processor.
The payoff is width. A desktop DDR5 module moves data over 64 bits at a time. One HBM3E stack uses an interface many times wider, and HBM4 doubles that width again. Each individual connection can actually run slower than in regular memory, which helps save power, because there are so many of them.
Why AI chips need HBM instead of regular DRAM
A modern AI model is, at heart, billions of numbers (its weights) that must be read again and again. A GPU can do math extremely fast, but if the numbers arrive slowly, those math units sit idle. Engineers call this the memory wall.
HBM attacks the memory wall in two ways. It delivers far more bandwidth per chip, and because it sits so close to the processor, it spends less energy moving each bit. In a data center running thousands of GPUs, power and heat matter as much as speed.
That is why every leading AI accelerator, from Nvidia’s data-center GPUs to AMD’s Instinct chips and the custom AI chips built by big cloud companies, uses HBM. Ordinary memory still matters in AI servers, but HBM is the part that sits on the accelerator itself.
How HBM compares with the memory in your PC and gaming GPU
| Memory type | Where it sits | Data path width | Typical use | Main trade-off |
|---|---|---|---|---|
| DDR5 | Separate modules on the motherboard | 64 bits per module | PCs, servers’ main memory | Cheap and large, but far slower per chip |
| GDDR (e.g. GDDR6, GDDR7) | Separate chips on the graphics card | 32 bits per chip, running very fast | Gaming GPUs, some AI cards | Fast and cheaper than HBM, more power-hungry |
| LPDDR | Inside phones and laptops, often near the chip | Narrow, low-power channels | Smartphones, thin laptops | Saves battery, limited bandwidth |
| HBM (HBM3E, HBM4) | Stacked inside the same package as the processor | Very wide per stack; HBM4 doubles HBM3E’s width | AI accelerators, supercomputers | Highest bandwidth, hardest and costliest to make |
HBM generations from HBM to HBM4
Each generation is a JEDEC standard that suppliers build to, then push further with their own versions. Here’s the short history.
- HBM (2013 standard): the first version, co-developed by SK Hynix and AMD, used in high-end graphics cards.
- HBM2 and HBM2E: faster and larger stacks that moved HBM into data-center accelerators and supercomputers.
- HBM3: the generation behind Nvidia’s H100, the GPU that powered the first big wave of generative AI. SK Hynix was the main supplier.
- HBM3E: an extended, faster HBM3, with 8-high and 12-high stacks. All three suppliers sell it, and it fills many of today’s AI accelerators.
- HBM4: doubles the interface width of HBM3E and allows the base die to be made on a more advanced logic process, which pulls foundries like TSMC deeper into memory production.
For exact speeds and stack capacities of a specific product, check the supplier’s own product page. The numbers change with each release, and press coverage often mixes up generations.
Why SK Hynix leads the HBM market
SK Hynix’s lead comes from timing and stubbornness more than luck. Four factors explain it.
It kept investing when HBM was a niche
For most of the 2010s, HBM sold in small volumes. SK Hynix kept developing each generation anyway, so when AI demand surged it already had working products, trained engineers and production lines.
It won Nvidia early
Being the main HBM3 supplier for Nvidia’s H100 put SK Hynix inside the most in-demand chip of the AI boom. Qualifying memory for a customer like Nvidia takes long testing, so an incumbent supplier holds a real advantage for the next generation.
Its stacking method suited tall stacks
SK Hynix uses a packaging approach it calls MR-MUF (mass reflow molded underfill), which fills the gaps between stacked dies with a liquid material in one step. Samsung has mainly used a film-based method called TC-NCF. SK Hynix credits MR-MUF with better heat dissipation and yields as stacks grow taller. Each company defends its own approach, so treat the comparison as a design choice, not a settled verdict.
Its earnings now reflect it
HBM sells for much more per bit than standard DRAM, which changes the economics of a memory maker. The Wall Street Journal reported that SK Hynix plans $75 billion in shareholder returns after blockbuster earnings. For the full figures, read SK Hynix’s own earnings release on its investor relations site.
Samsung vs SK Hynix vs Micron: who makes HBM
Only three companies produce HBM in volume. They all make standard DRAM too, so HBM competes with their other products for factory space.
| Company | Home base and listing | HBM position | What to know |
|---|---|---|---|
| SK Hynix | Korea, KOSPI (000660) | Largest supplier through the AI boom | Main Nvidia HBM3 supplier, MR-MUF stacking, memory is its whole business |
| Samsung Electronics | Korea, KOSPI (005930) | World’s largest memory maker overall, fighting to grow HBM share | Also makes phones, TVs and runs a foundry, so HBM is one part of a much bigger company |
| Micron Technology | US (Boise, Idaho), Nasdaq (MU) | Third supplier, entered at HBM3E | Skipped HBM3 to focus on HBM3E, and has supplied memory for Nvidia’s H200 |
Samsung’s case is the one Korea watches most closely. It is the bigger memory company, yet SK Hynix moved first in AI-grade HBM. Samsung has supplied HBM to AMD and others and has worked to win more Nvidia orders, a process the Korean press covers almost weekly. If you’re comparing the two as investments, remember that Samsung Electronics’ results also swing with smartphones, displays and its contract chipmaking arm.
What HBM means for SK Hynix stock and Korean markets
Because HBM is tied so closely to AI spending, SK Hynix stock tends to move with news about AI chips, Nvidia and US tech shares. Recent trading shows how sharp those swings can be.
According to Yahoo Finance data as of Nov 10, 2026, SK Hynix traded at 1,199,000 won per share, up 5.55% on the day but down 25.3% over one month. Samsung Electronics was at 197,800 won, up 4.16% on the day and down 21.04% over a month. The KOSPI index rose 3.52% to 5,195.46, still 18.62% lower than a month earlier. The Philadelphia Semiconductor Index, a benchmark of US-listed chip stocks, jumped 6.85% on Nov 9 after an 18% one-month drop.
KED Global reported that Korean stocks rebounded with the chip giants leading the recovery, and CNBC reported that Asian tech stocks tracked a Wall Street rebound. In plain terms, the day’s rise followed US chip stocks higher after a steep month-long fall.
The two Korean chipmakers carry heavy weight in the KOSPI, so when they move, the whole index often moves with them. That makes HBM news Korean market news. This is not investment advice; it explains what moved and why.
Risks investors watch in the HBM boom
HBM has made memory look like a growth business, but memory has a long history of boom-and-bust cycles. These are the pressure points to follow.
- AI spending: HBM demand depends on how much cloud companies and AI labs keep spending on accelerators. Any slowdown hits HBM first.
- Customer concentration: a large share of HBM goes into Nvidia products, so news around Nvidia ripples into Seoul. The Hankyoreh asked whether a scandal at Nvidia could rattle Korea’s semiconductor industry, while THE ELEC reported that Nvidia’s Korea partnerships are proceeding.
- Supply catching up: as Samsung and Micron add capacity, HBM pricing power could weaken. That is good for chip buyers and less good for margins.
- Next-generation qualification: each new generation, HBM4 now, reopens the race for customer approval. A delay can shift share quickly.
- Trade and export rules: advanced chips and the memory inside them are subject to US export controls and wider trade policy, which can change which customers suppliers can serve.
Common misconceptions about HBM
“HBM is a faster kind of chip, not memory”
HBM is DRAM. The speed comes from stacking and the wide connection, not from a new kind of memory cell.
“More HBM always means a better AI chip”
Capacity and bandwidth help, but the processor design, software (such as Nvidia’s CUDA platform) and networking between chips matter just as much.
“SK Hynix makes the whole AI chip”
It doesn’t. SK Hynix makes the memory stacks. The processor is designed by companies like Nvidia or AMD, usually manufactured by TSMC, and the two are joined in advanced packaging.
FAQ
What does HBM stand for?
HBM stands for high-bandwidth memory. It is a type of DRAM made by stacking memory dies vertically and connecting them with through-silicon vias. The stack sits right next to the processor, which gives it far more bandwidth than standard memory.
Is HBM the same as GDDR memory?
No. Both serve graphics and AI chips, but GDDR chips sit separately on the board and run each narrow connection very fast. HBM uses a much wider connection inside the processor’s package. HBM delivers more bandwidth per watt, while GDDR is cheaper and easier to make, which is why gaming GPUs use GDDR.
Why is HBM so expensive to make?
It uses more silicon per bit than standard DRAM, needs thousands of TSVs per die, and requires careful stacking where one bad die can spoil a whole stack. It also depends on advanced packaging capacity. For current prices, check supplier earnings reports, since HBM contracts are negotiated privately.
Does Nvidia use SK Hynix HBM?
Yes. SK Hynix was the main HBM3 supplier for Nvidia’s H100 and remains a major Nvidia supplier. Micron and Samsung also compete for Nvidia’s HBM orders, and each new GPU generation is a fresh contest.
Can I buy SK Hynix stock from the US or Europe?
SK Hynix is listed on the Korea Exchange under 000660, and many international brokers offer Korean market access. Some platforms also list depositary receipts. Check your own broker for what is available, and remember that this article is not investment advice.
Will phones and laptops use HBM?
Not in the near term. Phones and laptops use LPDDR, which is optimized for low power and low cost. HBM’s size, cost and heat make it a fit for AI accelerators and supercomputers, not consumer devices.
What to watch next for HBM and SK Hynix
Three signals will tell you where HBM goes from here. First, HBM4 adoption: watch which suppliers each new Nvidia and AMD accelerator names, usually announced at product launches and earnings calls. Second, quarterly results from SK Hynix, Samsung Electronics and Micron, where HBM revenue and capacity plans show up first; their investor relations pages post the official numbers. Third, AI spending plans from big cloud companies, which set the demand that everything else rides on.
If you follow Korea for culture as much as for chips, the same export story runs through entertainment too, from K-pop tours to webtoons like The Remarried Empress, which you can read legally here. For chips, though, the scoreboard is simple: who ships the next HBM stack, in volume, on time.
Related reading
Sources news.google.com · k-korea.blog
Prices, opening hours and availability change. Check the official page before you buy or travel.