Why did SK Chairman become Korea's new richest man?
The latest Korean rich list has been released, and Chey Tae-won, chairman of SK Group, has topped the list. And on March 17, at the NVIDIA GTC Conference, Chey Tae-won stated:
"In the AI wave, HBM demand will increase more than tenfold by 2030"
This is not just random talk-his wealth is already built on this AI boom of HBM. Today, starting from basic concepts, let's explain this thoroughly in one go.
First, understand three basic concepts: What are DRAM, NAND, and HBM?
|
Name |
Type |
Volatile? |
Core Function |
Where You See It in Daily Life |
|
DRAM |
Dynamic Random Access Memory |
Yes (data lost when power off) |
Running memory, stores temporary data currently being processed by CPU/GPU |
DDR5 memory in computers, LPDDR memory in mobile phones |
|
NAND |
Flash Memory |
No (data retained when power off) |
Long-term storage, stores files, photos, system images |
Solid State Drives (SSD), USB drives, internal storage in mobile phones |
|
HBM |
High Bandwidth Memory (high-end variant of DRAM) |
Yes |
Super VRAM dedicated for AI GPUs, 3D stacking process, provides ultra-high bandwidth |
Data center AI graphics cards (A100/H100), high-end computing cards |
What exactly is the relationship between HBM and GPU?
One sentence summary: HBM is a "super oil pipeline" custom-made for GPU
- he characteristic of GPU is massive parallel computing (an H100 has tens of thousands of computing cores)
- So many cores working simultaneously require ultra-large bandwidth to continuously feed data
- Traditional GDDR video memory bandwidth is insufficient, HBM achieves more than 3 times the bandwidth through 3D stacking
- Without HBM, even the strongest GPU computing cores will "starve" and cannot perform at full capacity
- Intuitive analogy:
- GPU = High-horsepower engine
- HBM = Large-diameter oil pipeline
- The greater the engine horsepower, the more it needs the pipeline to keep up with oil supply
Map of major global manufacturers: Three companies monopolize, two are in Korea
1️⃣ DRAM (Running Memory)
|
Manufacturer |
Region |
Status |
|
Samsung |
South Korea |
Global No. 1 |
|
SK hynix |
South Korea |
Global No. 2 |
|
Micron |
United States |
Global No. 3 |
|
ChangXin Memory |
Chinese Mainland |
DDR4 mass production achieved, DDR5 in R&D |
2️⃣ NAND Flash (Long-term Storage)
|
Manufacturer |
Region |
Status |
|
Samsung |
South Korea |
Global No. 1 |
|
SK hynix |
South Korea |
Global No. 2 |
|
Micron/Kioxia/Western Digital |
US/Japan/US |
First echelon |
|
Yangtze Memory |
Chinese Mainland |
Breakthrough in 3D NAND, 232-layer mass production |
3️⃣ HBM (High Bandwidth Memory) - The Highest Barrier
|
Manufacturer |
Region |
Status |
|
SK hynix |
South Korea |
Technologically leading, first to mass produce HBM3, core supplier for NVIDIA H100/H200, nearly half of the market |
|
Samsung |
South Korea |
Close behind, HBM3/HBM3E already in mass production |
|
Micron |
United States |
Third, supplies North American customers |
|
ChangXin Memory |
Chinese Mainland |
In R&D, still some distance from mass production |
Why is HBM so difficult to manufacture that only three companies worldwide can do it?
HBM is the "Jewel in the Crown" of the memory chip industry, three major barriers block almost all other players:
1.3D Stacking Process
- Stack 8-12 layers of DRAM dies vertically together
- Use TSV (Through-Silicon Via) to achieve interlayer connection
- Alignment accuracy must be controlled at nanometer level, yield improvement is very difficult
2.Silicon Interposer
- Requires high-density wiring on silicon wafers to connect HBM and GPU
- Complex process, extremely high cost, not something ordinary manufacturers can afford
3.Ultra-high Bandwidth Design
- HBM3E bandwidth has already exceeded 1TB/s, signal integrity design is extremely challenging
- Requires decades of DRAM technology accumulation, cannot be caught up overnight
Why has SK hynix become the biggest winner in this AI wave?
4.Bet early, hit the tuyere
- SK hynix began layout of HBM more than ten years ago, leading in technology accumulation
- First to mass produce HBM3, just caught this wave of AI large model explosion
- Now HBM capacity is in short supply, prices have increased 2-3 times, profit margins have soared
5.Latest judgment from GTC Conference
- SK Chairman Chey Tae-won clearly stated at GTC that AI demand for HBM is exponential growth
- Predicts that by 2030, HBM market size will increase more than tenfold
- SK is already accelerating expansion of HBM3E production to seize the opportunity
6.Profits translated into wealth
- HBM's explosive profits directly pushed up SK Group's market value, making Chey Tae-won the new richest man in Korea
- This is the reward for early technological layout-if you see the trend correctly,the trend will reward you
Where are we now?
- NAND Flash: Yangtze Memory has already achieved breakthrough, mass produced 232-layer 3D NAND, gained a firm foothold, domestic SSDs are already sold globally
- DRAM: ChangXin Memory has mass produced DDR4, is promoting DDR5 R&D, gradually catching up
- HBM: ChangXin has completed technical R&D, still needs time before mass production, and is making every effort to catch up
Final Summary
7.In this AI boom, the most upstream HBM is the real "gold basin"-no matter how much GPU manufacturers earn, HBM manufacturers are the inevitable "shovel sellers" who make money.
8.SK hynix won with early layout-insisted on R&D more than ten years ago when no one valued HBM, today reaps the largest AI dividends and became Korea's new richest man.
9.In the chip war, memory is the main battlefield-whoever masters HBM holds the "faucet" of AI computing power. We have already broken through in NAND, are steadily catching up in DRAM, and still need to continue working hard in HBM.
