SK hynix's turnaround ran on customer relationships, not on HBM technology
SK hynix's comeback was not a bet on HBM as a technology but a bet on customer relationships and sustained investment; Samsung, betting on PIM for edge inference, happened to hand its rival the sole-supplier position at the ChatGPT moment.
The argument · tap a timestamp to hear it
Vetoing the Micron acquisition is what left room for HBM to exist
After creditors took control of the restructuring, Micron offered to buy Hynix's memory business for 108 million shares. Most creditors were in favor, because Micron's purpose was not to operate Korean fabs but to shut the capacity down and eliminate a competitor. In late April 2002, the ten-member board voted unanimously against it: the price undervalued the business, the restructuring terms were unacceptable, and memory prices were already recovering. The public was also opposed to selling a Korean company to an American one. Hynix was left to sell its TFT-LCD and system IC businesses to stay alive. The video also offers the counterfactual: had Micron's offer been accepted at the time, the later history of HBM might not exist at all.
Buying Hynix was one chairman's decision, not the group's
Inside the SK Group the opposition was strong: chips had no synergy with the group, and Hynix was effectively bankrupt, having been under creditor management for more than a decade. Chey's judgment came from industry consolidation: Qimonda had already collapsed and Elpida was on the brink, so the survivors would be more resilient across the cycle, and the mobile market represented by iPhone and Android could hedge the PC cycle. He did the due diligence himself, read up on semiconductors, and flew to Taiwan to ask Morris Chang whether he intended to go into foundry. Only after Chey answered that he would not did Morris relax. In 2012 SK bought a 21% controlling stake for $3 billion. The point is that the entire deal was driven personally by the chairman, against internal resistance so strong that the bid was not submitted until the deadline itself.
Running alongside AMD on HBM followed a principle set a decade earlier
Chey explained the goal of buying Hynix: to turn a "commodity memory producer" into a mainstream semiconductor company whose products were indispensable. The lesson from the previous cycle was that a pure commodity DRAM business cannot escape the cycle, and that rivals can cross-subsidize it from other businesses. He wrote a memo by hand demanding that technology be used to create economic value, and that the company work with customers from the design stage onward. That directly explains why SK hynix was willing to run alongside AMD when AMD wanted to build HBM: even though HBM made no money early on and might not have earned back the cost of the equipment, Chey kept backing continued investment. This stretch is about strategic continuity: set the customer principle first, then wait for the market to validate it.
Bandwidth is stacked up from channel count, not from per-channel speed
The physics of HBM: bandwidth is set jointly by the number of channels and the rate of each channel, and a large number of slow channels is a better deal than a small number of fast ones. The TITAN Black's GDDR5 had 384 channels at 7 Gbps, for roughly 336 GB/s of total bandwidth; four HBM1 stacks provided 4,096 channels at only 1-2 Gbps each, and the total bandwidth was higher anyway. TSVs shorten the interconnect, which lowers resistance and power. HBM1 debuted in AMD's R9 Fury X, a flagship GPU of roughly 1,011 square millimeters integrating 22 dies, the highest-performing part but not a high-volume one. SK hynix and AMD handed the specification to JEDEC to build an ecosystem, laying the groundwork for the standards fight that followed.
HBM2 proved you get cost or performance, and only one of them
The HBM2 disaster is worth taking apart: to save cost the team skipped the 2y node and went straight to 2z; the customer changed the specification midstream, forcing a redesign; TSV speed came in at only 500 Mbps, a quarter of the 2 Gbps target; and the high-end version was cut. The rush meant dies cracked or were miscut during stacking, yield was around 50%, and the parts did not even pass customer qualification. The CEO heard "bitter criticism" from a customer to his face, and the design lead was replaced three times in a year and a half. Then volume production in 2018 ran straight into the crypto winter: bitcoin fell 75%, demand for mining GPUs and for HBM vanished, and the newly built P&T4 packaging line sat underfed. The team finally distilled it into a single dilemma: cost or performance, pick one.
Layer-by-layer thermal compression cannot hold HBM together; it has to be molded in one shot
For HBM2E the team chose performance: double the TSVs, plus dummy bumps for heat spreading. But the old process, TC-NCF, is layer-by-layer thermal compression, pressing one layer at a time, so an 8-9 layer stack has to be heated and pressed over and over; the more TSVs there are, the easier they crack under pressure, and the process does not scale. Mass reflow stacks all the dies first and then reflows them in one pass to melt the microbumps, which solved the bonding; but capillary underfill was too slow, and after six months it was abandoned. The team switched to MUF, using EMC molding compound to do the underfill and the overmold at the same time. MUF had previously only been used on single small chips; nobody had done it on an 8-layer DRAM stack. SK hynix started from zero, building the equipment with Towa of Japan and the materials with NAMICS, iterating the EMC material roughly fifty times. MR-MUF became the differentiating weapon of HBM2E.
Samsung's turn to PIM handed HBM3 to its rival
Before ChatGPT, HBM accounted for only 1.5%-2% of the DRAM market, and most people did not consider it mainstream. At the decisive fork Samsung turned toward HBM-PIM, putting a programmable AI engine inside the memory and aiming at edge inference for speech recognition and machine vision rather than at large language models. PIM also demanded heavy software investment. The result was that when ChatGPT launched on November 30, 2022, SK hynix was the only company able to supply HBM3 for Nvidia GPUs. This is not a technology fairy tale; it is a rival choosing a direction and, in doing so, ceding the long-term market.
SK rose from fifth chaebol to second on the back of memory
In July 2026 SK hynix listed on Nasdaq via ADR, raising $26.5 billion, the second-largest offering in U.S. history. The video's judgment is that Korean companies rarely step outside their home market this way: the previous generation of the SK Group became the fifth-largest chaebol on energy and telecom, and this generation became the second-largest on SK hynix. The company has already committed $4 billion to an advanced packaging plant in Indiana, and has announced a mid-term investment plan of 1.1 quadrillion Korean won covering Cheongju, Yongin and a new cluster in southwestern Korea. DRAM supply is already booked out for the next two years; the video's view is that more memory fabs are needed, and that SK hynix is going all in.
In their own words · checked verbatim
The deeper the downturn, the better you must connect with your customers; and in an upturn, never act as though you are above them.
Morris Chang12:13
What I really wanted to accomplish when we acquired Hynix was to transform it from a commodity memory producer into a mainstream semiconductor company whose products are indispensable
Chey Tae-won14:17
The design philosophy is that many slower channels will beat a few fast ones.
narration16:23
We will see more HBM in more AMD products. We didn't develop the technology just for this, we developed it for the future
Bryan Black18:28
Cost or Performance? Choose one. You can't have it all.
narration23:38
Few companies have swung this hard between crisis and success. After years of losses, SK hynix was nearly written off. SK bought it anyway over many internal and external objections. It made HBM and for years, HBM sat as this niche product few cared about. But the company kept working on it, even when the market gave no sign that it will pay off.
narration37:02
Figures
| Nasdaq ADR proceeds | $26.5 billion; second-largest offering in U.S. history | 0:02 |
| Micron's initial acquisition consideration (non-binding MOU) | 108 million Micron shares | 3:02 |
| Share price range, 2004-2005 | from 11 cents a share to roughly $21 | 6:04 |
| Average HBM2 yield | around 50% | 22:36 |
| Investment in the P&T4 packaging plant | more than $400 million | 23:38 |
| HBM's share of the DRAM market before ChatGPT | 1.5%-2% | 31:54 |
| SK hynix operating loss in the quarter ChatGPT launched | $1.5 billion, the first in a decade | 33:58 |
Glossary
- HBM / High Bandwidth Memory
- Multiple layers of DRAM stacked on a logic base die and interconnected with TSVs and microbumps, delivering far more bandwidth than GDDR.
- TSV / Through Silicon Via
- An interconnect running vertically through the silicon, carrying signals and power between stacked dies.
- TC-NCF / Thermal Compression with Non-Conductive Film
- Stacking layer by layer, using a non-conductive film plus heat and pressure to complete both bonding and underfill; SK hynix used it for HBM1 and HBM2.
- MR-MUF / Mass Reflow Molded Underfill
- Reflowing all the microbumps in a single pass, then using molding compound to do the underfill and overmold at once; developed by SK hynix for HBM2E.
- HBM-PIM / Processing-in-Memory
- Putting an AI compute engine inside the memory chip to cut data movement; Samsung turned this way and gave up on HBM3.
- ADR / American Depositary Receipt
- A certificate that lets a non-U.S. company's shares trade on U.S. markets.
How to listen
Semiconductor founders, AI infrastructure investors, memory and packaging engineers, and anyone currently making long-horizon decisions about locking in customers.
If you only care about HBM technology and the investment case, you can fast-forward through 9:13-10:13, the history of SK Telecom. It does, though, explain where Chey's instinct for the industry came from.