The AI Collapse and the Dangerous Gamble of Samsung and SK Hynix Going All-In on HBM

To start with the conclusion: Samsung-Hynix going all-in on HBM is sweet for short-term results, but it is astonishingly similar to the path of decline that Japanese semiconductors once took. The moment they emptied out the unprofitable general-purpose memory, China struck exactly into that gap — and even pocketed a hefty sum in listing proceeds. Let us verify with numbers.

1. The Illusion That There Is No HBM Supplier Other Than Samsung-Hynix

Let's start with reality. In Q2 2026, the global HBM revenue share was SK Hynix 50%, Samsung Electronics 33%, and Micron 18%, according to Counterpoint Research. Samsung pulled up 12 percentage points in a single quarter from 21% in Q1, halving the gap from 37 percentage points to 17. After shipping HBM4 to mass production for the first time in the industry in February, it reached $1 billion (about 1.36 trillion KRW) in HBM revenue in just four months.

In other words, HBM is not a Korean monopoly. Samsung and Hynix are in a civil war eating into each other, Micron is holding on at 18%, and China's CXMT is pushing HBM3E toward mass production. The industry worries that China may close even a one-generation gap after next year. Starting with HBM4, custom designs and a TSMC packaging alliance become variables, and there is no guarantee that Korea holds an absolute advantage in this area.

2. The Moment They Emptied General-Purpose Memory, China Came In

This is exactly where the user's point hits the mark. In Q2 2026, the global DRAM share was Samsung Electronics 39.4%, SK Hynix 24.9%, Micron 23.3%, and CXMT 9.5-10%. CXMT broke through double digits for the first time in its history: below 1% in 2023, 4% in Q2 of last year, 8% in Q1 of this year, and now 10%. It broke early past Counterpoint and UBS forecasts that it would not reach 10% by shipment volume until 2028 — and it did so on a revenue basis.

How was that possible? Because while Samsung-Hynix was absorbed in the 1c DRAM and HBM competition and cut supply, CXMT focused its attack on legacy general-purpose markets such as 1a DRAM. As the explosive HBM demand from AI servers led memory makers to increase HBM production, ordinary DRAM became severely scarce, and China took that vacant spot. A place they had emptied because it did not make money started to make money — only now the owner has changed.

Hynix's price was stamped in numbers. Its DRAM share evaporated by 7.2 percentage points in two quarters, from 32.1% in Q4 of last year to 24.9% in Q2 of this year, a 14 percentage point drop year over year. The gap with Micron (24%) has narrowed to 1 percentage point. The result of missing out on the windfall from surging general-purpose DRAM prices while defending its No. 1 spot in HBM (50%). BofA forecast that HBM's share of Hynix's DRAM revenue would fall from 40% last year to 18% this year. HBM revenue itself grows 63%, from $20.95 billion to $34.05 billion, but because the general-purpose DRAM market grew 59.5%, its share was diluted.

3. China's Ammunition Is Abundant

CXMT broke through a market cap of 3 trillion yuan on its first day of trading on the Shanghai stock exchange, securing roughly 12 trillion KRW in firepower. YMTC also pushed ahead in earnest with a 33 billion yuan (about 7 trillion KRW) IPO. That Samsung Electronics and SK Hynix plunged 13-14% in a single day right after the listing is a signal that the capital markets have begun to price China's pace of pursuit into the discount rate.

The capacity expansion is fearsome. CXMT will go from the current 300,000-320,000 wafers per month to 420,000 next year and 600,000 in 2028-2030. That is a scale rivaling SK Hynix's current capacity (about 590,000 wafers). Its target is a 15% share, and a Counterpoint director called 15% a baseline that must be crossed to secure investment funding. YMTC rose to No. 3 in the world for the first time ever in Q2 of this year with 14% of NAND bit shipments, overtaking Japan's Kioxia, and presented to investors a goal of global No. 1 by the end of next year. NAND wafer output expands from 2.01 million wafers this year to 2.496 million next year, while Samsung-Hynix stands still.

The technology chase also probes the gaps. CXMT eyes the cutting edge by equipping Xiaomi's new smartphone with LPDDR6, and pushes HBM3E toward mass production. YMTC is building two additional new plants in Wuhan at 100,000 wafers per month each, and is even considering producing HBM for AI. The assessment of a domestic industry official is painful: the place to watch more than the CXMT listing is YMTC, whose presence is growing in both share and technology.

4. A Carbon Copy of the Fall of Japanese Semiconductors

Japan surpassed a 50% share of the world semiconductor market in 1988. Hitachi, NEC, Toshiba, Fujitsu, and Mitsubishi dominated 80% of DRAM. That Japan plunged to a 7% share in 2016, and fell below 5% after the sale of Toshiba Memory. The process is a déjà vu of today.

First, the obsession with high quality lost to Korea on yield and cost. Japan clung to 25-year-warranty DRAM for mainframes, while Korea poured out cheap 3-year-warranty PC DRAM. In 1998 it handed the No. 1 spot to Korea. It overlaps with how Samsung-Hynix today clings only to the high-value HBM and empties out general-purpose lines.

Second, it was late to U.S. pressure and structural change. Through the U.S.-Japan semiconductor agreements of 1986 and 1991, it accepted a clause for a 20% foreign share, and missed the transition to the PC era. Today, the supply-chain diversification demands of big tech play the same role.

Third, the allied force Elpida went bankrupt in 2012. The Japanese DRAM alliance — formed by NEC and Hitachi merging and joined by Mitsubishi — filed for court protection leaving 448 billion yen (about 6 trillion KRW) in debt, and was acquired by Micron. Its share just before bankruptcy was 12.2%. It was pushed out by Samsung's chicken game. The failure of merger synergies, dependence of investment funds on the parent company, and absence of leadership are cited as causes.

Fourth, joint research actually grew Samsung. Of the Selete project, in which 12 Japanese companies and one Samsung participated, it was Samsung Electronics in 2001 that became the world's first to succeed in 300mm wafer mass production. The fact that the beneficiary of shared technology was the pursuer overlaps exactly with today's fears of technology leakage to China.

5. There Is a Counterargument: The HBM4 Price Reversal

For balance, let me present opposing numbers too. According to Eugene Investment & Securities, the average selling price per Gb of HBM this year is $1.61, lower than general-purpose DRAM ($1.82). But next year, as HBM4 supply gets into full swing, it rises to $3.52, surpassing general-purpose ($2.36). Next year's HBM4E is more than double that. In other words, one reading is that today's all-in bet is a move aimed at next year's price reversal. Samsung stated it would expand Q3 HBM4 revenue to more than triple Q2, and Hynix has also entered an expansion phase.

But the weakness of this counterargument lies in its premise. How long will HBM keep winning? The moment the AI boom turns, expensive HBM inventory takes a direct hit, and the general-purpose market they emptied is already China's. Just as Japan in its 1980s golden age never realized it was sick with over-quality, it is time to ask whether Samsung-Hynix is now sick with HBM.

Conclusion

The warning the numbers speak is clear. By emptying general-purpose lines in the HBM civil war, the result that came back was CXMT at 10%, YMTC at No. 3 in NAND, and 19 trillion KRW in listing firepower. It took Japan less than 30 years to fall from 50% to 7%. Before the HBM price reversal becomes reality, they must keep at least a minimum force to defend the general-purpose lines. Betting the entire production line on one thing is a gamble, not a strategy.