Last week, this column reported that CXMT had taken the commodity DRAM base. This week the story moves upstream into HBM proper — and the headline number everyone is repeating is that Korea's lead has narrowed to two years.
That claim comes from Shin Chang-hwan, a professor at Korea University, who told Korean financial media that the China-Korea HBM technology gap has shrunk to roughly two years. The reason isn't a single technical miracle, he argues. It's that HBM generations turn over faster than standard DRAM, compressing the window in which a laggard can be lapped.
Other Korean academics in the same report are less generous. Kim Yong-seok of Gachon University and Baek Seo-in of Hanyang University put the gap at about three years. Baek adds that if China attacks "without regard to cost," another year could come off.
Product generation may point to one to two generations, but yield, stacking, qualification, and ecosystem tell a four-to-five-year story.
Now the deeper contradiction. TechInsights senior vice president Choi Jung-dong looked at the same landscape and said that by product generation alone, China and Korea are one to two generations apart. Factor in mass-production yield, stacking and thermal management, customer co-design, and AI accelerator qualification — and the real ecosystem gap is four to five years.
Choi expects China to reach Korea's current level around 2030, and to match Korea's most advanced HBM products only by 2032 to 2034. Kwon Seok-jun of Sungkyunkwan University estimates a roughly six-year lag between what CXMT is currently pushing and what Korean firms have already developed and mass-produced.
These are not small disagreements. They are the difference between "the lead is safe" and "the lead is gone."
What triggered this round of anxiety is money. On the evening of September 28, CXMT disclosed a 349 billion yuan investment plan — about $7 billion — through a stock exchange filing. Of that, 241 billion yuan goes to technology research and development, and 108 billion yuan to a backend testing facility. Within the R&D bucket, 223.98 billion yuan is earmarked for equipment purchases alone. That is 92.94% of the total investment.
The project is scheduled to run 30 months. CXMT has already entered development and production of HBM3E, and the filing explicitly frames this spending as preparation for HBM4.
The backend facility is not a footnote. CXMT said external testing capacity was becoming insufficient as production expands, and that securing in-house DRAM testing and module assembly is a measure to prepare for HBM production expansion. In HBM, backend processing is not a support function — it is where yield is won or lost.
The United States technology outlet The Information reported on August 31, citing two sources, that CXMT has begun limited HBM3E production. The chips are undergoing qualification testing at two Chinese companies: Alibaba's Pingtouge and Cambrian. If qualification succeeds, CXMT's HBM could appear in products as early as 2027.
Korea's own industry association has already adjusted its numbers. The Korea Semiconductor Industry Association said in mid-September that the technology gap with China has narrowed to three years in HBM, two years in DRAM, and one year in NAND flash. A gap once estimated at more than five years has compressed to one technology generation or less.
The irony, per the Korea JoongAng Daily, is that Samsung and SK hynix may have opened this door themselves. By shifting resources toward high-margin HBM and server memory, they left the mainstream DRAM market open. Chinese firms began generating cash from commodity memory and reinvesting it in next-generation capacity and HBM research. That is a virtuous cycle — for China.
The numbers bear it out. CXMT's revenue surged 873.6% year on year to 150.31 billion yuan in the first half. Net profit swung to 77.6 billion yuan from a year-earlier loss. Its second-quarter operating margin hit 82%, topping SK hynix's 76% and Samsung's semiconductor business at 70%, according to the JoongAng Daily.
CXMT accounted for 10% of the global DRAM market in the second quarter, more than double its 4% share a year earlier. Samsung held 38%, SK hynix 25%, and Micron 24%, per Counterpoint Research.
Meanwhile Korea's incumbents are posting numbers that strain credulity. SK hynix's operating margin exceeded 70% — 72% in the January-March quarter, with net profit of 40.35 trillion won, five times the prior year. Most of that came from HBM.
Samsung's third-quarter operating profit consensus is 111.8 trillion won, which would mark a fourth consecutive record quarter. Kim Jae-joon, a vice president in Samsung's memory division, said on the second-quarter earnings call that HBM4 revenue in the third quarter would more than triple from the prior quarter. Next year's HBM supply negotiations are already complete.
Samsung's HBM share of total DRAM capacity is projected to rise from 27% last year to 33% this year and 40% in 2027. Conventional DRAM shrinks from 73% to 65% to 57% over the same period.
That reallocation is why DDR5 lead times for server-grade memory now stretch to 52 weeks — eight times the normal six-week cycle, according to KB Securities. The supply crunch in commodity DRAM is not temporary. It is structural.
But here is where the CXMT threat gets concrete. According to the Chinese-language report from September 27, CXMT entered its fifth-generation process platform in September 2026, boosting die output per wafer by at least 50%. Third-quarter DRAM price increases moderated to 13-18%, down from 53-58% in the prior quarter.
CXMT's three fabs — Hefei Phase 1, Hefei Phase 2, and Beijing — have a combined investment budget of 282.81 billion yuan. Construction progress on all three exceeds 96%. In the first half of 2026, 12.09 billion yuan of construction-in-progress converted to fixed assets, with another 22.77 billion yuan still pending. That capacity is coming.
The HBM4 crunch itself is worse than the headlines suggest. HBM4 mass-production yield is approximately 25%. Each stack uses 16 dies. If one die is defective, the entire stack is scrapped — and the dies above and below it go with it.
Stack yield degrades exponentially. If single-die yield is 99%, a 16-high stack yields about 85%. If single-die yield drops to 95%, the stack yield falls below 45%. The industry's 25% figure reflects that math at higher layer counts and more complex structures.
CoWoS advanced packaging capacity has a roughly 30% monthly gap. Global HBM demand is growing 67% year on year, with major suppliers booked through the end of 2027. Long-term contracts lock up more than 70% of capacity one to two years out. The spot market, when it exists, gets bid into the stratosphere.
This is the context for CXMT's $7 billion. It is not a bet on catching up tomorrow. It is a bet that the cash flowing from commodity DRAM — where China now has real share and real margin — can be converted into advanced packaging capability, stacking know-how, and customer qualification.
The two-year gap claim depends on how you measure. If you measure by product generation and ignore yield, ecosystem, and customer qualification, two years is defensible. If you measure by what can actually be shipped in volume to a qualified AI accelerator customer, the number is four to five years.
Both numbers can be true simultaneously. That is the uncomfortable part.
Korea's current profitability is extraordinary. SK hynix at 72% operating margin, Samsung at record quarterly profits. But the same dynamic that produced those margins — shifting capacity to HBM and leaving commodity DRAM tight — is what gave CXMT the cash and the market opening to accelerate.
The question is not whether CXMT's HBM qualifies. It is when, at what yield, and at what cost. CXMT has incorporated X-Stacking, a 3D NAND bonding technology from YMTC, into its HBM design, per DigiTimes. That approach directly bonds two wafers, shortening circuit distance and improving bandwidth and heat without EUV. Wuhan Xinxin handles wafer packaging; Tongfu Microelectronics handles assembly and testing.
China cannot buy EUV. It is building workarounds. Whether those workarounds scale is the only question that matters.
Korea's lead is not gone. But the assumption that it is structurally protected — that the silicon cycle will always favor the incumbent with the best equipment — is being tested. The gap is narrowing. The only disagreement is how fast.