Cutting-Edge Technologies
NextFin News - South Korea's lead over China in memory chip production is set to widen by 2028, even as China's largest memory maker doubles its capacity, after the Bank of Korea said Samsung Electronics and SK Hynix will add about 600,000 wafers a month of new output. The symmetry of the headline numbers — both sides adding roughly the same headline capacity — is deceptive: Korea starts from a much larger base, and the real gap is measured in revenue and technology, not wafer starts.
The central bank said in a report Friday that new chip plants due to come online by 2028 will lift South Korea's memory output capacity by about 600,000 wafers per month. Over the same period, China's leading memory maker, ChangXin Memory Technologies, or CXMT, is likely to raise capacity to as high as 600,000 wafers a month from its current level of 300,000. The report's message is not that China is standing still. It is that Korea is running faster on a longer track.
The Scoreboard: Korea Still Dominates Revenue and the AI Frontier
On the revenue scoreboard, the Korean duopoly still commands the market. In the first quarter of 2026, Samsung held 38% of the global DRAM market and SK Hynix held 29%, according to Counterpoint Research. CXMT accounted for 8%, up from 3% in the same period of 2025 — fast growth from a small base, but still less than a quarter of the combined Korean share. The entire DRAM market grew 80% quarter-on-quarter and 260% year-on-year in the first quarter, a surge that lifted every producer but lifted the leaders most of all.
The gap is wider still in high-bandwidth memory, the scarce product that powers AI accelerators and commands the industry's highest margins. In the first quarter of 2026, SK Hynix led HBM revenue with a 58% share, down from 69% a year earlier, while Samsung and Micron each held 21%. CXMT does not yet appear in the HBM revenue tally. Samsung, as the first supplier of HBM4 to Nvidia, is expected to gradually increase its share as shipments of the new generation materialize in the second half of the year. Most HBM revenue today comes from HBM3E, the generation before the newest.
These figures matter because revenue share, not capacity share, is what funds the next cycle of research and development. A wafer of HBM3E generates far more revenue than a wafer of commodity DDR4, and the Korean makers are the ones shipping it at scale. The price gap between the two tiers is stark: industry reporting indicates a 36GB HBM3E module has traded at about 2.87 million won, or roughly $2,100, on the spot market, compared with 500,000 to 700,000 won, or $300 to $400, under long-term agreements. That premium is what pays for the next generation of fabs.
Why the Gap Is Structural, Not Just Cyclical
The memory cycle is turning up, and a rising tide lifts every producer. Server DRAM contract prices rose 45% to 50% in the fourth quarter of 2025, with TrendForce projecting a further 55% to 60% quarter-on-quarter gain in the first quarter of 2026. That upcycle helps CXMT as much as it helps Samsung. But the Korea-China gap is widening for reasons that will outlast the cycle. Four structural advantages anchor it: process technology, advanced packaging, customer qualification, and capital.
First, technology. CXMT has narrowed the gap with the Korean leaders from more than five years to two to three years, but it is still chasing a moving frontier. The company mass-produced four-layer HBM2 in 2024, is advancing eight-layer HBM3 production in 2026, and targets 12-layer HBM mass production in 2027. By then, SK Hynix and Samsung will be shipping HBM4 and developing HBM4E. Catching up requires not just reaching today's node but arriving before the leaders have moved on.
Second, packaging. HBM is not simply a DRAM die; it is a stack of dies bonded together, and the bonding process is where yields are won or lost. SK Hynix's lead rests heavily on its mass reflow molded underfill process, known as MR-MUF, which produces better yields than the traditional thermal compression non-conductive film method at scale. CXMT is attempting to replicate this packaging know-how, but process integration is accumulated experience rather than a blueprint that can be copied.
Third, customers. Nvidia, Google, Meta, and other hyperscalers qualify memory suppliers through lengthy validation cycles, and once qualified they tend to stay. SK Hynix's full-year 2026 HBM capacity is sold out, and industry sources say Samsung, SK Hynix, and Micron have sold out their DRAM and HBM supply through 2027. A Chinese supplier without access to US customers faces a smaller domestic market — and even Chinese AI chipmakers cannot absorb the output at premium prices.
Fourth, capital. Samsung and SK Hynix are investing on a scale that few competitors can match. SK Hynix plans to invest 100 trillion won in Cheongju, including 80 trillion won in a new NAND fabrication facility and 20 trillion won in an advanced packaging plant due for completion by the end of 2027. Samsung's P5 facility in Pyeongtaek is expected to be operational by 2028, and SK Hynix's M15X facility in Cheongju is slated for use by mid-2027. Total semiconductor fab investment in Korea is projected to reach $35 billion this year, up 29.6% from a year earlier, and to remain elevated at $38 billion in 2027 and $37 billion in 2028.
The scale of the commitment reflects a belief that this demand is not a cycle but a regime change. SK Group Chairman Chey Tae-won said in March that the global chip wafer shortage is likely to persist until 2030, as AI demand continues to outpace supply.
"So we need some time to build up more wafers, at least four to five years. The current shortage could continue until 2030, so we expect more than a 20% shortage of the wafers."
He later warned that memory prices have to normalize, "otherwise, the market shrinks and competitors flood in" — an admission that high prices are themselves a signal to rivals like CXMT to keep building.
The Second-Order Effect: A Two-Tier Memory Market
The immediate consequence of Korea's expansion is more supply. The second-order consequence is market segmentation — and it is the part most investors are missing. As Samsung and SK Hynix add 600,000 wafers a month, much of that capacity will flow into HBM and high-end server DRAM, where margins are widest. That leaves standard DDR4 and DDR5 tighter than the headline capacity numbers suggest.
That dynamic is already visible in prices. DDR4 spot chips nearly tripled between late 2025 and August 2026, and TrendForce projects DDR5 server modules could double year-over-year by late 2026. A rare inversion has appeared in the spot market: chip quotes have exceeded module quotes, a condition that has historically preceded rapid module price surges, because module houses are sitting on inventory bought at lower prices and have not yet been forced to reprice. The spot market is pricing in a sustained upcycle, not a transient squeeze.
The mechanism is straightforward. AI-driven HBM demand cannibalizes standard DRAM capacity. Samsung's reallocation of its most advanced capacity to HBM production tightens standard DDR5 supply at exactly the wrong moment for PC and smartphone buyers. Korea's expansion, paradoxically, reinforces the shortage in legacy nodes — because the best wafers go to AI.
This two-tier structure has a geopolitical dimension. US export controls limit the equipment CXMT can buy, which constrains the advanced nodes it can reach. But those same controls also limit how much high-end memory Chinese AI chipmakers can consume, which caps the premium market CXMT can serve. The result is a bifurcated world: a high-margin tier where Korean makers sell to hyperscalers at contract prices locked in years ahead, and a commodity tier where Chinese makers compete on price for domestic and non-US customers. The revenue gap between those tiers is wider than the capacity gap.
There is also a macro channel back to Korea itself. Moody's raised its 2026 growth forecast for South Korea to 3.5% from 2.5%, citing the AI-driven memory chip boom, and projected the chip cycle to remain strong at least into the middle of 2027. A stronger chip cycle improves Korea's trade balance, supports the won, and eases fiscal pressure — which in turn gives Seoul more room to underwrite the subsidy packages that keep the Korean champions investing. The memory upcycle is not just a corporate story; it is a national balance-sheet story.
The Counter-Thesis: China Catches Up Faster Than Expected
The strongest argument against the widening-gap thesis is that China's catch-up is accelerating, not slowing. Goldman Sachs projects CXMT will supply 50% of China's domestic DRAM demand by 2028 and 40% of its HBM demand by the same year. Seoul Economic Daily reports that CXMT's capacity is approaching SK Hynix's roughly 590,000 wafers a month by 2028. If Chinese AI chipmakers mature in parallel, CXMT could carve out a self-sufficient supply chain insulated from US restrictions.
There is evidence that the door is already opening. Industry reporting indicates major PC makers and Apple are testing or sourcing CXMT DRAM, and the company's average selling price in the first quarter of 2026 was only about 5% to 10% below the Korean and US leaders — a much smaller discount than a laggard would have to accept. CXMT is also pursuing a high-profile Shanghai listing, with a target valuation that has ranged as high as roughly $85 billion at the IPO price, giving it the capital to keep expanding even if technology lags.
That scenario is plausible but not the base case. It requires CXMT to solve yield at advanced nodes, master advanced packaging at scale, and find customers willing to pay premium prices — all while US export controls tighten. CXMT began mass-producing LPDDR5 and DDR5 only late last year, products long established among the three leaders, and its yields are still considered unstable. The gap in revenue terms, not capacity terms, is what matters for profitability — and revenue follows technology.
There is a second risk inside the counter-thesis: overcapacity. If CXMT and its domestic peers build capacity faster than demand materializes, the commodity tier could flood first. Morningstar analyst Jing Jie Yu, who holds the only sell recommendation on CXMT, predicts that rising industry capacity will drag down DRAM prices in late 2027 and 2028, hurting profits for all producers. In a price war, the incumbent with the deepest pockets and the highest-margin product mix usually wins — which is to say, Korea again.
What to Watch
Three signals will confirm or break the widening-gap thesis. First, CXMT's HBM yield rates and its 12-layer production timeline in 2027 — if the company hits that target on schedule, the technology gap is closing faster than expected. Second, whether Samsung's HBM4 supply to Nvidia ramps without quality setbacks — a stumble there would hand share back to SK Hynix and narrow the Korean internal gap even as the China gap widens. Third, the DRAM price curve: if legacy DDR4 prices stop rising while HBM keeps climbing, the two-tier market is entrenched and Korea's revenue advantage compounds.
The base case is that Korea extends its revenue and technology lead through 2028, even as China's share of unit capacity grows. The upside case for China is that CXMT's 2027 12-layer HBM ramp succeeds and domestic AI demand absorbs the output at premium prices. The downside case for Korea is that HBM4 yields disappoint and Chinese packaging closes the gap faster than expected, compressing the pricing power that funds the next generation of investment.
History offers a warning for both sides. Memory is the most cyclical sector in semiconductors, and every boom has ended in a bust. The leaders of the last great memory upcycle learned that capacity discipline matters as much as technology leadership. This time, the discipline is being tested by AI demand that looks structural but carries a cyclical price tag. The Korean makers are betting they can stay ahead of the curve long enough for the next one to arrive.
South Korea is not winning the memory race by producing more wafers. It is winning by producing the wafers nobody else can yet make at scale — and by the time China reaches that frontier, the leaders will have moved on again.
Explore more exclusive insights at nextfin.ai.
