China's memory makers are catching up. What happens when they do?

The memory business has rarely had it this good, thanks to the artificial intelligence (AI) boom.

Korea’s chipmakers are reporting extraordinary earnings and unusual confidence that the good times will last. SK hynix’s third-quarter operating profit rose more than sevenfold to 83.3 trillion won ($62 billion), an operating margin of almost 80 percent. Samsung Electronics’ rose nearly ninefold to 107.4 trillion won, with its memory business setting records of its own.

Across the West Sea, China’s challengers are sharing in the windfall.

ChangXin Memory Technologies (CXMT) remains far smaller, but its quarterly revenue rose more than sevenfold to 118.7 billion yuan ($17.8 billion), with a net profit attributable to shareholders of 54.4 billion yuan, reversing years of losses.

CXMT and Yangtze Memory Technologies (YMTC) are the twin pillars of China’s drive for memory self-sufficiency, part of Beijing’s broader pursuit of technological independence facing U.S. export restrictions. Fresh capital, new factories and increasingly sophisticated technology are helping them challenge the overseas incumbents that have long dominated the industry.

The remaining tech gap is “hard to close,” according to Laila Khawaja, research director at Gavekal Technologies. She expects Chinese and non-Chinese memory markets to diverge in supply chains and technology paths, but sees Chinese suppliers’ gains at home spilling over into global pricing and bargaining power.

“Chinese suppliers do not need to flood overseas markets to create that pressure,” she said.

CXMT, founded in Hefei, capital city of China’s eastern Anhui province, in 2016, has become the world’s fourth-largest producer of DRAM, the working memory used in computers and smartphones. Its share of global DRAM revenue reached 9 percent in the third quarter of 2026, up from 5 percent a year earlier, Counterpoint Research's latest data shows.

Government funding helped finance its entry into the capital-intensive industry. The memory boom is now funding its ramp-up: revenue rose more than eightfold to 269 billion yuan in the first nine months of 2026. Its blockbuster Shanghai IPO in July raised 66.6 billion yuan, the biggest in Asia of the year.

Morgan Stanley estimates that CXMT’s monthly wafer-processing capacity could reach 500,000 by 2028, roughly 15 percent of the global DRAM total, with bit shipments growing at a compound annual rate of 40 percent between 2025 and 2028.

CXMT is also moving up the technology ladder. At an earnings briefing on Oct. 30, the company highlighted it had begun mass production using its fifth-generation process, which it claims produces at least 50 percent more chips per wafer than its predecessor.

CXMT also said it had validated prototypes using its sixth-generation 4F2 design, which could theoretically shrink memory cells by a third compared with conventional layouts. It plans to send customers samples of server memory modules using the technology by year-end.

YMTC, founded in Wuhan the same year with central and provincial state backing, has made similar strides in NAND flash, which retains data without power. By the second quarter of 2026, it ranked third globally by shipment volume, with a 16.5 share, ahead of Japan’s Kioxia, according to Counterpoint.

“With AI driving the NAND super cycle, YMTC has become the fastest-growing vendor, with prudent investments in capability and capacity despite geopolitical headwinds,” said Choi Jeong-ku, senior analyst at Counterpoint Research, noting that the chipmaker’s revenue in 2026 is bound to grow six times faster than the industry and its peers.

Its strength lies partly in Xtacking, an architecture unveiled in 2018 that builds memory cells and their control circuitry on separate wafers before bonding them together. This allows denser storage and faster data transfer. Samsung has licensed YMTC’s hybrid-bonding patents and SK hynix was also said to be in talks, local media outlets reported, while Micron faces patent infringement lawsuits brought by the Chinese company.

YMTC is seeking 33 billion yuan through a Shanghai listing without a clear timeline.It reported a first-quarter net profit of 33.4 billion yuan — more than twice its earnings for all of 2025, prospectus shows.

Equipment orders tell the same story. Advanced Micro-Fabrication Equipment China (AMEC), which makes tools that etch microscopic structures into chips, booked record orders in the first nine months of 2026, with memory manufacturers accounting for about 60 percent.

Harder part

The harder test is turning those advances into chips that can be made reliably, at competitive cost and in volume.

CXMT’s earlier announcement on 4F2 rattled investors, sending shares in SK hynix, Samsung and Micron tumbling that day, although broader concerns over AI spending also weighed on the sector.

However, its rivals were more circumspect. SK hynix said it would keep extending its proven 6F2 architecture before switching to a new design.

“The timing of technological development and the timing of commercialization are distinct from each other,” management said during its third-quarter earnings call.

Samsung is also developing 4F2. Micron has excluded it from its road map, arguing that smaller transistor layouts do not necessarily improve density and the capacitors that store electrical charge remain the limiting factor.

Nomura expects CXMT’s 4F2 chips to reach volume production only in 2027-28, and estimates that its manufacturing technology still trails the leaders by at least three years.

The gap is also evident in manufacturing economics.

“CXMT remains about two generations behind the industry leaders,” Macquarie’s Daniel Kim wrote in an October report. He estimates its conventional DRAM process delivers only about 65 percent as much memory per wafer as the industry’s leading processes.

The divide is wider in high-bandwidth memory (HBM), which stacks DRAM chips to feed data rapidly to AI processors.

“CXMT remains primarily a conventional DRAM maker, and is not yet capable of mass-producing HBM profitably,” Khawaja said. Its preceding DRAM platform still falls short of mature industry yields — the proportion of working chips — of 85-90 percent, she estimates.

YMTC is closer to the leaders in NAND, but still trails by about a generation in cutting-edge layer counts. Adding more layers requires equipment capable of etching ever deeper, narrower structures through the stack — another area where Chinese toolmakers lag.

Equipment remains a central constraint for both companies. Beyond gaps in technology, manufacturing know-how and research talent, domestic toolmakers face “patent walls, not to mention the rigorous validation period required and the high requirement for tool stability,” Khawaja said.

“These are not easy gaps to close.”

Disruption starts at China market

“My base case remains further erosion of foreign suppliers’ market share in China,” Khawaja said. That displacement can affect global prices even without Chinese chips leaving the country.

The first openings are in mainstream memory for smartphones, PCs and storage, supported by Beijing’s localization mandates for state-backed users and Chinese companies’ efforts to hedge against supply risks, she said.

As established manufacturers prioritize AI products, shortages are also prompting customers outside China to qualify Chinese alternatives. Once qualified, Chinese suppliers can gain a foothold as second or third sources, which “could affect the big three's [Samsung, SK Hynix, Micron] bargaining power over the long run,” she said.

The pressure will be uneven. CXMT poses a more immediate challenge in conventional DRAM than advanced HBM, where the manufacturing gap remains wide.

YMTC’s push into enterprise SSDs could bring it into closer competition with industry leader Samsung in higher-value data center storage. Counterpoint Research expects server-related products to generate nearly 39 percent of its revenue by 2028, up from virtually nothing in 2024.

“The bigger question is whether YMTC can convert this into design wins with customers outside China, not just domestic hyperscalers,” said MS Hwang, a research director at Counterpoint.

The greater medium-term risk is that AI demand slows after new capacity comes online, Khawaja said. That could push incumbents to redirect DRAM capacity towards conventional products, where stronger Chinese competition could weaken their pricing power. Equipment constraints make an imminent China-driven glut unlikely, she added.

For memory chipmakers, prosperity usually arrives with an expiry date. Rising prices finance new factories; the factories produce a glut; the glut destroys the profits that built them.

AI has interrupted the cycle. SK hynix expects demand to outstrip supply through 2030.

For the boom to last, AI customers must keep spending faster than chipmakers can add usable capacity.

Industry analysts disagree over how durable those supply constraints will prove.

Macquarie’s Daniel Kim sees structural limits as “supply growth is capped by DRAM’s slowing scaling and worsening HBM wafer penalty.”

Morgan Stanley’s Shawn Kim argues that supply will eventually catch up. “Structurally higher AI secular growth riding on top of a cycle... is still a cycle,” he wrote in an October note.