China's new memory chips may stay home for AI

What happened

On September 24, 2026, TrendForce reported that CXMT and YMTC are speeding up capacity. The English piece uses Commercial Times for the wafer figures and Chosun Biz, plus TrendForce’s own tables, for the competitive picture. Commercial Times expects CXMT’s monthly DRAM wafer starts to rise from around 320,000 now to around 420,000 by 2027. Those are reported expectations, not an audited spending filing. YMTC’s third fab in Wuhan is described as beginning operations by the end of 2026 and ramping through 2027, with further fabs planned after that.

Institutional investors, as TrendForce relays them, do not expect the new capacity to reverse the global supply-demand imbalance, given the scale of AI memory demand. The same account says China’s expansion may not meaningfully move global supply and demand until 2027 or later: Commercial Times has the added DRAM aimed at China’s market, and the investors expect much of the extra NAND to be absorbed by domestic AI and cloud demand for enterprise SSDs and high-capacity NAND.

Commercial Times also says CXMT is raising the share of LPDDR5 and DDR5 in its mix. The revenue table has already moved. TrendForce puts CXMT’s global DRAM revenue share at 9.5 percent in the second quarter of 2026, up from 7.6 percent in the first quarter, fourth behind Samsung at 39.4 percent, SK hynix at 24.9 percent, and Micron at 23.3 percent. Chosun Biz, via TrendForce, says Samsung and SK hynix have put more production into HBM for AI demand, which tightened DRAM for phones and PCs and left room for CXMT in conventional memory. That 9.5 percent is a conventional DRAM figure. Prior DigiTimes and Seoul Economic Daily reporting still puts CXMT’s early HBM trial yields near 25 percent, which is what keeps a commodity share gain from becoming an HBM claim.

The two companies are also stepping into each other’s markets. On September 18, Reuters reported that three people with knowledge of the plans say CXMT is preparing a NAND flash research-and-development production line at its new Beijing plant. One source said a Beijing research institute includes NAND projects. A third person said CXMT has discussed the plans with customers, including a new storage startup that wants NAND for AI systems and supercomputers. The sources would not be named because the plans are private. Reuters could not establish when that line would start, or whether CXMT intends to move from research and trial work to large-scale commercial manufacturing. Reuters reported in April that YMTC had sent low-power DRAM samples to customers. Commercial Times, via TrendForce, says YMTC could begin trial production of LPDDR5 as early as the end of 2026. Industry observers cited by Commercial Times expect the portfolios could meet in direct competition around 2028, eventually including advanced manufacturing, hybrid bonding, and patents.

On September 20, Reuters reported CXMT’s statement that its fifth-generation DRAM platform had entered mass production, including two 24-gigabit LPDDR5X parts already in production for phones and portable devices. Separately, the first-half tally TrendForce published on September 23, drawing on ijiwei, ranked CXMT first in both revenue and attributable profit among listed Chinese semiconductor companies.

Why it matters

For readers outside China, fourth place in DRAM revenue is not the HBM chapter. Samsung and SK hynix still dominate the stacked memory that frontier training buys. CXMT’s climb on the ordinary memory table is the opening Chosun Biz described after those Korean lines directed more starts into HBM. It is not evidence that Chinese stacked memory has caught Korea. YMTC’s Wuhan timetable is a NAND build. CXMT’s Beijing NAND line, in the Reuters account, still has no date for commercial manufacturing, while YMTC has already sent low-power DRAM samples the other way. A meeting of the two portfolios around 2028 is observer color from Commercial Times, not a joint product announcement. YMTC has also been on the United States Entity List since 2022, Reuters noted, a constraint on tools and customers that a new fab date does not erase.

What to watch is a named buyer for the Beijing NAND line, a YMTC LPDDR5 trial that actually starts by the end of 2026, and a primary report that moves CXMT’s HBM trial yields out of the early band already described.

The September 20 mass-production report is CXMT’s Gen-5 DRAM. The first-half ranking is CXMT and SMIC on the A-share tape.