China's CXMT Begins Mass Production of 5th‑Gen DRAM and Launches 24‑Gb LPDDR5X Chip

SentiSense · Published Sep 19, 2026, 11:54 PM ET · Updated Sep 28, 2026, 2:45 AM ET

CXMT said its fifth-generation DRAM platform has entered mass production, built on an 11.95 nanometre half-pitch using quadruple patterning, and unveiled two 24-gigabit LPDDR5X products. The company says the platform yields at least 50% more dies per wafer and holds 50% more data than its fourth generation. A CXMT executive claimed parity with the industry's most advanced mass-produced nodes, not superiority. Its HBM3E work remains in trial production.

CXMT said its fifth-generation DRAM platform has entered mass production, announcing the milestone at the World Manufacturing Convention in Hefei. The company reduced the active area half-pitch to 11.95 nanometres using quadruple patterning, and unveiled two 24-gigabit LPDDR5X products built on the platform. The unit is gigabit, not gigabyte: a single 24Gb die holds 3 gigabytes of data.

The manufacturing claims are the substance. CXMT says the platform produces at least 50% more gross dies per wafer than its fourth-generation process, and that the new parts hold 50% more data than the equivalent products they replace. Reuters reported the announcement and framed it as something that could help China build a stronger competitor to Samsung, SK Hynix and MU.

That competitive framing belongs to the press rather than to CXMT. The only on-record executive language claims parity, that the company's process capability is now on par with the most advanced mass-produced nodes in the industry, which is a different and more modest claim than overtaking the incumbents. No source maps 11.95 nanometres onto a named competitor node, and analysts quoted in Korean coverage say the figure is hard to compare directly with how Samsung, SK Hynix and Micron name their own processes.

On market position the numbers circulating are approximate and do not agree. Counterpoint-sourced coverage puts CXMT near 10% of global DRAM, up from about 4% a year earlier, while Korean reporting cites 9.5% of DRAM revenue and fourth place. Both are directionally the same story of rapid share gain from a small base, and neither should be quoted as a precise figure.

The constraint worth watching is yield economics rather than density. Quadruple patterning adds process steps and cost, so the real test is whether CXMT can hold yields at volume and price competitively. One area where it has not yet arrived is high bandwidth memory: its HBM3E is reported to be in trial production, not mass production, which is the segment where AI server demand actually concentrates.