China’s CXMT to start advanced AI memory production despite poor yields, sources say - The Japan Times

CXMT forcing mass production of AI memory at sub-optimal yields is a capital-intensive replication of the 2015 commodity DRAM flood strategy, applied to a fundamentally harder product. HBM's 3D stacking and reliability requirements mean every 10-point yield deficit inflates unit cost by over 30%—this is a pure capital burn. The technical signal matters more than the headline: functional output implies CXMT has partially circumvented advanced DUV/EUV restrictions via multi-patterning. Expect the next US-Japan-Netherlands export tightening within two quarters. Downstream, domestic AI accelerator designers finally gain a non-gray-market memory source, materially de-risking China's AI supply chain. Competitively, SK Hynix and Samsung will accelerate HBM4 timelines and tighten Chinese allocations. Micron may exploit CXMT's mid-tier gap. The three-player oligopoly faces its first credible fourth entrant in a decade. 12-24 month projection: by mid-2027, CXMT yields should reach 55-65%, qualifying as a second source for mid-tier AI accelerators. This won't break the cutting-edge oligopoly but will compress HBM pricing 15-25% in the Chinese market, formalizing a bifurcated global memory landscape.