CXMT Hit The Tier That Matters — And The Tier Moved

That card argued the China memory scare landed in standard DRAM, not in the memory the thesis is actually long. CXMT has now begun small-scale production of HBM3E. That is the tier that matters. The card was right about where the scare was aimed and it is now overtaken on the one axis it drew the line. What makes this worth writing rather than quietly dropping is that the timeline moved, not just the product. The published roadmap had CXMT mass-producing HBM3 by the end of 2026 and reaching HBM3E in 2027, with HBM3 8-high yields reported around 25%. Producing HBM3E now skips a rung and puts them roughly a year ahead of what analysts modeled. A company that was two full tiers behind is one tier behind. But the frontier moved too, and it moved further. Samsung has scaled HBM4 sales and shipped the industry's first HBM4E samples. NVIDIA's Rubin Ultra ships with HBM4 8-Hi. The memory that matters for frontier training is no longer HBM3E — it is HBM4, and HBM4 requires 2nm or 3nm logic base dies. That is exactly where lithography export controls bite hardest and where CXMT's constraint is physical rather than a matter of engineering time. So the honest scorecard: the gap narrowed from two generations to one, faster than anyone modeled, and the remaining generation is the one China cannot buy the tools for. The original card said the scare missed the tier that matters. It no longer misses. It is one tier short of the tier that matters, and closing. Caveats that stay attached: small-scale production, single-sourced to The Information, no yield figure disclosed. Small-scale HBM3E at 25%-class yields does not move global HBM pricing. It does change what CXMT is. What would settle it: a disclosed HBM3E yield above 50%, a named Chinese AI accelerator shipping CXMT HBM3E in volume, or evidence of CXMT taping out an HBM4 base die.