CXMT IPO Fundraising Excludes HBM, Cooling AI Memory Expectations
nashnova research
ChangXin Memory Technologies' STAR Market IPO targets ~¥29.5 billion (~$4.1 billion), yet the prospectus commits none of it to HBM capacity — a stark gap from market expectations of a push into AI memory. The narrative of a near-term Chinese challenge to Samsung and SK hynix needs dialing back.
Where does the ¥29.5 billion go?
The prospectus splits funds three ways: mainstream DRAM line upgrades, next-gen DRAM R&D, and manufacturing infrastructure — none earmarked for HBM.
This means → CXMT is betting entirely on commodity memory; HBM did not get a ticket in this funding round.
In plain terms = the money is for making existing products better and bigger, not for chasing the AI-memory wave.
99% of revenue from commodity DRAM — what does the customer list tell us?
Roughly 99% of 2025 revenue comes from mainstream DRAM serving PCs and smartphones; no standalone HBM customer contracts are disclosed.
Key customers include Alibaba Cloud, Tencent, ByteDance, Lenovo, and Xiaomi — all buyers of commodity DRAM and DDR5.
This reflects a revenue structure that is still a traditional memory company, not an AI-memory contender.
HBM capacity under 2%, yield at 25% — where is the bottleneck?
SemiAnalysis estimates CXMT's total DRAM capacity at roughly 265,000 wafers/month by end-2025, with only about 5,000 allocated to HBM — under 2%.
The HBM3 eight-layer stack product in development has a combined yield of roughly 25%. This means → three out of four units fail, and at that quality level exports outside China are impractical.
TSV — through-silicon via, a process that drills vertical channels through chips to interconnect stacked layers — is technically demanding. U.S. export controls on advanced equipment further narrow the window for rapid scale-up.
Why does the prospectus stay silent on HBM? It may not just be "not ready yet"
CXMT faces scrutiny from both the U.S. Department of Defense and the Commerce Department. Explicitly earmarking IPO funds for HBM capacity aimed at AI accelerators could trigger additional restrictions.
In plain terms = the silence on HBM is part technical reality, part regulatory risk management.
Some analysts argue HBM investment has not stopped; future funding may come from channels outside the IPO, including China's National Semiconductor Industry Fund or local-government financing.
How should we read CXMT's roadmap?
HBM capacity is projected at roughly 30,000 wafers/month by end-2026 and ~55,000 by end-2027 — but these are longer-term plans, outside this IPO's funding scope.
The strategic logic: solidify the commodity DRAM base, build DDR5 competitiveness, accumulate financial strength, then move deeper into HBM when the moment is right.
This means → whether the HBM roadmap materializes hinges on four things advancing in sync: yield improvement, customer qualification, equipment access, and government support.
市场有风险，内容仅供研究参考，不构成投资建议。