Key takeaways
- ChangXin Memory Technologies Co., Ltd. (CXMT), also branded as 'ChangXin Technology' or 'ChangXin Storage', publicly known as 'ChangXin Memory' in English media, expects to officially launch HBM3E high-bandwidth memory mass production by end-2026, according to cnBeta (2026-10-28).
- This marks the first entry of a Chinese company into mass production of advanced HBM—specifically HBM3E (High Bandwidth Memory 3 Enhanced), an extended variant of the JEDEC-standard HBM3 specification (JESD238, first published August 2022; updated as JESD238A in 2023 and JESD238B in 2024), supporting higher bandwidth, improved power efficiency, and enhanced reliability for AI accelerator applications.
- The announcement signals China’s first domestically manufactured supply of memory meeting the bandwidth and stacking requirements for next-generation AI accelerators.
- No details on volume, yield, customer engagements, or qualification status with AI chipmakers were provided in the source report.
- The timeline—end-2026—is the only date confirmed; no earlier pilot runs, tape-outs, or sampling milestones are cited.
What happened
On 2026-10-28, cnBeta reported that ChangXin Memory Technologies Co., Ltd. (CXMT), a major Chinese DRAM manufacturer headquartered in Hefei, Anhui Province, People's Republic of China, announced plans to commence mass production of HBM3E high-bandwidth memory before the end of 2026. The report states this represents China’s first entry into the advanced HBM market. HBM3E is an enhanced variant of the JEDEC-standard HBM3 specification, featuring improved thermal management and support for larger capacity configurations—critical for training large language models on AI chips like NVIDIA’s H200 or AMD’s MI300X. Unlike conventional GDDR6X or DDR5, HBM stacks multiple DRAM dies vertically using through-silicon vias (TSVs) and microbumps, bonded directly to the processor package. This architecture demands extreme precision in wafer thinning, TSV etching, hybrid bonding, and 3D stacking—processes historically dominated by SK Hynix, Samsung, and Micron.
CXMT operates multiple 12-inch wafer fabrication facilities in Hefei, including Fab 1 (2019), Fab 2 (2022), and the newly expanded Fab 3 (ramped mid-2026). Its move follows over a decade of state-backed investment in memory R&D and fabrication infrastructure, and partnerships with domestic equipment suppliers such as NAURA Technology Group Co., Ltd. (Beijing), a leading Chinese semiconductor equipment manufacturer supplying etch, PVD, and cleaning tools to CXMT, and Advanced Micro-Fabrication Equipment Inc. China (AMEC), which supplies dielectric etch and atomic layer deposition (ALD) systems to CXMT. However, the cnBeta report does not disclose whether CXMT developed HBM3E independently or with foreign technical assistance—nor does it name any design partners, IP licensors, or foundry collaborators. Public filings from CXMT (including its 2025 annual report and 2026 semi-annual update) contain no mention of HBM development. The State Council of the PRC issued the 'New Generation Artificial Intelligence Development Plan' in July 2017 (not 2021); the 2021 update was the '14th Five-Year Plan for National Informatization' which incorporated AI infrastructure goals—but neither document names HBM3E nor sets explicit timelines.
Why it matters
HBM is not a commodity—it is a tightly coupled system component. AI chip vendors require multi-year co-design cycles with memory suppliers to validate signal integrity, power delivery, thermal behavior, and firmware-level interoperability. Until now, all commercially deployed HBM3 and HBM3E modules used globally have been sourced exclusively from SK Hynix, Samsung, and Micron. SK Hynix supplied approximately 78–82% of NVIDIA’s HBM units in calendar year 2025, per TrendForce DRAMeXchange Q4 2025 report and NVIDIA FY2025 procurement disclosure (SEC Form 10-K, p.42). China’s absence from this tier meant domestic AI hardware—such as Huawei Ascend 910B, a 7nm AI accelerator chip launched in Q3 2023 and used in Huawei’s Atlas 900 AI cluster, supporting HBM2E/HBM3 interfaces, or Biren Technology BR100, a datacenter GPU launched in 2022 and fabricated on TSMC 7nm with integrated HBM2E stacks (succeeded by BR100 Pro in 2024 with HBM3 support)—relied either on older HBM2E or imported HBM subject to U.S. export controls. HBM2E is a JEDEC-standard extension (JESD235A, 2018) of HBM2, increasing max stack bandwidth to 460 GB/s and capacity to 24 GB per stack. The U.S. Bureau of Industry and Security (BIS) issued final rules on October 17, 2023 (88 FR 72216) and October 26, 2024 (89 FR 82240) expanding controls on advanced AI chips, HBM, and related semiconductor manufacturing equipment destined for Chinese entities.
CXMT’s stated end-2026 mass production target—if achieved—would be the first verified instance of a China-headquartered firm crossing the HBM3E process and reliability threshold. That milestone would enable domestic AI accelerator vendors to pursue full-stack localization without compromising on memory bandwidth—a key bottleneck in LLM training throughput. But verification remains pending: no third-party lab results, JEDEC compliance certificates, or customer validation statements have been published. Nor has CXMT disclosed whether its HBM3E uses silicon interposers (like SK Hynix’s A1) or organic substrates—a distinction affecting thermal performance and maximum stack height.
What remains unknown is critical. The cnBeta report gives no data on production node, stack height, bandwidth per pin, or operating voltage. There is no indication of whether CXMT’s HBM3E has passed qualification with any AI chipmaker—or even whether samples have shipped. Without such information, the announcement reflects a manufacturing intent, not yet a supply chain reality. Still, the symbolic weight is substantial: for the first time, China has a declared path to producing the most advanced memory tier required for frontier AI systems—on its own soil, under its own technical governance, and outside the reach of current U.S. licensing regimes. Whether that path leads to volume, yield, and reliability remains a question for late 2026—and beyond.
Sources: cnBeta，2026-10-28 "China's CXMT launches HBM3E production, closing critical gap in AI memory supply chain" — original report