China’s leading memory manufacturers CXMT and YMTC have begun collaborating on high-bandwidth memory (HBM) development, a strategic move to secure domestic AI infrastructure supply. According to MK, the partnership leverages YMTC’s hybrid bonding technology to help CXMT improve HBM yields, addressing a critical bottleneck as advanced HBM from Korean suppliers remains subject to US export regulations. This cooperation signals a concerted effort to build a self-sufficient memory ecosystem for AI workloads, directly challenging the market dominance of Samsung and SK hynix in the high-end segment.
State capital and local equity stakes
The push for domestic HBM is underpinned by massive state capital. Since 2014, the Chinese government has invested approximately 680 billion yuan through the National Integrated Circuit Industry Investment Fund, with local governments providing additional equity and subsidies. Filings indicate that local government shares account for 50% of CXMT’s total, while 70% of YMTC’s parent company CCSH is owned by local government funds. This financial backing has allowed both firms to survive previous deficits and scale production, creating a unified front that combines DRAM and NAND expertise to tackle the complex packaging and yield challenges of HBM.
Huawei Ascend 950DT price surge
The urgency of this domestic push is driven by severe supply constraints and rising costs for Chinese AI chipmakers. With advanced HBM from SK hynix, Samsung, and Micron restricted or expensive, manufacturers like Huawei are turning to gray market sources at significant premiums. Reports indicate that the quoted price for Huawei’s Ascend 950DT accelerator chip has increased by 20% to 50% in just two months, exceeding 250,000 RMB. Similar price hikes are reported for chips from Cambricon and other domestic players, highlighting how memory scarcity is directly impacting the cost structure of China’s AI hardware ecosystem (TechNews).
SanDisk HBF NAND stacking capacity
While the CXMT-YMTC alliance focuses on DRAM-based HBM, the industry is also exploring NAND-based alternatives to expand memory capacity for AI inference. SanDisk has promoted High Bandwidth Flash (HBF), a module that stacks NAND dies to offer 10 to 100 times the capacity of DRAM-based HBM at lower cost and power. This approach addresses the growing need for large memory footprints in inference systems, where traditional HBM capacity may be insufficient. By leveraging the higher density of NAND, HBF offers a potential path to more cost-effective memory expansion for next-generation AI accelerators.
Kioxia GP1 PCIe 6.0 IOPS performance
Kioxia is also positioning flash storage as a critical tier in AI architectures with its GP1 PCIe 6.0 SSD. Designed for GPU direct access, the KIOXIA GP1 delivers up to 10 million random read IOPS, serving as a high-performance memory extension that improves GPU utilization. This strategy aligns with the broader trend of using flash to offload data from costly in-node HBM or DRAM, enabling AI systems to access larger datasets efficiently. Kioxia’s technology, built on XL-FLASH generation 2, aims to scale to 100 million IOPS in future generations, reinforcing the role of high-speed flash in AI infrastructure.
Samsung HBM4 share and revenue mix
The global memory landscape is shifting as Korean suppliers consolidate their lead in advanced HBM while Chinese firms accelerate domestic production. Samsung’s HBM market share is expected to rise to around 40% in the fourth quarter, with HBM4 accounting for over 60% of its HBM revenue in the second half of the year. Meanwhile, the CXMT-YMTC collaboration represents a significant step toward reducing China’s reliance on imported memory, potentially altering the competitive dynamics in the AI storage market. This dual-track development, with Korean firms advancing in high-end HBM and Chinese firms building domestic capacity, underscores the intensifying geopolitical and technological competition in the memory industry.