Micron, CXMT Unveil New Memory Push

Micron Presents EIM at SEDEX on Oct. 20, aiming for mass production in second half of 2028. (Reuters-Yonhap) A new variable has emerged for Samsung Electronics and SK hynix, which have been leading the artificial intelligence (AI) memory market spearheaded by high bandwidth memory (HBM). US-based Micron unveiled next-generation memory specialized for physical AI, and China’s ChangXin Memory Technologies (CXMT) also surprised the world by sharing news of its development of a next-generation memory architecture. In contrast, Samsung Electronics and SK hynix appear to be focusing on further advancing existing technologies, such as the next model of their current flagship HBM, Compute Express Link (CXL) which was announced for release several years ago, Processing-in-Memory (PIM), and High Bandwidth Flash (HBF). Neither the industry nor experts dispute that the memory technology capabilities of Samsung Electronics and SK hynix are the best in the industry. Given the nature of the semiconductor industry where ‘stability’ is crucial, they do not expect that new technologies will immediately overturn the market dynamics. Nevertheless, some experts pointed out that the companies must demonstrate a ‘super gap’ through overwhelming new technologies, citing the case of Samsung Electronics, which had developed HBM earlier than SK hynix in the past but failed to respond to the AI market in a timely manner. The reason is that doing so is necessary to reassure investors. According to the industry on Oct. 20, US-based Micron presented the development direction for its ‘Edge Inference Memory’ (EIM), which it recently unveiled as a next-generation product, at the 18th Semiconductor Exhibition (SEDEX 2026) held in Seoul last week. Micron’s EIM involves stacking low-power DRAM (LPDDR) on top of a logic die responsible for AI computations. The goal is to reduce data processing latency between the AI semiconductor and the memory while simultaneously increasing the bandwidth. This is a different concept from PIM, which embeds computing units inside the memory, and the specific computational architecture and target performance have not yet been disclosed. Micron stated that it is discussing the development of products suitable for the physical AI market, such as autonomous driving, humanoids, augmented reality (AR), and virtual reality (VR), with major clients. The industry expected that Micron will begin mass production in the second half of 2028. For physical AI, LPDDR may be more advantageous than HBM as the power efficiency of battery-operated robots is crucial. However, as AI models become more advanced, the demand for memory bandwidth and capacity increases, meaning that existing LPDDR alone may hit a limit. Micron forecast that high-performance humanoids will require 200GB-level DRAM and 4TB of storage in the future. Considering this, Micron’s new products appear to be targeting new markets such as physical AI and autonomous driving rather than servers. CXMT, which recently brought down the stock prices of global semiconductor companies by sharing news of its next-generation DRAM process development, expressed confidence in the long-term growth of the DRAM industry at an extraordinary general meeting of shareholders on this day. The company stated that it has introduced a self-aligned quadruple patterning (SAQP) process and has made progress in the 4F² architecture, a next-generation structure, announcing that it will pursue process miniaturization and architecture innovation. In the case of 4F², Samsung Electronics and SK hynix have not yet released detailed launch plans either. Of course, the general opinion is that ‘Samsung and SK hynix’, which are ahead in micro-processes, can achieve higher performance with the existing 6F² structure than CXMT’s 4F², but the situation could change if US semiconductor regulations are lifted and China becomes able to bring in extreme ultraviolet (EUV) lithography equipment from ASML. CXMT officially announced the mass production of LPDDR6 in August, and entered mass production in the 5th generation DRAM process last month. It announced that it will also introduce next-generation server products utilizing the architecture by the end of the year. The company forecast that its revenue for the first to third quarters of this year will increase by 725% to 741% compared to the same period last year. However, it is difficult to conclude that the technological gap with leading companies has been resolved solely based on such revenue growth projections. Actual product performance, yield rates, and whether the company secures clients are the variables that will determine future competitiveness. The problem is that the center of the memory market is not fixed. HBM was also not the mainstream of the DRAM market in the past, but it emerged as a core product as demand for generative AI and high-performance graphics processing units (GPUs) surged. If physical AI spreads in the future, the importance of low-power, high-bandwidth memory could grow, and there is a possibility that memory with new architectures will form separate markets. Although the results of new technologies such as EIM have not yet been verified, if the market grows rapidly, the possibility that pursuers like Micron could seize leadership in one fell swoop cannot be ruled out. In fact, domestic battery companies had a similar experience. Korean companies, which led the early electric vehicle market, were caught off guard believing that China’s low-cost lithium iron phosphate (LFP) could not surpass their nickel, cobalt, and manganese (NCM), only to have their leadership completely taken away by China, which leveraged its strong domestic market. As a result, the market share of Korean electric vehicle batteries, which accounted for half of the market excluding China until the early 2020s, has now plummeted to the 10% range. This shows that it is difficult to maintain market leadership with technological competitiveness alone, and that responding to demand changes as well as price and production structures is important. The conclusion drawn is that companies must maintain leadership in the existing market while simultaneously possessing various new technologies that can immediately respond to diverse demands. Moon Song-cheon, a professor emeritus at KAIST, said, “We must not be dragged along by the US strategy with only semiconductors,” adding, “An industrial strategy that looks to the future of the nation, such as fostering software companies, is needed.”