The Zhitong Finance App learned that SK Hynix (SKHY.US), one of the world's largest memory chip manufacturers, is evaluating the feasibility of building a joint venture storage wafer manufacturing plant (Memory Fab) in Japan and listed sufficient, low-cost electricity and water resources as key site selection conditions. The potential Japanese project will complement South Korea's 54 trillion won production expansion plan and the advanced packaging base in Indiana, USA, while leveraging Japan's advantages in high-end semiconductor equipment, semiconductor raw materials, large customers, government subsidies, and the NAND flash memory leader Kioxia industry chain. This move is more like SK Hynix's global production capacity upgrade in the face of a serious shortage of memory chips, and capacity allocation after comprehensive cost and geographical risks. Instead of AI computing power fundamentals showing signs of oversupply or weakening demand.
The boom in memory chips is still dominated by investment in AI computing power infrastructure, and the AI computing power industry chain level continues to release positive signals at the level of AI computing power demand. South Korea's exports in July increased 62.8% year over year to 98.89 billion US dollars, of which semiconductor exports, which are critical to the Korean economy, surged 178.8% to 41.01 billion US dollars; according to the survey, economists expect South Korea's exports to increase sharply by 62.6% on a high base in August, but official data will not be released until September 1.
The actual AI semiconductor production capacity and order side seem to be more tight. SK Group CEO Choi Taiyuan said that customers are requesting a 60% to 100% increase in AI semiconductor supply in 2027 compared to 2026, and the additional production capacity is very limited, possibly up to 50%; SK Hynix CEO Guo Luzheng expects that 2027 may experience the worst shortage of memory chips in the history of the storage industry, and customer demand may still be higher than the company's supply capacity even after 2030.
SK Hynix is considering a joint venture to build a memory chip fab in Japan to expand supply for the AI boom
SK Hynix is studying the feasibility of setting up a joint venture to produce memory chips in Japan; this is one of the various solutions the company is considering to meet the ever-surging demand for AI semiconductors while controlling production costs.
SK Group Chairman Choi Tae-won said in an interview on the sidelines of a meeting of members of the Korea-Japan Chamber of Commerce and Industry in Sendai, that the South Korean dynamic random access memory (DRAM) and flash memory (NAND) supply giant is examining multiple candidate sites for a large Japanese memory chip manufacturing plant that is jointly operated.
The boom in artificial intelligence has made SK Hynix, headquartered in Icheon, South Korea, the brightest star asset of SK Group under Choi Tae-won. Its products are helping Nvidia and the largest data center developers to provide the most core AI computing power resources. SK Hynix held a groundbreaking ceremony for a large-scale advanced chip packaging facility in Indiana, USA last week, and the company is also exploring further expanding overseas business.
“We are speeding up our investigation all over Japan,” Choi Tae-won told reporters. “Any place with good electricity and water resources will do.” When asked about potential partners or specific locations within Japan, he declined to elaborate further. SK Group's business spans many fields such as energy, telecommunications, and chips, and is responsible for the overall capital allocation of the Group's major investments.
In order to meet AI's huge demand for data storage, and at the same time respond to America's efforts to curb China's technological ambitions, SK Hynix is searching for possible chip production and packaging bases around the world. The company plans to spend 54 trillion won (about 39 billion US dollars) to expand chip manufacturing facilities in South Korea, while also building a large-scale advanced memory chip packaging facility in West Lafayette, Indiana.
The search for a factory site in Japan is part of SK Hynix's broader production capacity expansion efforts to further cooperate with Japanese customers and suppliers. The company already has an indirect shareholding in Kioxia, a global leader in flash memory manufacturing headquartered in Tokyo; its CEO Guo Luzheng said last week that the company is “carefully studying how to develop the NAND flash chip market together with customers and suppliers.”
As with high-bandwidth memory, demand for NAND has also increased dramatically; global cloud computing giants such as Facebook's parent company Meta Platforms and Amazon, and companies that have invested heavily in AI data center construction in recent years are competing to secure larger data NAND storage resources in the next few years and pursue more powerful performance than traditional HDD enterprise hard drives.
One of SK Hynix's competitors, Micron Technologies, headquartered in the US, has a large manufacturing plant in Hiroshima in western Japan. Japan has brought together many suppliers of SK Hynix — from unlisted Namics to semiconductor equipment supergiant Tokyo Electronics, and the largest customers including Sony Group and Nintendo. The Japanese government has also shown a willingness to provide huge subsidies to international chipmakers such as TSMC for projects to expand production capacity in Japan.
“Now is the time for Korea and Japan to form an economic circle,” Choi Tae-won, who is also the chairman of the Korea Chamber of Commerce and Industry, said at an event focusing on the shrinking labor force between the two countries before the annual meeting of the presidents of the Korea-Japan Chamber of Commerce and Industry was held.
AI swallows memory chips, HBM eats more wafers! SK Hynix's bull market trajectory has not stopped
As of the beginning of the Asian session on Monday, SK Hynix's Korean stock market trading price fell 1.45%, hovering around 1.63 million won per share. During the night trading session of US stocks on Monday, SK Hynix's US stock ADR (SKHY.US) hovered around $157.8.
Wall Street is still generally optimistic about the AI storage supercycle, but the ranking of stock price elasticity is closer to “SK Hynix's HBM leadership — Micron's pure storage attributes for US stocks — Samsung's catch-up and repair elasticity.” Goldman Sachs gave SK Hynix a target price of up to 3.5 million won, Citi gave 3.1 million won, and Morgan Stanley offered a target price of 2.6 million won.
The investment logic of the three major storage giants is not exactly the same: SK Hynix is the original target with HBM's most prominent leadership position and the most sensitive AI storage price; Micron is the pure standard for the memory chip with the highest liquidity in the US dollar memory chip market; and Samsung is the target of high operating leverage brought about by the advantage of scale and HBM's catch-up.
Large AI training sets, model weights, checkpoints, vector databases, RAG corpus, multi-modal data, logs, and inference results all need to reside in high-capacity storage for a long time; during training, large-scale data needs to be continuously fed from object storage and local NVMe SSDs to the GPU cluster; the inference era further generates massive KV caches, long contexts, agent states, and retrieval data.
HBM is responsible for the “thermal data layer” with the highest bandwidth, enterprise-grade high-performance DRAM is responsible for the working memory of the system, while enterprise-grade NAND SSDs carry a “thermal data and persistence layer” with far greater capacity than HBM and lower cost. Therefore, NAND is not a replacement for HBM, but rather co-expands the capacity with HBM and DRAM in the AI server's storage tier.
In the AI era, there is an urgent need not only for computational models, but also for continuous handling, storage, and retrieval of massive state data with low latency. HBM and DRAM are responsible for the highest bandwidth working set near the GPU, but model weights, training data sets, checkpoints, vector databases, RAG knowledge bases, inference logs, long-term memory of agents, and some KV cache layers make it impossible to reside all in expensive and limited capacity HBM; as a result, enterprise-grade NVMe SSDs assume a high-performance persistence layer between HBM/DRAM, HDD, and object storage.
According to a research report by TrendForce, a well-known market research agency, the agency expects a cumulative increase of about 270% in server DRAM contract prices and a cumulative increase of about 235% in 2026; HBM contract prices may still rise 70% to 140% in 2027. The contract price for data center server-level DRAM in the third quarter of 2026 is expected to rise 13% to 18% month-on-month from last week's high base in the second quarter, while the supply of RDIMM bits will only increase by 15% to 20% in 2027, which clearly lags behind the total demand for server DRAM shipments.
The latest storage industry chain and dynamic data on demand for the entire AI computing power industry chain show that the expansion of demand for memory chips brought about by AI computing power demand is rapidly spilling over from HBM/DRAM to NAND, and sky-level AI inference workloads are becoming the core engine of this round of changes in demand structure for NAND.
An unprecedented surge in storage demand driven by the continued explosive expansion of AI computing power demand, the increasingly crowded out production capacity of HBM storage with extremely complex manufacturing and packaging processes, and the long-term lack of flexibility in general DRAM/NAND supply have all catalyzed the current “super supply shortage cycle” in the field of AI computing power-related memory chips.