One of the world's largest memory chipmakers, SK hynix, is evaluating the feasibility of establishing a joint-venture memory wafer fabrication plant in Japan, with the availability of ample, low-cost electricity and water resources as key site-selection criteria. A potential Japanese project would complement a 54 trillion won expansion plan in South Korea and an advanced packaging base in Indiana, USA, while leveraging Japan's advantages in high-end semiconductor equipment, raw materials, major customers, government subsidies, and the industry chain relationships of NAND flash leader Kioxia.

This move appears more like a global capacity upgrade by SK hynix in response to a severe memory chip shortage, alongside a capacity allocation strategy weighing comprehensive costs and geopolitical risks, rather than a signal of oversupply or weakening demand in the AI computing fundamental landscape. The memory chip sector's momentum remains driven by AI infrastructure investment, and positive signals regarding AI computing demand continue to emerge across the supply chain. South Korea's July exports surged 62.8% year-over-year to $98.89 billion, with semiconductor exports—critical to the nation's economy—soaring 178.8% to $41.01 billion. Surveys indicate economists expect August exports to grow substantially by 62.6% on a high base, though official data won't be released until September 1st.

Actual AI semiconductor capacity and order books appear even tighter. SK Group CEO Chey Tae-won stated that customers are requesting a 60%–100% increase in AI semiconductor supply in 2027 compared to 2026, while new capacity is very limited, possibly maxing out at a 50% increase. SK hynix CEO Kwak Noh-Jung predicts that 2027 could witness the most severe memory chip shortage in the industry's history, with customer demand potentially outstripping the company's supply capacity well beyond 2030.

SK hynix is studying the feasibility of establishing a joint venture in Japan to produce memory chips, marking one of several options the company is considering to meet surging AI semiconductor demand while managing production costs. SK Group Chairman Chey Tae-won, speaking on the sidelines of a Korea-Japan business conference meeting in Sendai, revealed that the South Korean DRAM and NAND supplier is examining multiple candidate locations for a large, jointly-operated memory fab in Japan. The AI boom has transformed SK hynix, based in Icheon, South Korea, into the crown jewel of Chey's SK Group, with its products powering the core AI computing resources for Nvidia and the largest data center developers.

Last week, SK hynix broke ground on a major advanced chip packaging facility in Indiana, USA, and the company is exploring further overseas expansion. "We are scouting around Japan at a fast pace," Chey told reporters. "Anywhere with good electricity and water conditions would work." He declined to elaborate on potential partners or specific locations within Japan. SK Group's business spans energy, telecommunications, and chips, with overall capital allocation for major investments managed centrally. To meet AI's massive data storage demands while navigating US efforts to curb China's technological ambitions, SK hynix is scouting potential chip production and packaging sites globally. The company plans to invest 54 trillion won (approximately $39 billion) to expand domestic manufacturing facilities, while also building a large advanced memory packaging facility in West Lafayette, Indiana.

The Japanese site search is part of a broader capacity expansion effort by SK hynix to deepen collaboration with Japanese customers and suppliers. The company already holds a significant indirect stake in Tokyo-based NAND flash giant Kioxia, and CEO Kwak Noh-Jung stated last week that the firm is "prudently studying how to jointly develop the NAND flash market with customers and suppliers." Like high-bandwidth memory, NAND demand has surged as global cloud giants such as Meta Platforms and Amazon, along with enterprises heavily investing in AI data centers, compete to secure ever-larger NAND storage resources for years ahead, seeking stronger performance than traditional HDDs.

One of SK hynix's competitors, US-based Micron Technology, operates a major manufacturing plant in Hiroshima, western Japan. Japan is home to many of SK hynix's suppliers—from privately-held Namics to semiconductor equipment giant Tokyo Electron—as well as major customers including Sony Group and Nintendo. The Japanese government has also shown willingness to provide substantial subsidies for international chipmakers like TSMC to expand capacity in the country. "Now is the time for Korea and Japan to form an economic bloc," said Chey, who also heads the Korea Chamber of Commerce and Industry, at an event focused on the two nations' shrinking workforces ahead of the annual Korea-Japan chamber presidents' meeting.

As of early Asian trading on Monday, SK hynix shares in South Korea were down 1.45%, hovering near 1.63 million won per share. During Monday's US overnight session, the company's ADR was trading around $157.80. Wall Street remains broadly bullish on the AI memory supercycle, though rankings of stock resilience lean toward "SK hynix's HBM leadership—Micron's pure-play memory status—Samsung's catch-up and repair elasticity." Goldman Sachs has set a target price of 3.5 million won for SK hynix, Citi targets 3.1 million won, and Morgan Stanley targets 2.6 million won. The investment theses for the three memory giants differ: SK hynix stands out for its HBM leadership and sensitivity to AI memory pricing, enhanced by a 40 trillion won buyback and cancellation plan; Micron offers the purest exposure to memory chips with the highest liquidity; and Samsung presents operating leverage from scale advantages and HBM catch-up potential.

AI's massive training sets, model weights, checkpoints, vector databases, RAG corpora, multimodal data, logs, and inference results all require long-term residency in high-capacity storage. During training, large-scale data must continuously flow from object storage and local NVMe SSDs into GPU clusters; the inference era further generates massive KV caches, long contexts, agent states, and retrieval data. HBM handles the highest-bandwidth "hot data layer," enterprise high-performance DRAM serves as system working memory, and enterprise NVMe SSDs handle the larger, lower-cost "warm data and persistent layer." Therefore, NAND isn't replacing HBM but rather scaling alongside HBM and DRAM within the AI server storage hierarchy.

The AI era demands not only computational models but also continuous movement, preservation, and low-latency retrieval of massive state data. While HBM and DRAM manage the highest-bandwidth working sets near GPUs, model weights, training datasets, checkpoints, vector databases, RAG knowledge bases, inference logs, agent long-term memory, and parts of KV cache layers cannot all reside permanently in expensive, capacity-limited HBM. Enterprise NVMe SSDs thus serve as the high-performance persistent layer between HBM/DRAM and HDD/object storage. A recent report from TrendForce projects server DRAM contract prices to cumulatively rise approximately 270% by 2026, with enterprise SSD prices up about 235%; HBM contract prices in 2027 could still rise 70%–140%. In Q3 2026, data center server-grade DRAM contract prices are expected to rise 13%–18% quarter-over-quarter from an already elevated Q2 base, while 2027 RDIMM bit supply growth of only 15%–20% lags significantly behind total server DRAM shipment demand.

The latest demand data from the memory supply chain and the broader AI computing ecosystem collectively show that storage demand from AI is rapidly spilling over from HBM/DRAM into NAND, with massive AI inference workloads emerging as the core engine driving this structural shift in NAND demand. The unprecedented storage demand surge driven by explosive AI computing growth, capacity constraints from increasingly complex HBM manufacturing and packaging, and chronically inelastic general DRAM/NAND supply together fuel the current "super supply shortage cycle" in AI-related memory chips.