NextFin News - SK Hynix is weighing a memory chip fabrication plant in Japan's Miyagi Prefecture, a move that would make it the first South Korean chipmaker to make a large-scale semiconductor manufacturing investment in Japan, as the world's leading supplier of AI memory races to add capacity amid a shortage that has left high-bandwidth memory sold out through 2026. The project is still under review - SK Hynix says no final decision has been made - and a potential investment could run into tens of trillions of won. But the mere existence of the talks marks a strategic pivot: the company whose supply chain was restricted by Japan's 2019 export controls is now considering building inside the very country that imposed them.
The shares tell part of the story. SK Hynix closed near 1.65 million won on the Korea Exchange on August 28, down about 4.5% on the day, as investors weighed the company's heavy capital-spending commitments against a memory market trading at a premium to its cyclical history. The pullback came after a week in which the Miyagi talks moved from rumor to reported fact, and it underscores the central tension of the moment: SK Hynix is being asked to spend tens of billions of dollars on capacity that will not produce a chip until 2029, in a market where today's scarcity could be tomorrow's glut.
In an official statement, SK Hynix said any location with the necessary infrastructure could be considered a candidate, but no decision has been made regarding Miyagi. Industry sources, however, say the company is moving forward with plans for the site, and SK Group Chairman Chey Tae-won has personally visited the region. If realized, the fab would place SK Hynix in Japan's northeastern Tohoku semiconductor hub alongside plants operated by Taiwan's TSMC in Kyushu, U.S. memory maker Micron in Hiroshima, and Rapidus - the Japanese government-backed venture working to produce 2-nanometer logic chips on Hokkaido.
The timing is the point. Research estimates put SK Hynix's share of the global HBM market at roughly 62% as of mid-2025, with NVIDIA estimated to account for about 90% of its HBM customer base. That capacity is already committed: HBM production for 2026 is reported sold out across all three major suppliers - SK Hynix, Micron and Samsung - and data centers are estimated to consume about 70% of all memory chips made worldwide. Building a fab takes years, not quarters. SK Hynix is not deciding where to make chips for next year; it is deciding where its AI supply chain will sit for the next decade.
A Fab in Miyagi Is About Supply Chains, Not Just Capacity
On the surface, a new memory fab is a simple capacity story: demand exceeds supply, so build more. That reading misses the mechanism. Memory fabrication is only the final assembly of a much longer chain, and Japan controls the upstream choke points. Although Japan no longer produces a large share of finished semiconductors, it remains globally dominant in the materials, components and equipment required to manufacture chips at the most advanced process nodes.
Miyagi Prefecture is not an empty greenfield. It is already a production base for Tokyo Electron, Asia's largest semiconductor equipment maker, which is investing 104 billion yen in a new equipment plant there. The prefecture hosts Sony's Shiroishi-Zao Technology Center for image sensors, Toyota Motor East Japan's compact-car plants, and Tohoku University in Sendai, regarded as a training hub for semiconductor specialists. The Japanese government has designated Miyagi as one of three national semiconductor industry hubs, alongside Kyushu and Hokkaido.
For SK Hynix specifically, the supply-chain logic is intimate. The underfill - a specialized liquid adhesive essential to stacking HBM dies - is exclusively supplied by Japanese chemical and materials company Namics. SK Hynix's main suppliers of silicon wafers and photoresists, the light-sensitive materials that define microscopic circuits, are also Japanese firms, including Shin-Etsu Chemical and Tokyo Ohka Kogyo. Locating fabrication next to those suppliers shortens lead times, reduces logistics risk and, critically, changes the company's legal posture.
"Japan, as a chipmaking country, has the entire ecosystem that would be needed, including electric power and materials. When considering [investment locations] outside Korea, it's certainly an excellent candidate," SK Group Chairman Chey Tae-won said in a June 10 interview.
The legal posture matters because Japan has already shown it is willing to weaponize the supply chain. In 2019, Tokyo imposed export controls on three semiconductor materials to South Korea following a Korean Supreme Court ruling on forced-labor compensation from the colonial period. That episode is not ancient history for chip executives; it is the reference case for every location decision made since.
"Japan designated semiconductors as strategic products when it passed the Economic Security Promotion Act in 2022, paving the way for imposing export controls on the equipment used to manufacture them as well. Running a semiconductor fab inside Japan could allow [Korean companies like SK Hynix] to avoid such restrictions," said Kim Yang-paeng, a researcher at the Korea Institute for Industrial Economics and Trade.
That is the first-order mechanism: a fab inside Japan is partly an insurance policy against the very controls Japan itself can impose. The second-order effect is subtler. By establishing a local production base, SK Hynix gains political capital in Tokyo - the kind of capital that becomes decisive when export licenses, subsidies or antitrust approvals are on the line.
The Kioxia Card: Leverage, Ownership and a Negotiating Table
That capital is being accumulated in more ways than one. SK Hynix has emerged as the de facto largest shareholder of Kioxia, Japan's NAND flash memory champion and the company formerly known as Toshiba Memory. A special purpose company that SK Hynix can control - BCPE Pangea Cayman2, affiliated with Bain Capital - now holds 77.4 million Kioxia shares, a 14.19% stake, edging past Toshiba's 14.06% after Toshiba sold shares in seven batches between July 15 and August 3. On paper the SPC is the top holder; in substance, SK Hynix holds convertible bonds that could secure nearly all of the SPC's shares.
SK Hynix cannot exercise voting rights immediately. Conversion requires antitrust approvals in multiple countries and clearance under Japan's foreign investment procedures. But the position gives SK Hynix something rarer than dividends: influence over a Japanese national champion whose market value briefly overtook Toyota's in June on the AI-driven memory rally.
Here the negotiation cuts both ways. A Japanese semiconductor expert, speaking on condition of anonymity, put the leverage plainly: when SK Hynix eventually seeks to sell its stake in Kioxia, Japan's Ministry of Economy, Trade and Industry may attempt to block the sale on grounds of economic security or technology transfer. The implication is that SK Hynix could use a promise to reinvest the proceeds from any eventual sale into Japan - into a place like Miyagi - as bargaining chips with Tokyo.
Put the two moves together and the shape of the strategy appears: a Korean memory champion embedding itself in Japan through both equity (Kioxia) and physical capital (a Miyagi fab), converting what was once a purely adversarial relationship into a web of mutual dependence. It is a hedge against geopolitics by making disentanglement costly for both sides.
The Capacity Race: Korea, the U.S., and Now Japan
The Japan talks sit inside a much larger capital-spending wave. On August 7, SK Hynix said it would invest 54 trillion won ($38.1 billion) to build two new memory plants in South Korea: the Y2 fab in Yongin, a 35.2 trillion won facility for HBM and next-generation DRAM, and the M17 fab in Cheongju, a 19.1 trillion won NAND facility. Construction is scheduled to begin in July 2027, with the first cleanroom expected to open in June 2029. These investments are the next tranche of a master plan that commits 600 trillion won to the Yongin Semiconductor Cluster and 100 trillion won to expand Cheongju.
Three weeks later, on August 27, the company broke ground on its Indiana fab in the United States. And on August 19, SK Hynix accelerated a 40 trillion won share repurchase and cancellation program while pledging to expand shareholder returns to "over 50% of FCF." The company is simultaneously spending tens of billions of dollars on future capacity, rewarding shareholders today, and - now - scouting a fourth country for production. That is not a company managing a cyclical upswing. It is a company positioning for a structural shift in where AI memory is made.
The demand side justifies the ambition. HBM is expected to account for roughly 25% of total DRAM wafer production by 2026, with demand growing around 70% year on year. HBM4 is entering production this year, with 16-high stacks targeting the fourth quarter. Producing one bit of HBM displaces several bits of conventional DRAM output, so every wafer shifted to HBM tightens the rest of the market. Reports citing company guidance suggest the shortage could persist past 2027, and in some assessments past 2030. The market's total addressable value for HBM is projected to reach $100 billion by 2028, up from $35 billion in 2025 - a compound annual growth rate near 40%.
This is the second-order point the market has not fully absorbed: memory, not logic silicon, is becoming the binding constraint on AI infrastructure buildout. NVIDIA and its peers can design ever-larger accelerators, but each new generation depends on HBM stacks that take years of fab construction to bring online. The bottleneck has moved upstream, and whoever controls HBM capacity controls the pace of the AI buildout.
The Counter-Thesis: A Cyclical Supercycle Priced to Perfection
The strongest case against this reading is the oldest rule in semiconductors: memory is cyclical, and every supercycle ends in oversupply. The 2017-2018 memory boom collapsed into a downturn in 2019, precisely the year Japan imposed its export controls. Today's rally is being driven by AI data centers rather than consumer electronics, but the economics of capacity are unforgiving - when the fabs now under construction come online in 2028 and 2029, they will arrive simultaneously, and any slowdown in AI capital expenditure would turn committed capacity into a glut.
There are already signs that the pricing peak may be nearer than the bulls assume. Analysts at Mirae Asset Securities expect HBM bit shipments to grow 16% in 2026 but project the average selling price for HBM products could fall about 7.4% from the prior year as competitors qualify their processes. A price decline on a product line that is supposed to be sold out would signal that the scarcity premium is already eroding. Chinese producers CXMT and YMTC, while still confined to lower-end segments with an estimated 5-10% share and no leading-edge HBM capability today, represent a longer-term margin threat that the current pricing power does not reflect.
There is also the question of whether the Miyagi project is real strategy or negotiating theater. SK Hynix has said no decision has been made. A fab announcement that never breaks ground can still extract Japanese subsidies, Kioxia sale clearance, or export-control concessions - and then quietly lapse. The company's own capital discipline argues against overreach: it just committed 54 trillion won at home and broke ground in Indiana. Adding a Japan fab on top of that within a single quarter would stretch management bandwidth and capital allocation.
The falsifying signal is specific: if SK Hynix does not file a formal investment application with Japanese authorities or announce a subsidy agreement with the Japanese government within 12 months, the Miyagi talks should be read as leverage, not a buildout. And if HBM average selling prices fall more than 10% year on year in any quarter of 2027 while bit shipments grow below 10%, the structural-scarcity thesis is wrong - the market would be entering the downleg of a cycle, and today's capacity bets would be timed to the peak.
What Comes Next: Three Horizons, Three Scenarios
Short term (sentiment and supply-chain de-risking). The news itself is accretive: it signals that SK Hynix is treating Japan as a viable production base, which reduces the geopolitical discount investors have applied to Korea-based memory capacity. Watch for any statement from Japan's Ministry of Economy, Trade and Industry on subsidies or strategic designation, and for movement on the Kioxia convertible-bond conversion.
Medium term (capacity, 2027-2029). The binding question is timing. The Yongin Y2 and Cheongju M17 fabs do not produce their first cleanroom output until mid-2029. A Miyagi decision would sit on a similar timeline. The risk is that all three major suppliers - Samsung, SK Hynix and Micron - bring new HBM capacity online in the same window. The beneficiary in that scenario is the customer, not the supplier; the exposed party is whichever producer built the highest-cost capacity.
Long term (structural realignment). This is where the thesis stands or falls. If SK Hynix builds in Miyagi, it would mark the beginning of a structural re-knitting of the Korea-Japan semiconductor relationship - from the adversarial 2019 export-control episode to a mutually dependent production network. That would be a regime change, not a cycle. If the talks lapse, the 2019 playbook remains in force, and SK Hynix's Japan exposure stays limited to the suppliers it cannot avoid.
Base case: SK Hynix announces a phased Miyagi investment tied to Japanese government subsidies, with a partnership structure that keeps it below the thresholds that would trigger a full economic-security review. Upside case: a deeper tie-up with a Japanese partner - potentially involving Kioxia capacity or a joint venture - that gives SK Hynix preferential access to Japanese materials and equipment. Downside case: Tokyo blocks or delays the Kioxia conversion on national-security grounds, poisoning the atmosphere and pushing the Miyagi plan into indefinite review.
The specific signals to watch: a formal investment filing in Japan; METI's subsidy decision; the conversion of SK Hynix's Kioxia-linked convertible bonds; HBM4 qualification results for SK Hynix, Samsung and Micron; and quarterly HBM average-selling-price data. Any two of those printing against the base case should shift the thesis.
The real story here is not a new factory. It is that the AI memory shortage has become severe enough, and durable enough, that a Korean champion is willing to build inside the country that once restricted its supply chain - and that Japan, for its part, may prefer a Korean fab on its soil to a Korean shareholder in its national champion. In the semiconductor age, supply chains are not just logistics. They are foreign policy.
Explore more exclusive insights at nextfin.ai.
