HIROSHIMA — The ground-breaking ceremony on July 4 drew a modest crowd: local officials, plant managers, and a delegation from Micron Technology’s headquarters in Boise, Idaho. Behind the podium, construction equipment idled at the edge of a hillside site on the outskirts of Hiroshima. By the time the ribbon was cut, Micron had set in motion the largest single investment in the history of the plant it runs here.
The expansion carries a total investment of ¥1.5 trillion, roughly $9.3 billion, and is designed to produce high-bandwidth memory, or HBM, along with other advanced memory chips. Micron expects the site to begin shipping product around the summer of 2028. HBM — stacks of DRAM dies wired together and placed directly beside an AI accelerator — has become the most contested component in the artificial-intelligence supply chain, and only three companies in the world can make it at scale: Micron, Samsung Electronics, and SK Hynix.
All three are building at once. Last week, SK Hynix said it would invest 80 trillion won, about $51.5 billion, in a new NAND flash factory in Cheongju, South Korea. Samsung has its own expansion under way and has described it as the largest capacity increase in its history. Within roughly a month, the three memory giants have committed well over $100 billion to new production, most of it aimed at the same customer: the data centers where AI models are trained and served.
Hiroshima is a strategic choice for Micron. The plant is the company’s main development site for HBM in Japan, and Micron has deepened its roots there since acquiring the facility from the bankrupt chipmaker Elpida Memory in 2013. Tokyo has backed the project with subsidies, part of a national campaign to rebuild advanced semiconductor manufacturing on Japanese soil that has also drawn TSMC to Kumamoto. For the prefecture, the expansion means thousands of construction jobs now and thousands of manufacturing jobs later.
The demand picture explains the scale of the bet. Nvidia’s newest AI accelerators each require several HBM stacks, and the memory industry has been unable to keep up for two consecutive years. By the start of 2026, the three major producers had effectively sold out their HBM output for the year, according to industry estimates. Prices for conventional DRAM have followed, with contract prices rising by double-digit percentages in each of the past two quarters, a trend Samsung confirmed this month when it told customers it planned another substantial increase in the third quarter.
Goldman Sachs analysts have described the global memory market’s supply-demand imbalance as the most severe in nearly 15 years, driven by AI servers and inference cards. The memory industry, they note, has become the main bottleneck in the AI build-out — the place where chip orders actually meet factory capacity. That has made memory makers the clearest beneficiaries of the AI boom, but it has also loaded them with the heaviest capital obligations in their history.
The risk is that everyone is building at once. The memory business has a long record of overbuilding during booms and then watching prices collapse when new factories come online together. Most of the capacity now under construction is scheduled to arrive in 2027 and 2028, which means the industry could face a glut just as the first chips from this Hiroshima expansion reach the market. Micron’s own history is cautionary: the company cut production and idled lines in 2023 when a sharp downturn wiped out profits across the sector.
Micron enters this cycle in better shape than in past booms. Revenue and operating profit have hit records in each of the past several quarters as prices recovered, and the balance sheet, which carried heavy debt through the 2023 downturn, has been repaired by two years of strong cash flow. The company has also narrowed the gap with SK Hynix, which dominates HBM with roughly half the market, and is working to qualify its next-generation memory with Nvidia and other accelerator makers.
The industry is already moving toward HBM4, a generation that ties memory directly into the logic chip beneath it and requires even tighter cooperation between memory makers and chip designers. Micron has said its next-generation products will be ready in time for the 2028-2029 product cycle, a schedule that makes the Hiroshima build’s timing all the more important. The company has also signaled that it will expand packaging capacity alongside the fab, since stacking and testing the chips is where much of the value now sits.
For Japan, the project is a rare piece of good news in a sector it largely lost decades ago. For Micron, it is a bet that the AI demand wave will last long enough to fill the factory. The machines will not arrive until 2028 — a long wait in an industry where demand can turn in a quarter. If the AI build-out continues, Hiroshima becomes one of the most valuable memory plants on earth. If it falters, Micron will be left holding another monument to a cycle it has lived through many times before.
