Micron Technology broke ground Saturday on a roughly ¥1.5 trillion, or $9.3 billion, expansion of its factory in Higashi-Hiroshima, Japan, aimed at producing high-bandwidth memory for AI processors. Shipments from the new production line are expected to start around summer 2028, according to Bloomberg reporting carried by Moneycontrol and The Japan Times.
Japan’s Ministry of Economy, Trade and Industry has allocated up to ¥500 billion to support the project. That public backing matters because HBM is no longer a niche memory product attached to high-end accelerators. It is one of the constraints shaping how quickly AI data centers, model providers, chip designers, and cloud platforms can expand.
Why Hiroshima matters to AI capacity
HBM sits close to GPUs and other AI accelerators, giving them the memory bandwidth needed to move huge amounts of model data quickly. More powerful AI systems do not only need more compute. They need memory capacity, bandwidth, power efficiency, advanced packaging, and enough finished supply to match accelerator roadmaps.
Micron has been unusually direct about that shift. In its fiscal Q3 2026 prepared remarks, the company said data center revenue exceeded $25 billion in the quarter, an annualized run rate above $100 billion, and that DRAM and NAND demand continues to run ahead of supply. Micron expects tight memory conditions to persist beyond calendar 2027 because of AI-driven demand and structural supply limits.
The company’s explanation is useful because it gets past the generic phrase “AI chip shortage.” New fabs take years to build, cleanroom space is scarce, permitting and energy infrastructure can slow projects, and each new generation of memory process technology is harder to manufacture. Micron also said HBM’s growth and its increasing trade ratio with every generation put additional pressure on non-HBM supply, since capacity shifted toward HBM cannot simply be treated as ordinary commodity memory output.
The 2028 date is the point
The Hiroshima expansion will not fix today’s allocation problem. A summer 2028 shipment target means hyperscalers and AI hardware companies still have to plan through at least two more years of constrained memory supply while model training, inference, agent workloads, and context storage keep increasing demand.
That is why the project is better read as a supply-chain signal than as a single factory story. AI infrastructure planning is moving from quarter-to-quarter GPU access toward multi-year control over the components around the accelerator: HBM, server DRAM, SSDs, networking, power, cooling, and packaging capacity. Memory is becoming part of the contract strategy, not just the bill of materials.
Micron has already tied that strategy to customer commitments. In the same Q3 materials, the company said it had signed 16 strategic customer agreements across data center, consumer, and automotive markets. Those agreements typically run from 2026 through 2030, cover roughly 20% of Micron’s DRAM volume and one-third of its NAND volume over the period, and include take-or-pay commitments for specific volumes. Fourteen of the agreements carry roughly $100 billion in remaining performance obligations at minimum contracted pricing, with projected cash deposits and related commitments of $22 billion.
For AI buyers, the implication is straightforward: advanced memory access is being reserved years in advance by the largest customers. Smaller cloud providers, enterprise infrastructure buyers, and hardware startups may find that HBM availability and pricing are as important to deployment plans as accelerator supply itself.
Japan’s role is not only the fab
The Hiroshima site also gives the project a supply-chain angle beyond Micron’s own capacity. Japan remains strong in semiconductor materials and manufacturing equipment even though it lost leadership in finished logic and memory chips decades ago. Kota Nosaka, representative director of Micron’s Japan unit, told Bloomberg that roughly 80% of the materials needed by the Hiroshima factory now come from Japan.
Micron took possession of the Hiroshima factory through its 2013 acquisition of bankrupt Japanese DRAM maker Elpida Memory. The new expansion turns that older memory-manufacturing footprint into part of the AI hardware race, backed by a government that wants more advanced semiconductor production tied to domestic suppliers.
The Japan project sits alongside Micron’s broader global buildout. The company’s U.S. expansion plan includes two leading-edge fabs in Idaho, up to four in New York, modernization in Virginia, investments in R&D, and end-to-end advanced HBM packaging capabilities. In its Q3 remarks, Micron also pointed to Taiwan and Singapore capacity, including Singapore HBM packaging expected to contribute meaningfully beginning in the first half of 2027.
What readers should watch next
The most important question is not whether Micron can announce enough construction. It is whether the new capacity comes online fast enough, with yields and packaging capacity good enough, to meet demand from AI accelerator roadmaps that are also moving quickly.
Watch for three practical signals. First, whether Micron’s Hiroshima timeline holds as equipment, labor, energy, and permitting pressures build. Second, whether HBM4 and later HBM generations ramp at yields that support high-volume AI systems rather than only premium deployments. Third, whether long-term supply agreements make memory pricing more predictable for big customers while leaving smaller buyers exposed to allocation pressure.
The AI infrastructure race has often been described as a GPU race. Micron’s Hiroshima expansion is a reminder that the next bottleneck may be the memory stacked beside those processors, the cleanrooms needed to make it, and the customers able to reserve it before everyone else.
