Micron Breaks Ground on a $9.3 Billion Bet to Crack SK Hynix's Grip on AI Memory

Micron broke ground on a $9.3 billion expansion of its Hiroshima plant to produce high bandwidth memory for Nvidia's AI chips, aiming to challenge SK Hynix's dominant market share. The project, subsidized by Japan's government, is expected to start shipments around summer 2028, as the company seeks to close the gap in the HBM market.

Micron broke ground on a $9.3 billion Hiroshima expansion built for one purpose: making the high bandwidth memory that keeps Nvidia's AI chips running, and challenging SK Hynix's grip on that market. On Saturday, in Higashi-Hiroshima, Micron turned the first soil on a 1.5 trillion yen expansion of its existing plant there, according to Bloomberg. The company plans to use the new capacity to build high bandwidth memory, the stacked chips that sit next to processors like Nvidia's and feed them data fast enough to keep up with AI workloads. Shipments are expected to start around the summer of 2028. You don't build a plant this size on a hunch. Micron and Samsung are locked in a distant tie for second place in HBM, well behind SK Hynix, and Micron is spending nearly $10 billion to close that gap. Japan's Ministry of Economy, Trade and Industry is subsidizing the project too, kicking in up to 500 billion yen, or roughly $3.2 billion, according to Bloomberg. That's not charity. Tokyo has spent years trying to rebuild a domestic chip industry it let slip to Taiwan and South Korea in the 1990s, and HBM is the one segment where a Japanese win is still plausible. SK Hynix isn't standing still. The company held around 58 percent of the global HBM market in the first quarter of 2026, with Samsung and Micron splitting most of the rest near 21 percent apiece, according to data compiled by TrendForce. SK Hynix has already locked in Nvidia as its lead customer for years and is racing to double its own wafer capacity by 2030. Micron's Hiroshima bet is the clearest sign yet that it isn't content to stay in third place. Micron already runs a DRAM fab in Higashi-Hiroshima, acquired when it bought Elpida Memory out of bankruptcy back in 2013. Building HBM capacity there instead of at a greenfield site cuts years off the timeline and gives Micron a second geography outside Taiwan, where the bulk of the world's advanced chip packaging still happens. A single earthquake, a single flashpoint near the Taiwan Strait, and a huge share of the AI industry's memory supply is exposed. Spreading HBM production across Japan, Korea and the United States isn't a nice to have anymore. It's how a supplier survives a bad year. Micron CEO Sanjay Mehrotra has been blunt about what HBM actually is. He has called it a technical marvel, not your grandfather's memory, and noted it takes far more wafers to produce than standard DRAM for the same output. That's the real bottleneck behind the AI chip shortage everyone keeps reading about. Fabs can't just spin up more HBM overnight. Building the equipment, qualifying the process and getting yields high enough to ship in volume takes years, which is exactly why Micron started digging in Hiroshima now for chips it won't sell until 2028. Nvidia's next Rubin platform is expected to run on HBM4, and TrendForce has reported that SK Hynix, Samsung and Micron are all racing to qualify 12 layer stacks this year, with 16 layer versions targeted for 2027. Whoever ships qualified HBM4 to Nvidia first locks in the next multi year contract. That's why the 2028 shipment date attached to Hiroshima matters as much as the dollar figure. Frankly, the subsidy math says as much about Tokyo's ambitions as it does about Micron's. Chip manufacturing is a business where the winner is often whoever can borrow money most cheaply and wait longest for a return. Japan is willing to eat roughly a third of Micron's bill because a functioning HBM supply chain on its own soil is worth more than the subsidy check. The same logic explains why the US CHIPS Act, South Korea's K Chips Act and now this Japanese package keep showing up wherever advanced memory or logic gets built. State money has become a normal input cost in this industry, not an exception. For the companies buying HBM, the practical upshot is capacity they can plan around. Nvidia, AMD and the hyperscalers building their own AI accelerators have spent the past two years absorbing every HBM chip SK Hynix, Samsung and Micron can produce, and prices have climbed accordingly. A third credible supplier ramping toward volume by 2028 doesn't fix the shortage this year or next. But it gives chipmakers something they haven't had in this cycle, a second source they can point to when they're negotiating with SK Hynix. Micron hasn't disclosed monthly output targets for the new lines, only that production will scale through the back half of the decade. Whether that's enough to meaningfully dent SK Hynix's lead by 2030 is still an open question. What's not in question is that Micron just committed nearly a decade of capital and Tokyo's money to finding out. Also read: SK Hynix is racing to list on Nasdaq just as the AI chip trade it built its fortune on wobbles https://startupfortune.com/sk-hynix-is-racing-to-list-on-nasdaq-just-as-the-ai-chip-trade-it-built-its-fortune-on-wobbles/ • Hackers Just Showed How Fragile the AI Software Supply Chain Really Is https://startupfortune.com/hackers-just-showed-how-fragile-the-ai-software-supply-chain-really-is/ • Three nuclear startups beat Trump's deadline and one already powers an Nvidia chip https://startupfortune.com/three-nuclear-startups-beat-trumps-deadline-and-one-already-powers-an-nvidia-chip/