Micron, Samsung fabs peak 2028–2030

Memory is the part of the AI stack people keep underestimating. GPUs get the headlines, but the useful work only happens if those GPUs can stay fed with enough DRAM and especially enough HBM — the stacked memory bolted right next to the accelerator. What changed is that the three companies that matter here — Samsung, SK hynix, and Micron — have all laid out timelines that look fast on paper but slow in factory reality. The new capacity is coming, but a lot of the real relief still looks like a 2027-to-2028 story, and some of the biggest additions don’t become meaningful until closer to 2028. ### What exactly is the bottleneck? HBM is high-bandwidth memory — dense stacks of DRAM linked with through-silicon vias and advanced packaging so AI chips can move data absurdly fast. The catch is that HBM burns far more wafer capacity than ordinary DRAM for the same bit output, because stacking, testing, yields, and packaging all get harder. SK hynix has said HBM needs at least twice the capability of general DRAM to secure the same production, which is the cleanest way to understand why “more demand” does not translate into “easy supply.” ### Why are 2028 and 2030 showing up? Because fabs do not matter the day they are announced. They matter after shell completion, tool move-in, process qualification, yield tuning, customer validation, and then a long ramp. Samsung’s own Korean release says its Pyeongtaek P5 line is slated to become a core HBM production base from 2028. SK hynix said in 2024 that M15X would come first, with the first Yongin cluster fab completing in the first half of 2027. (news.skhynix.com) That means the next real step-up in supply is staggered — not immediate. ### Where does Micron fit? Micron is earlier in the U.S. buildout than a lot of people assume. Its Idaho fab is under construction, but outside reporting tied to Micron’s own updates still points to a 2027 production start. Meanwhile, Micron’s business has already shifted hard toward AI memory — in its June 25, 2025 quarter, it said HBM revenue jumped nearly 50% sequentially and data-center revenue more than doubled year over year. (news.samsung.com) In plain English, demand is arriving before the new bricks-and-mortar capacity does. ### Isn’t Samsung already shipping HBM4? Yes — but shipping a leading-edge product is not the same thing as flooding the market with it. Samsung said it has started commercial HBM4 shipments and expects 2026 HBM sales to more than triple versus 2025 while proactively expanding HBM4 capacity. That sounds huge, and it is, but it also tells you demand is outrunning the current base. If sales can triple before the big 2028 line is online, the system is still tight. (trendforce.com) ### Why does AI make this worse? Because AI memory demand compounds in two directions at once. Training clusters need massive HBM per accelerator, and inference is starting to do the same as models get larger and context windows expand. TrendForce’s 2026 HBM outlook says the DRAM industry is now focusing capacity expansion around HBM, while conventional DRAM growth leans more on technology migration than giant fresh wafer additions. (news.samsung.com) Basically — the industry is reallocating scarce capacity toward AI first. ### So is compute no longer the constraint? Not exactly. Compute is still constrained. Power is constrained. Networking is constrained. But memory is becoming the sneaky one — the part that can cap cluster growth even when GPU roadmaps look healthy. If HBM stays sold through and new fabs ramp on a 2027-to-2028 schedule, then the practical limiter for large AI deployments is not “can Nvidia ship chips?” but “can the memory ecosystem keep up?” Micron already said its HBM output was sold out for calendar 2025 and saw strong 2026 demand. (trendforce.com) ### What should people watch next? Watch three things — HBM customer qualification, not just fab construction; packaging capacity, because memory stacks are useless if they cannot be integrated fast enough; and whether Samsung, SK hynix, and Micron start talking about 2027 supply in the past tense rather than the future tense. That is when the bottleneck starts to ease. ### Bottom line? The story is not that memory companies failed to build. (investors.micron.com) It’s that AI demand hit a part of semiconductor manufacturing that ramps slowly and painfully. New fabs are coming. But the market is telling you relief is late enough that memory could be the thing that sets the pace for AI through the next few years.