Micron says HBM capacity effectively full

Micron’s news is simple on the surface but pretty important underneath: the memory going into top-end AI chips is basically already spoken for. In March, Micron said it had completed price and volume agreements for its entire calendar 2026 HBM supply, including HBM4, while posting record fiscal Q2 2026 results — $23.86 billion in revenue and a 74.4% GAAP gross margin. ### What is HBM, and why does it matter? HBM means high-bandwidth memory. It is the stacked, ultra-fast memory that sits right next to AI GPUs and feeds them data fast enough to keep giant training and inference workloads moving. If regular DRAM is system memory, HBM is the premium fuel line for the most expensive compute in the rack. Micron’s own HBM3E and HBM4 products are aimed straight at that market. ### What exactly did Micron say? The key line was not vague. Micron said it had “completed agreements on price and volume” for its entire 2026 HBM supply, including HBM4. That means customers are not just expressing interest — they have reserved output on negotiated terms. The company also pulled forward its HBM market forecast, now expecting the market to grow from about $35 billion in 2025 to around $100 billion in 2028, two years earlier than its prior outlook. ### Why is that a bigger deal than one company selling well? Because HBM is not easy to add in a hurry. You need advanced DRAM, advanced packaging, clean yields, and close co-design with GPU platforms. Micron said its Singapore advanced packaging facility is on track to contribute meaningfully in 2027, which is another way of saying more supply is coming — but not fast enough to loosen 2026. Micron also began volume shipment of its 36GB 12-high HBM4 in the first quarter of calendar 2026 for NVIDIA Vera Rubin, so this is already tied to the next AI platform cycle, not some distant roadmap. ### Why do margins tell the same story? A 74.4% GAAP gross margin is wild for a memory company. Memory has usually been cyclical and brutally price-sensitive. Numbers like this tell you Micron is selling into a market where demand is outrunning available supply, especially in premium products. The company’s Q3 guide was even stronger, and management explicitly tied the step-up to AI demand, structural supply constraints, and execution. Basically, if customers had easy alternatives, margins would not look like this. ### Does this only hit HBM buyers? No — that’s the catch. When leading-edge DRAM and packaging lines are heavily loaded with HBM, the squeeze can spill into adjacent memory categories and into full server build schedules. The bottleneck stops being “can I buy a GPU?” and becomes “can I get the whole system delivered with the right memory, thermals, and power envelope?” Micron’s results also showed record DRAM and data-center revenue more broadly, which fits that tighter-supply picture. ### What does this change for AI customers? It shifts the buying logic. If the best memory is allocated a year or more ahead, customers care more about guaranteed delivery, usable throughput per rack, and software efficiency than about headline chip counts alone. A faster model stack or better inference efficiency can matter as much as adding another box if memory-rich systems are scarce. That is an inference from the supply picture, but it follows directly from how constrained HBM has become. ### Is Micron alone here? Not really. Samsung said in February it had started mass production of HBM4 and shipped commercial products, and SK hynix has also been described as sold out well into the AI cycle. So Micron’s statement is not an isolated boast — it fits a broader pattern where all three major HBM vendors are racing to expand capacity against demand from Nvidia-linked AI systems. ### Bottom line? Micron is telling the market that AI memory is no longer the cheap, swappable part of the server. It is the constraint. And when the memory constraint is booked out through 2026, the real scarce product is not just chips — it is delivered AI capacity.