HBM: Stacked DRAM for AINarrow moat

High-bandwidth memory is the product at the center of Micron's transformation, and understanding what it is explains why it matters so much. HBM is DRAM — the same fundamental technology as ordinary computer memory — but arranged in a radically more capable form: many memory chips are stacked vertically atop one another and connected through the stack by thousands of microscopic vertical wires, then placed right beside the processor. This architecture delivers vastly more memory bandwidth than conventional memory, which is precisely what an AI accelerator needs, because the speed of training and running AI models is gated by how fast data can be fed to the processor.

That makes HBM not a luxury but a necessity for high-end AI: every leading AI accelerator from Nvidia and AMD depends on it, and the amount of HBM per system keeps rising with each generation. Because HBM is far harder to make than ordinary memory — the stacking is delicate, yields are lower, and it consumes several times the wafer capacity per bit — it is genuinely scarce, and that scarcity, combined with its indispensability to the AI buildout, gives it economics unlike anything in commodity memory. Micron is one of only three companies on earth that can make it. Yet HBM is still, at bottom, a form of DRAM: it is differentiated and scarce today because it is hard to make and demand is exploding, but it is not protected by a brand or a proprietary standard, and if the three makers collectively build enough capacity, even HBM's scarcity could ease. For now, though, HBM is the most valuable memory ever made, and Micron makes it — one of only three companies on earth that can¹.