The HBM Memory Supercycle

For decades, memory chips were a brutal commodity business: cyclical, margin-crushing, mostly ignored. Then AI accelerators turned high-bandwidth memory into the hottest product in tech. Nvidia's Blackwell B200 packed roughly 192 gigabytes of HBM per card, more than double its predecessor, and multiplied across hundreds of thousands of GPUs, demand became staggering. SK Hynix, the South Korean maker that moved early and hard, held over half the market by early 2026 and had sold out its capacity for the year, while Micron and Samsung scrambled to catch up.

For founders and operators, this case sharpens how you spot a sleepy, undifferentiated input that a single shift in demand can suddenly make strategic. It trains you to look at the parts of your own supply chain everyone treats as interchangeable and ask who is positioned to capture the value when one of them becomes the bottleneck. Why SK Hynix's early bet paid the way it did is the part to uncover.