HBM

HBM, or high-bandwidth memory, is DRAM stacked directly beside an AI accelerator so the chip is fed data fast enough to stay busy rather than sitting idle waiting for inputs. It matters because raw compute is useless without the bandwidth to supply it, which has quietly made memory — not logic — the gating component for how many accelerators actually ship. Capacity has been sold out across multiple generations, so HBM output, more than wafer supply, sets the real ceiling on the industry.

If HBM stays scarce or yields disappoint, GPU shipments are capped no matter how much foundry capacity exists — the bottleneck lives below the obvious part. Because only a very small club can make it, that concentration is both a source of pricing power and a source of fragility. SK hynix anchors supply, NVIDIA depends on it, and the tool and inspection makers that enable the stacking — Advantest, Disco, Onto Innovation, Tokyo Electron — are pulled in alongside.