Bandwidth is the headline; capacity is the wall. High-bandwidth memory gets the attention because bytes-per-second is what AI training screams for, but every HBM stack is built from ordinary DRAM dies, and a DRAM die's capacity comes from one thing: how small you can make the memory cell. That cell — one transistor and one capacitor (1T1C), storing a single bit as charge — is where the real memory wall lives.
Samsung's grant US12652792B2, "Semiconductor device" (issued June 9, 2026; Samsung Electronics Co., Ltd.; CPC H10B12 DRAM device classes), is cell-level structure work. The reason a generic-sounding title hides hard engineering is that at these dimensions the capacitor has to hold enough charge to be reliably readable while occupying almost no footprint — a geometry problem that gets worse every node. The claim is part of how you keep the cell readable as it shrinks.
The companion grant, US12652789B2, "Semiconductor devices having landing pad structures" (issued June 9, 2026; Samsung; CPC H10B12/0335 and H10B12/315), is about the connections. A "landing pad" is the contact structure that lets you wire the buried cell up to the rest of the circuit without shorting to its neighbors. At advanced DRAM pitches, the contacts are as hard as the devices — misalign a landing pad by a hair and you kill the cell or bridge two of them.
Why does this gate the AI story? Because HBM capacity per stack is DRAM density times the number of dies you stack. The stacking (covered elsewhere on this site) gives you the wide, fast bus; the cell scaling gives you the gigabytes. A 16-high stack of denser dies is how you push both bandwidth and capacity at once — and the capacity half depends entirely on grants like these, made at the single-cell level.
The deflationary point a memory specialist has to make: DRAM cell scaling does not follow the logic node cadence, and it has been getting harder for years. There is no easy gate-all-around equivalent that resets the curve; you are fighting capacitor physics in shrinking volumes. That is why the patent flow concentrates on incremental structural tricks — better capacitors, better contacts, better isolation — rather than a single clean reinvention.
So read an HBM capacity number as two achievements bolted together: a stacking achievement and a cell-scaling achievement. Samsung's June 2026 grants are the cell-scaling half, the unglamorous device-level work that decides how many gigabytes ride in each stack. Bandwidth is the new node story; capacity is the old, hard one that never went away.