Most high-bandwidth memory IP comes from the giants - Samsung, SK hynix, Micron, Intel. But the field also attracts well-funded startups betting on more aggressive 3D integration than HBM offers. Kepler Computing, Inc., a ferroelectric-and-stacking startup, filed heavily in this area, and its late-2021 grant claims a 3D-integrated ultra-high-bandwidth memory.
US11152343B2, "3D integrated ultra high-bandwidth multi-stacked memory" (issued 2021-10-19), is classified in H01L 25/18 (memory-device assemblies) with H01L 23/3128/3672 (encapsulation/thermal), H01L 23/481 (TSV), H01L 24/17, and H01L 25/50. The claim covers a multi-stacked 3D memory arrangement engineered for ultra-high bandwidth.
Construe the multi-stack limitation. HBM stacks a handful of DRAM dies; the claim's value is in a specific multi-stacking arrangement - how the memory layers, the interconnect (TSVs at H01L 23/481), and the thermal handling (H01L 23/3672) combine to push bandwidth past the conventional stack. The independent claim defines which arrangement is fenced.
The startup-IP-strategy angle is the interesting part. A young company with no fab builds value through patents - establishing an early position that makes it either an acquisition target or a licensing player. Kepler's dense 2021 filing across stacked-memory and chiplet claims is exactly this playbook: stake the territory before the incumbents fully occupy it.
The design-around space is broad in a crowded field. Conventional HBM stacking, or a different 3D-integration topology, sits outside a claim tied to this specific multi-stack arrangement. But the claim adds to Kepler's portfolio depth, which is the point for a startup building negotiating leverage.
For a portfolio analyst, Kepler's appearance among the HBM-and-stacking assignees is a reminder that the bandwidth-memory thicket includes ambitious newcomers, not just the memory establishment, and their early filings can become surprisingly load-bearing if their approach catches on.