The moat fact first: Broadcom's AI story is custom silicon — chips co-designed with hyperscalers — and the value of custom silicon lies partly in integration IP the customer cannot easily replicate. US12166027B2, granted December 2024 to Avago/Broadcom, patents copper-bonded memory stacks and interconnection systems (CPC H01L 25/18).
Gloss it once. Copper bonding joins dies by directly fusing copper pads, allowing far more and finer connections than solder bumps. Applied to memory stacks, it widens the bandwidth between memory and logic. For a custom-silicon vendor, owning a defensible way to bond memory stacks is part of the package it sells to a hyperscaler.
Why a moat read cares: custom silicon competes on the system, not just the core. The integration techniques — how memory is bonded and connected — are part of what makes a custom design better than a merchant part for a given workload. Patents here protect the differentiator Broadcom monetizes with its largest customers.
The period framing matters. By late 2024, custom AI silicon was a major growth driver and memory bandwidth was the binding constraint. A copper-bonded memory-stack grant at that moment reflects Broadcom holding IP in exactly the layer that gates custom-accelerator performance.
The caveat we attach: this is an integration patent and a defensive asset. It evidences Broadcom's position in copper-bonded memory integration; it does not quantify bandwidth or confirm which products use it.
For the period investor, the durable point is that the custom-silicon edge is built in integration. A 2024 copper-bonded memory grant is a piece of the moat behind Broadcom's AI-infrastructure business.