Broadcom - through its Avago Technologies International Sales entity - is one of the largest beneficiaries of the AI build-out, supplying custom accelerators and networking silicon that sit beside high-bandwidth memory. A late-2024 grant shows it claiming the memory-stacking IP that complements that business: copper-bonded memory stacks.

US12166027B2, "Copper-bonded memory stacks with copper-bonded interconnection memory systems" (issued 2024-12-10), is classified in H01L 25/18 (memory-device assemblies) with H01L 24/05/08/80 (bonding), H01L 23/367 and 23/481 (thermal and TSV). The claim covers a memory stack using copper-bonded interconnection between layers - direct Cu-Cu bonds replacing microbumps - with attention to the thermal path.

Construe the copper-bonded limitation plus the thermal element. Direct copper bonding gives finer pitch and lower resistance than solder microbumps, and it shortens the thermal path out of the stack - critical when memory sits next to a hot accelerator. The claim's value is the specific arrangement coupling the copper-bonded interconnect to the heat-removal structure (note the H01L 23/367 thermal classification).

The commercialization grounding is what makes this a 'From Claim to Market' case. Broadcom ships custom AI silicon and the networking that connects it, and copper-bonded memory integration is directly relevant to the bandwidth and thermal envelope of those products. The patent defends know-how tied to a real, high-margin product line.

The design-around space is the bonding and thermal scheme. A competitor using microbump-attached HBM, or a different thermal-management structure, reaches a working memory stack outside a claim that requires copper bonding coupled to this specific thermal arrangement. But as the industry moves toward bonded HBM base dies, this kind of claim grows more relevant.

For an IP-risk reader, the takeaway is that the HBM-integration thicket includes the system houses - Broadcom, NVIDIA, AMD - not only the DRAM makers. Broadcom's copper-bonded-stack claims sit alongside Adeia's bonding IP and the memory vendors' stack IP, and any AI-memory integration program navigates all three.