Hybrid bonding - the simultaneous fusion of dielectric-to-dielectric and copper-to-copper interfaces between two dies - is the enabling technology behind today's 3D-stacked logic and memory. Taiwan Semiconductor Manufacturing Co., Ltd. has built one of the densest portfolios in this area, and its mid-2020 grant is a useful claim-construction case because the title is broad while the claims are narrow.
US10727205B2, "Hybrid bonding technology for stacking integrated circuits" (issued 2020-07-28), is classified in H01L 25/0657 (assemblies of stacked devices), H01L 23/481 (through-substrate vias), and H01L 24/83 (bonding processes). The claim is not on hybrid bonding as a concept - that would read on a generation of prior art - but on a particular arrangement that ties a through-via on one die to an interconnect and bonding pad that meet across the bond interface.
Construe the limitation set carefully. What is being fenced is the geometric and electrical relationship among the via, the topmost interconnect layer, and the bonding pad that mates with the opposing die. A competing 3D stack that routes its vertical connection differently - landing on a redistribution layer rather than a direct pad, for instance - may sit outside the literal claim even while using hybrid bonding.
This matters because hybrid bonding is now a shared industry technique with contributions from Adeia (formerly Invensas/Xperi), Intel, Samsung, and the memory makers. No single assignee owns 'hybrid bonding.' Ownership lives in the specific surface preparation, the pad metallurgy, the via topology, and the alignment scheme. Each is a separately patentable sliver.
For freedom-to-operate work, the practical takeaway is that a hybrid-bonding product is rarely blocked by one patent. It navigates a thicket of narrow claims, and the design-around question is always 'which specific interconnect arrangement does this grant actually require, and can we route around it?'
TSMC's accumulation here is strategic: it pairs the bonding IP with its CoWoS and SoIC packaging programs, so the claims defend a commercialized stack rather than sitting idle. That commercialization is what turns a narrow claim into a meaningful competitive moat.