Samsung collaborates with TSMC for Nvidia’s custom HBM chips - www.verdict.co.uk

The Samsung-TSMC HBM co-development is not a partnership headline—it is the structural collapse of the memory/foundry boundary that has defined the industry since the 1990s. HBM was a commodity: three DRAM houses built stacks, TSMC packaged them. Now Samsung is embedded inside Nvidia's architecture loop, converting HBM into system-level custom IP. For SK Hynix, the 'default Nvidia supplier' moat is eroding structurally; the HBM3E differentiation window may close within two quarters. TSMC is migrating its value anchor upstream from pure foundry to memory-logic co-design, and CoWoS allocation leverage shifts decisively toward Nvidia. Within 12-24 months, HBM4 16-hi stacks become the new battleground, and 'who controls custom HBM architecture' replaces 'who has the most capacity' as the defining variable in the AI supply chain. The deeper signal: advanced packaging is no longer a back-end process—it is the new front-end of compute.