Nvidia pushes glass substrates for HBM packaging as CoWoS-L warpage limits 4-die designs
Nvidia is encouraging TSMC and others to develop glass substrates that would let AI GPUs stack more HBM than silicon interposers allow. Jensen Huang met SK Group chairman Chey Tae-won to discuss cooperation on the technology.
Glass is cited as raising chip speeds about 40% at roughly half the power, but through-glass-via metallization remains a hurdle and customer qualification is incomplete. Reported timelines run from TSMC CoPoS volume production in 2028-2029 to Samsung Electro-Mechanics readiness in the second half of fiscal 2027 and Intel in the latter half of the 2020s. TSMC's CoWoS-L reportedly could not control warpage at 4-die scale, which pushed Rubin Ultra back to a dual-die design.
For AI-datacenter buyers, packaging limits shape when higher-HBM GPUs ship and what they cost.