Google's expanding AI chip business opens new opportunities for Samsung, SK hynix beyond Nvidia - ET Enterprise AI

Google's TPU expansion signals a structural re-architecture: AI compute demand is migrating from an Nvidia-centric monopoly to a multi-pole custom ASIC landscape. HBM is no longer a GPU-ecosystem afterthought—Google, Meta, and Amazon's in-house accelerators are opening a parallel demand channel. This is market restructuring, not incremental growth. Technically, TPU v5p/v6 demands HBM3E at 12-16 layer stacks with extreme bandwidth, directly pulling utilization on Samsung's 1c DRAM node and SK hynix's 12-Hi lines. More critically, logic die foundry work may shift from TSMC to Samsung's 4nm/3nm, eroding single-fab bargaining power. Compliance-wise, US export controls on CoWoS packaging and HBM shipments to China are creating a bifurcated market. Korean memory makers must maintain dual production tracks, inflating capex while their China hyperscaler revenue faces further compression. Strategically, Nvidia will likely lock SK hynix HBM capacity via multi-year agreements. AMD's MI325X and Intel's Gaudi 3 will compete for the same HBM pool. Memory pricing floors will rise through 2025-2026 as demand diversifies. Within 24 months, HBM4 (16-Hi) becomes the new battleground. If Samsung achieves yield leadership, SK hynix's HBM monopoly fractures. Custom ASICs will exceed 30% of AI accelerator shipments by 2026, turning HBM from an "Nvidia-exclusive" into a multi-architecture commodity.