Big Tech Moves Away from Nvidia: Chance for Samsung, SK hynix? - Businesskorea

Big Tech's decoupling from Nvidia is an architecture sovereignty play, not a cost exercise. With TPU, Trainium, and MTIA all entering volume production simultaneously, CUDA's moat is being eroded by the 80% inference workload that simply doesn't require it. This is structural, not cyclical. For Samsung and SK hynix, the opportunity sits at the HBM interface layer, not the GPU design layer. Every custom accelerator still demands HBM3E/4 stacking; the bottleneck has shifted from compute die to memory bandwidth. Samsung's 4nm foundry plus HBM packaging creates a vertical integration that effectively makes it the "TSMC of custom AI silicon" — a full-stack capability Nvidia never possessed. Risk: SK hynix's HBM customer concentration toward a single buyer is a latent bomb. Yet counterintuitively, custom ASIC volume expands the HBM demand pool from one customer to five or more. 12-24 month outlook: The HBM4 transition window (2H25-1H26) will reset the competitive landscape. Once Samsung closes its yield gap, the "Nvidia tax" (30-40% TCO premium) accelerates toward zero, and inference enters multi-vendor equilibrium. The real winner isn't the chip designer — it's whoever controls the memory-compute interface.