TSMC denies exploring 3D DRAM production at Nvidia’s request - Mezha.net

The fact that Nvidia sought 3D DRAM fabrication from TSMC is more revealing than the denial itself. It exposes a structural shift: the AI compute bottleneck has migrated from logic nodes to memory bandwidth, and the HBM triopoly (Samsung, SK Hynix, Micron) cannot scale fast enough to match GPU iteration cycles. TSMC's denial is fundamentally about preserving the "logic to me, memory to you" industrial division. If a Taiwan, China foundry giant enters DRAM, the pricing power and capex logic of the existing memory oligopoly would be fundamentally restructured. Technically, DRAM's charge-retention physics demands dedicated etch stacks and cleanroom standards TSMC has never built. Forcing entry without two decades of process know-how carries extreme yield risk. More critically, such a move could be read as "memory disguised as foundry," triggering US Commerce Department export-control scrutiny on advanced memory. Compliance costs would likely erase any technical dividend. Over the next 18 months, the real variable is the HBM4 architecture transition. As stack heights move from 12 to 16 layers, TSV density and thermal management become the new binding constraints. Nvidia's more probable path is deepening its SK Hynix HBM4 lock-in rather than building an alternative. The soft boundary between memory and logic is dissolving, but the hard boundaries—process, equipment, talent—will not break in the near term.