Nvidia Tells Samsung, SK Hynix and Micron to Build Its Custom HBM Base Dies on TSMC's 3nm Process

Nvidia has asked its three HBM suppliers — Samsung Electronics, SK Hynix and Micron — to manufacture the logic base die of its custom high-bandwidth memory on TSMC's 3nm process, according to a report by Seoul Economic Daily on Sunday that was relayed by Reuters. The report, attributed to industry sources, says Samsung has completed an internal review and agreed to the request, marking the first time Samsung has worked with its foundry rival at any stage of HBM manufacturing.

The component in question is the base die of NVHBM, the customized HBM product Nvidia unveiled in August. Built on the HBM4E architecture, NVHBM integrates Nvidia's own memory controller directly into the base die, the logic layer that sits beneath the DRAM stack and connects it to the GPU. Nvidia's stated goal is to unify implementations across memory vendors so that one design performs consistently regardless of supplier, and to cut the per-vendor validation work that today slows down deployment.

Under the reported arrangement, Nvidia wants all three suppliers to use TSMC's 3nm process for the base die while standardizing the core DRAM dies on the industry's sixth-generation (1c) node. Samsung's memory division accepted the requirement after weighing it internally, the report said. The company retains its one-stop production model for standard HBM products and for customized HBM built for other customers, AMD among them.

The concession has a capacity logic behind it. Samsung's SF4 (4nm) foundry lines produce roughly 30,000 wafers per month, and half of that — around 15,000 wafers — is already allocated to base dies for Samsung's own HBM4, which the company began mass-producing first in the world this year. HBM4E is in production validation on the same in-house process. Squeezing Nvidia's NVHBM base dies onto those lines was never realistic, which is one reason the company kept a door open: Seoul Economic Daily reports Samsung is still evaluating whether its own fabs could produce NVHBM base dies if TSMC's capacity ever falls short of demand.

The move matters because the base die has quietly become the most contested layer of the AI memory stack. Once a passive routing layer, it now carries buffers, interconnect logic and, increasingly, processing functions that GPU designers want to shape themselves. That has pulled leading-edge logic foundries into a product traditionally owned entirely by DRAM makers, and it is re-drawing the division of labor: the GPU designer specifies the stack, foundries build the logic layer, and memory vendors integrate and supply finished modules.

For Samsung, the decision is double-edged. Its foundry unit has trailed TSMC in leading-edge share for years, and a Samsung memory product being fabricated on TSMC silicon is a visible vote of no confidence in-house. But the memory division sells complete HBM stacks, not wafers, and staying inside Nvidia's platform — the largest HBM socket in the industry — is worth more than the foundry pride. TSMC, for its part, deepens an AI supply-chain position that already spans GPUs, interposers and now the logic layer of its customers' memory.

No production volumes, timing or financial terms were disclosed, and Nvidia, Samsung and TSMC had not commented publicly at the time of reporting. Samsung's standard HBM business, including the HBM4 and HBM4E products now ramping, is unaffected. If the report holds, the next test will be yield: two rival foundries will be qualifying 3nm base dies for the same memory product, and the AI accelerator roadmap for 2027 will depend on both of them landing it.