NVIDIA has asked Samsung, SK hynix and Micron to build the base dies of their next HBM stacks on TSMC's 3nm process, according to a TrendForce report. The request would place all three memory makers on the same foundry platform for the dies, which carry the logic layer beneath the stacked DRAM in high-bandwidth memory.

The report centers on a package NVIDIA calls NVHBM. The name covers HBM built to NVIDIA's own specification rather than to the generic JEDEC standard. NVIDIA wants the base dies for this product fabricated at TSMC on its 3nm-class node, TrendForce says.

The move is unusual. Samsung and SK hynix compete directly with TSMC for logic wafer business in some segments, and each memory maker has its own logic options. Samsung in particular runs its own advanced foundry and could in principle produce base dies internally, as it has done for earlier HBM generations.

Why does one foundry for three rivals matter?

A single TSMC base-die platform would give NVIDIA a uniform interface across three HBM suppliers. Today, each vendor pairs its DRAM with its own logic die, which forces NVIDIA to qualify and route different packages for the same GPU slot.

Common base dies would change that arithmetic. NVIDIA could treat Samsung, SK hynix and Micron output as closer to interchangeable supply for its AI accelerators, easing allocation of HBM, a component that has constrained GPU shipments through the current AI buildout.

For the memory makers, the tradeoff is clear. They would gain access to a large, committed NVIDIA order while ceding the logic portion of the HBM stack to TSMC. Samsung faces the sharpest choice, because a TSMC-sourced base die would sideline its own foundry for that business.

What is confirmed and what is not?

The word "reportedly" carries weight here. TrendForce attributes the plan to supply-chain information, not to any of the four companies. No party has confirmed the arrangement publicly, and the report does not state committed volumes, pricing or a production start date.

What is established separately is the technical direction. Base dies for advanced HBM are moving from passive interposer duty toward logic-heavy designs, and the leading-edge nodes needed for that logic sit with TSMC and Samsung. NVIDIA's request, if accurate, would push the entire merchant HBM base for its products onto the most mature of those two platforms.

How does this reshape the supply chain?

The arrangement would concentrate more of the AI memory pipeline at TSMC. The foundry would already produce the GPUs themselves; adding the logic layer of every major HBM supplier's output would extend that grip to the memory interface.

It would also compress Samsung's foundry opportunity in HBM while giving its memory division a clearer path into NVIDIA sockets. SK hynix, the current leader in HBM supply to NVIDIA, would keep its DRAM position but shift part of the stack to a competitor's wafers. Micron, the newest entrant, would gain a level logic playing field defined by NVIDIA and TSMC rather than by its own logic design resources.

Pricing leverage is the quiet variable. NVIDIA already sets demanding cost targets for HBM suppliers, and a shared base-die design would give it one negotiation over the logic layer instead of three, strengthening its position as HBM supply catches up with demand.

What comes next?

Watch for qualification announcements from the three memory makers and for any TSMC disclosure of logic-wafer orders tied to HBM production. TrendForce's account suggests the plan targets the coming NVHBM generation rather than current shipping product, so volume signals would mark the first hard confirmation of the shift.