Samsung will fabricate the base die for NVIDIA's NVHBM memory on TSMC's 3nm process and will not use its own 4nm node for that component, according to a report published by XenoSpectrum. The decision puts a foundry contract for NVIDIA silicon on a rival's line rather than Samsung's.

By Amara Osei

What does the report say?

The core claim is narrow and specific. The base die that sits inside NVIDIA's NVHBM memory stack will be manufactured at TSMC on a 3nm-class process. Samsung, which is building the NVHBM package itself, will not produce that die on its own 4nm technology, the report states.

That single sentence carries three checkable facts:

• The component in question is the NVHBM base die destined for NVIDIA.
• The manufacturing process is TSMC's 3nm node.
• Samsung's 4nm process is out of the picture for this part.

Why does the base die matter?

The base die is the logic layer at the bottom of a high-bandwidth memory stack. It manages the data paths between the DRAM layers above it and the host chip below. As memory interfaces have widened, that logic layer has moved onto advanced logic nodes, which is why a memory product now depends on a leading-edge foundry process at all.

In this case, the report places that logic layer at 3nm — a node class that TSMC has shipped in volume since 2022 — rather than at Samsung's 4nm, which the company has run for its own Exynos processors and for select foundry customers.

What does it mean for Samsung Foundry?

The reported split is the notable part. Samsung keeps the NVHBM work — the packaging and memory-side integration — while sending the most process-sensitive piece, the base-die logic, to TSMC.

For Samsung Foundry, that means one fewer advanced-node socket on a high-profile AI-related product line. The company's 4nm output has served internal and external customers since 2021, but the report indicates it did not win the NVHBM base-die slot against TSMC's 3nm.

For TSMC, it adds an NVIDIA-linked component to a 3nm demand base that already spans mobile and high-performance computing customers.

What remains unconfirmed?

One report, one claim. XenoSpectrum cites no named Samsung, NVIDIA or TSMC executive, and none of the three companies has published a statement confirming the allocation. Until Samsung, NVIDIA or TSMC comment on the record, the 3nm-for-4nm swap stands as a single-sourced report.

The financial terms are also absent: no wafer volumes, no pricing, no timeline for production start. The report establishes which process wins the socket, not how large that socket is in revenue terms.

How does this fit the wider picture?

Foundry allocation decisions for base dies have become a visible proxy for competitive standing in AI memory. Each generation of high-bandwidth memory has pushed the base die toward denser logic, and the choice of node — and of foundry — signals where suppliers place their confidence for yield and capacity.

The reported outcome gives TSMC the logic layer and Samsung the package. In one line: Samsung will build NVIDIA's NVHBM base die on TSMC 3nm, not on its own 4nm, per XenoSpectrum.