NVIDIA is developing custom HBM memory, and the project has produced an unusual result: TSMC and Samsung are working on the same piece of silicon. Both foundries are engaged to build the "base die" — the logic layer that sits beneath the DRAM stack in an HBM module — on their respective 3nm-class process nodes.

The arrangement breaks a pattern. TSMC and Samsung have competed directly for leading-edge logic orders for a decade, and base die work for HBM4 has become a new battleground between them. NVIDIA's decision to develop custom HBM shifts that contest into a parallel effort rather than a winner-take-all award.

Why does custom HBM need a foundry at all?

HBM is no longer a pure memory product. The base die carries logic functions — buffer, routing and, increasingly, processing features pushed down from the GPU package. As AI accelerators demand more bandwidth per watt, GPU designers want to shape that logic layer themselves.

NVIDIA's move into custom HBM follows that logic. A purpose-built stack lets the company tune the memory interface to its own compute architectures instead of accepting standardized modules from the three HBM makers — SK hynix, Samsung and Micron.

The base die is where the foundries enter. It is fabricated with logic processes, not DRAM processes, which places the work squarely in TSMC's and Samsung's core business. A report by Wccftech states that both manufacturers are now developing base die on the 3nm node for the effort.

What changes for the TSMC–Samsung rivalry?

The rivalry does not disappear. Each foundry still must deliver its own 3nm base die silicon and compete on yield, density and delivery. What changes is the frame: NVIDIA's custom HBM program gives both foundries a role in the same end product, reducing the risk that one side is shut out of a major AI supply chain socket.

For Samsung, the stakes are higher. The company has trailed TSMC in leading-edge foundry share and has sought high-profile AI-related orders to validate its 3nm gate-all-around processes. Base die work tied to NVIDIA's custom HBM would put Samsung's advanced nodes inside a flagship AI platform even if its foundry unit loses the main GPU wafer business to TSMC.

For TSMC, participation in base die work extends its reach deeper into the memory stack, a segment traditionally owned by the DRAM makers. TSMC already positions itself as the platform partner for AI compute; co-developing the logic layer of custom HBM strengthens that claim.

What it means for the memory makers

SK hynix, Samsung's memory division and Micron remain the HBM manufacturers. Custom HBM does not bypass them. It changes the division of labor: the GPU designer specifies and co-develops the stack, the foundries build the logic base die, and the memory makers integrate and supply the finished modules.

That division matters commercially. If custom HBM becomes standard practice for large AI customers, memory vendors could see their differentiation narrow toward stacking and DRAM process quality, while the value of the interface and logic layer migrates to the GPU designer and its foundry partners.

What comes next

No production volumes, timing or customer commitments have been disclosed for the base die work. The development stage is 3nm-class on both sides, per the report, and NVIDIA has not commented publicly on release plans for custom HBM.

The outcome will hinge on yield and integration. If both foundries deliver working 3nm base die for NVIDIA's stacks, the AI memory supply chain gains a second qualified source for its most critical layer — and the TSMC–Samsung rivalry will be fought inside one product line rather than across two camps.