Samsung will use TSMC to manufacture the base die for a custom HBM (high-bandwidth memory) package built for Nvidia, Seoul Economic Daily reports. The deal pairs Samsung's memory stacks with silicon from its strongest foundry competitor.

HBM modules stack DRAM (dynamic random-access memory) dies on top of a logic chip — the base die — which routes data between the memory and the GPU. Whoever fabricates that base die controls a critical piece of the AI accelerator supply chain, which is why this arrangement turns heads in Seoul and Hsinchu alike.

What does this change for Samsung?

Plenty. Samsung has spent years selling itself as a one-stop shop: memory, foundry, and packaging under one roof. Nvidia's custom HBM order now splits that story. The memory dies stay Samsung's; the base die goes to TSMC (Taiwan Semiconductor Manufacturing Company), the world's largest contract chipmaker and Samsung's direct rival for advanced logic.

It signals Nvidia wants qualified, proven logic manufacturing for the base die rather than waiting on Samsung's foundry nodes to clear whatever internal bar the GPU maker applies. For Samsung's foundry unit, losing the base die portion of a flagship AI memory product stings — the optics alone are rough.

What does this change for Nvidia?

Nvidia gains supply flexibility and de-risks the logic layer. Custom HBM — memory specified to Nvidia's own requirements rather than a standard catalog part — demands tight coordination between memory maker and foundry. TSMC's involvement gives Nvidia a base die partner with an unmatched track record on advanced process nodes.

It also keeps Samsung meaningfully inside Nvidia's HBM supply chain, alongside SK Hynix, the current market leader in HBM for AI accelerators. Competition among suppliers is exactly what a buyer like Nvidia wants when demand for AI memory outruns capacity.

The competitive picture

The report lands amid a three-way race:

• SK Hynix — leads HBM shipments to Nvidia today.
• Samsung — fighting to qualify its HBM and expand its slice of the AI memory market.
• TSMC — fabricates the logic, and now, per the report, part of Samsung's own HBM product.

A Samsung memory package carrying a TSMC-made base die blurs the neat vendor boundaries the industry likes to draw. Expect procurement teams and analysts to read this as proof that in custom HBM, the memory maker and the foundry no longer need to be the same company.

Why the base die matters

The base die isn't a passive spacer. It carries the Through-Silicon Via (TSV) interconnects and I/O (input/output) circuitry that move data between the stacked DRAM and the processor. Its performance caps the bandwidth the whole module can deliver. For Nvidia's custom specification, that layer is arguably the hardest engineering problem in the package — and the part Nvidia apparently wants on TSMC's lines.

Seoul Economic Daily's report doesn't specify volumes, timelines, or pricing. Watch for confirmation from the companies and for qualification milestones, which will tell us whether this becomes a lasting arrangement or a one-off workaround.

What's certain: Nvidia's custom HBM will combine Samsung memory with TSMC logic. Two rivals, one package, one customer calling the shots.