Samsung Weighs TSMC for Nvidia HBM4E Base Dies

Samsung is considering having TSMC manufacture the custom base dies for HBM4E memory that Nvidia has requested, according to a report by SamMobile. The detail is small. The implication is not.

Base dies sit at the bottom of every high-bandwidth memory stack and route signals between the DRAM layers above and the customer's accelerator below. With HBM4 and its successor HBM4E, those dies have absorbed more logic. That shift turns a passive interposer into a piece of silicon that demands fine-pitch foundry work, and fine-pitch foundry work is TSMC's home turf.

For Samsung, the calculus is uncomfortable but straightforward. Nvidia wants its custom base dies built on a logic process that Samsung's own foundry division has not convincingly demonstrated at the leading edge for this class of customer. If the memory unit insists on internal fabrication and loses the Nvidia socket, the foundry wins an argument and the company loses a customer. If the memory unit outsources to TSMC, the foundry swallows a public rebuke and the HBM business keeps its largest buyer.

Why would Samsung hand work to its chief rival?

Because the buyer holds the leverage. Samsung and SK hynix are racing for Nvidia's high-bandwidth memory orders, and SK hynix already works with TSMC for advanced logic die production on its HBM roadmap. A Samsung refusal to use the same foundry would hand its Korean competitor a clean talking point: same memory quality, fewer process risks.

The arrangement, if it proceeds, would also blur the neat corporate lines the industry likes to draw. Samsung would supply the stacked DRAM; TSMC would fabricate the logic die underneath; Nvidia would assemble the pair into its accelerator packages. One memory product, two rival foundries, three balance sheets.

What is confirmed and what is not?

Only the consideration itself is reported. SamMobile's account describes a possibility, not a signed wafer agreement. Nothing in the report fixes a volume, a process node, a timeline, or a price. Treat every specific figure currently circulating as analyst speculation rather than company guidance.

What the report does establish is direction. The base die discussion covers HBM4E, the generation after HBM4, which places any resulting wafers comfortably into the second half of the decade. This is roadmap positioning, not current-quarter revenue.

Who gains if the deal happens?

TSMC gains a second HBM-adjacent customer for logic die production and deepens its grip on the advanced packaging bottleneck that now shapes AI accelerator supply. Nvidia gains a second qualified source of stacked memory with the process pedigree it demands. Samsung gains, reluctantly, a better shot at the HBM4E socket.

The loser, on paper, is Samsung Foundry, which would watch a marquee advanced-packaging opportunity walk across town to Hsinchu. That division has bet heavily on winning logic customers for precisely this kind of work.

Samsung has not commented publicly, and no fab allocation has been announced. Whether the company formalizes the arrangement should become clearer as HBM4E qualification windows open; whichever way Seoul decides, the decision will signal whether Samsung values foundry pride or memory market share.