Samsung Turns to TSMC 3nm Lines for Nvidia HBM Base Die

Samsung has placed production of Nvidia's custom HBM base dies on TSMC's 3nm process lines, according to a Dealroom report. The decision means one foundry giant is buying advanced-node capacity from its chief rival to serve the same customer both firms compete to supply.

The deal concerns the logic base die that sits beneath high-bandwidth memory stacks in advanced AI accelerators. Nvidia packages these base dies together with HBM memory to form the memory subsystem of its top-end chips. The report indicates that the base die destined for Nvidia's custom HBM will be manufactured on TSMC's 3nm-class node.

Why does the base die matter?

The base die is the silicon layer that connects the DRAM stacks to the GPU or accelerator package. As memory bandwidth demands climb, this layer has absorbed more logic function. That shift has pushed base dies toward leading-edge process nodes, the same territory used for flagship processors.

Manufacturing them on a 3nm-class node gives the die faster logic and lower power delivery losses. It also means the supplier of the memory stack and the manufacturer of the base die can now be two different companies. In this case, Samsung supplies the memory integration while TSMC builds the logic layer underneath.

The takeaway: memory leadership no longer guarantees control of every layer in an advanced AI package.

What does the deal say about Samsung's position?

For Samsung, the move carries a clear commercial logic. Nvidia is the largest buyer of HBM, and securing its custom base die business strengthens Samsung's hold on the memory side of the package. Using TSMC's 3nm lines to build that die removes a potential objection from Nvidia if Samsung's own foundry node is not the customer's first choice.

It also signals restraint about Samsung Foundry's advanced-node standing. Rather than press Nvidia to use its own 3nm-class capacity, Samsung chose the rival's lines. That choice reflects where leading-edge logic customers currently place their trust, and it shows Samsung weighing package-level revenue against foundry pride.

The takeaway: Samsung is protecting the memory business first, even when that means funding a competitor's fab utilization.

What does it mean for TSMC?

For TSMC, the order adds a new customer segment to its 3nm demand base. Base dies for HBM packaging are a growing volume category as AI accelerator shipments rise, and every advanced package typically needs one. Winning this work from Samsung itself, rather than only from Nvidia directly, deepens TSMC's role in the AI supply chain regardless of whose memory wins the socket.

The arrangement also locks TSMC into the memory-adjacent part of the AI bill of materials. Even when a rival captures the HBM stack sale, the logic layer can still route to TSMC.

The takeaway: TSMC now earns advanced-node revenue from Nvidia packages even when Samsung supplies the memory.

How firm is the report?

The Dealroom item is a single-outlet report, and it does not state shipment volumes, timing, or financial terms. Details such as which 3nm variant TSMC will use, when production starts, and how the dies pair with specific HBM generations remain unconfirmed. Readers should treat the production arrangement itself as reported rather than officially announced by either company.

Neither Samsung, TSMC, nor Nvidia has publicly confirmed the arrangement in the source material, so official statements may adjust the picture.

The takeaway: the direction is plausible and consistent with packaging trends, but the specifics await confirmation from the companies.

What comes next?

If the arrangement holds, expect more split packages in which memory vendors and logic foundries divide a single AI product. Samsung's willingness to buy rival capacity suggests the base die business will concentrate further at TSMC until Samsung Foundry proves competitive at the leading edge for external logic customers.