Nvidia's Custom HBM Demand Pulls Samsung Toward TSMC's 3nm

By Marcus Bennett

A report from Digitimes claims Nvidia's push for custom high-bandwidth memory (HBM) is pulling Samsung onto TSMC's 3-nanometer process. If accurate, the move ties Samsung's memory business to a production line run by its chief foundry rival.

The claim rests on one mechanism. Nvidia designs its AI accelerators around HBM stacks that sit on the same package as the GPU. Custom HBM variants tighten the tolerance between the logic die and the memory. That pushes the supporting logic, and the integration work, toward whichever foundry already builds the GPU. For Nvidia, that foundry is TSMC.

What does this change for Samsung?

Samsung operates two large businesses in parallel. One makes memory. One makes logic chips as a foundry. The report suggests Nvidia's custom HBM requirements drag Samsung's memory operation into TSMC's 3nm orbit, even as Samsung Foundry competes for the same customers.

The practical consequence is a split role. Samsung would supply memory to a customer whose flagship logic goes to a competitor's fab. That is not unusual in semiconductors. It is unusual when the memory must be co-engineered with the logic at a leading-edge node.

Why the 3nm link matters

TSMC's 3nm family, including N3P and its derivatives, carries the bulk of high-end AI silicon. Nvidia's data center GPUs ride on these nodes. Custom HBM that interfaces directly with 3nm logic needs process-specific tuning.

For Samsung, working on TSMC's 3nm platform means:

• Engineering memory integration around a rival's process design kit.
• Aligning Samsung HBM production schedules with TSMC wafer output.
• Deepening dependence on Nvidia as a customer across two business units.

The competitive backdrop

Nvidia buys HBM at scale for its AI accelerators. SK Hynix has led that supply so far. Samsung has chased the same orders. Custom HBM gives Samsung a route into Nvidia's supply chain, but on terms set by Nvidia's foundry choice.

The report does not state order volumes, timelines, or financial terms. Digitimes attributes the account to supply-chain sources, as is its practice. Treat the specifics accordingly.

What to watch next

What does this change for buyers of AI hardware? Nothing immediately. But it signals tighter coupling between memory makers and foundries. If Samsung memory engineers start building against TSMC's 3nm platform, the line between memory supplier and foundry partner thins.

Watch for three markers:

• Confirmation from Samsung, TSMC, or Nvidia on custom HBM programs.
• Volume shipments of Samsung HBM tied to Nvidia product cycles.
• Any shift in how Samsung Foundry positions its own advanced packaging.

For now, the Digitimes report stands as a single sourced claim. It fits a clear industry logic. It still needs confirmation.