Sources: Nvidia asked Samsung, SK Hynix, and Micron to make base dies for its custom NVHBM memory on TSMC's 3nm process; Samsung agreed, its first use of TSMC
Samsung Electronics (005930.KS) will use TSMC's (2330.TW) 3-nanometre process to produce base dies for Nvidia's (NVDA.O) …
Context & Ripple Effects
Samsung’s pursuit of Nvidia memory business had progressed from a reported final HBM4 qualification phase in January to initial commercial HBM4 shipments in February. The reported base-die arrangement adds a foundry dependency to that supplier push rather than treating memory qualification and logic manufacturing as separate contests.
The pairing is notable because Samsung had previously positioned itself as an early 3nm producer through its own process technology. Using TSMC for a Nvidia-directed component would make the customer’s memory design requirements more important than the foundry rivalry between the two manufacturers.
First-order effects
- Samsung’s reported agreement would make TSMC the producer of base dies for Samsung’s Nvidia-focused NVHBM, while Samsung remains responsible for competing as a memory supplier.
- Nvidia’s request puts Samsung, SK Hynix and Micron under a shared requirement to support a custom-memory architecture; only Samsung is reported to have accepted TSMC-made base dies.
Second-order effects
- SK Hynix and Micron must weigh whether their NVHBM bids need comparable access to advanced foundry logic, shifting competition beyond DRAM stacking and qualification alone.
- TSMC gains a role in the memory supply chain through a component used by a rival foundry’s customer-facing memory product, tying its 3nm capacity more directly to Nvidia’s memory specifications.
Third-order effects
- If Nvidia’s custom base-die model is adopted broadly, HBM supply becomes more modular: memory vendors differentiate on stacks and packaging while leading-edge foundries provide the logic foundation.
- The arrangement points toward customer-designed memory systems that can redraw traditional boundaries between DRAM suppliers, packaging operations and foundries.
The trend: AI-memory procurement is moving toward workload-specific HBM designs that separate the base logic die from the supplier that builds and stacks the memory.