NVIDIA's custom NVHBM specification is pushing the High Bandwidth Memory base die toward TSMC's foundry process, with Micron and Samsung working to bring the world's largest contract chipmaker into the logic layer beneath their DRAM stacks, according to TrendForce.

The TrendForce headline frames the change as a structural break with the conventional HBM supply chain, in which the three memory suppliers — Samsung, Micron and SK Hynix — have historically fabricated the base die on internal logic processes tuned to their DRAM fabs.

What the base die does

The base die in an HBM package is the silicon logic layer that sits beneath the stacked DRAM dies. It carries the physical interface to the host processor, the test circuitry and the signal-routing blocks that connect each DRAM die to the package substrate.

In current HBM3 and HBM3E parts, the memory suppliers build that base die themselves. TrendForce's reporting suggests NVIDIA's NVHBM moves that work to TSMC, pulling the base-die fabrication step out of the memory makers' internal process flows.

Why TSMC is entering the picture

TrendForce does not name a process node in the headline, but the underlying logic is straightforward: TSMC operates the most advanced logic process generations commercially available, while memory companies run logic processes that lag the leading foundry by several years. As AI accelerators push for higher per-pin bandwidth, lower interface latency and tighter integration with the host GPU, the base die has to behave more like a leading-edge logic chip than a memory-side companion.

Pulling TSMC into the base-die fabrication step also aligns the memory stack with the same foundry where NVIDIA's GPUs and packaging flows (CoWoS and FOPLP variants) already run, simplifying process control, qualification and binning.

What Micron and Samsung each handle

Both Micron and Samsung remain the DRAM-stack suppliers for NVHBM, according to TrendForce. The base die becomes a TSMC-fabricated layer designed to NVIDIA's specification, with the two memory companies integrating it under their own packaging flows.

The headline does not name SK Hynix among the suppliers moving the base die to TSMC. SK Hynix retains its conventional base-die approach for the rest of the HBM market, as far as the TrendForce framing indicates.

What does the supply chain look like now?

Bringing TSMC into the base die adds a second foundry dependency to the NVHBM supply chain, on top of the CoWoS and other advanced packaging capacity that NVIDIA already consumes at TSMC for its GPU and accelerator dies. For Samsung and Micron, the shift trades internal logic-die revenue for closer alignment with NVIDIA's roadmap and continued access to the highest-end AI memory pool.

What to watch next

Whether NVHBM sets a template that the rest of the HBM market follows, or stays a NVIDIA-specific customization, will hinge on how quickly competing accelerator designers request equivalent base-die performance. The first public NVHBM-equipped product and the supply-chain footprint that ships with it will determine how broadly TSMC's role in HBM logic extends.