Samsung will use rival TSMC's 3-nanometer process to manufacture the base dies at the heart of Nvidia's next-generation NVHBM memory stacks, according to a report from Seoul Economic Daily, after the GPU maker asked its memory suppliers to standardize production on TSMC's most advanced logic node.
The report, which was picked up by Reuters over the weekend, says Nvidia has asked all three of its HBM suppliers — Samsung, SK hynix and Micron — to adopt TSMC's 3nm process for the base dies of NVHBM, the custom memory package Nvidia unveiled in August. The move effectively unites two of the world's fiercest foundry competitors, Samsung and TSMC, on the same piece of silicon. For more context on this story, see our ongoing artificial intelligence updates.
What NVHBM Changes About AI Memory
NVHBM is Nvidia's attempt to redesign high-bandwidth memory around the needs of AI accelerators. In a conventional HBM stack, the memory controller lives on the GPU itself. NVHBM moves that controller into a custom base die beneath the memory stacks, a shift that treats the base die less like passive wiring and more like a piece of advanced logic.
That design is what gives the technology its headline numbers. According to Nvidia, NVHBM offers up to 30% more memory bandwidth, 15% lower HBM power consumption and 25% more XPU compute die area than standard HBM4E — gains that come precisely because the base die must be built on a leading-edge logic process rather than a memory-optimized one.
Why TSMC, and Not Samsung's Own Fabs
The logic-node requirement is also what pushed Nvidia toward standardization. Seoul Economic Daily reports that Nvidia wants consistent HBM performance across its suppliers and a simpler validation process — goals that are easier to hit if every memory maker builds its base die on the same process, at the same foundry.
Yield is the other factor. Industry estimates cited by the newspaper place TSMC's 3nm yields above 90%, while Samsung Foundry's 2nm process (SF2) sits at around 60%. For a complex base die integrating a memory controller that once lived on the GPU, that gap is decisive: a base die built on a lower-yielding node raises costs precisely where the design is most demanding.
Samsung's Device Solutions division, which oversees the company's foundry operations, reportedly agreed to the requirements after an internal review — a notable concession for a conglomerate that has spent years trying to win high-volume advanced logic business away from TSMC.
Rivals on One Chip
Cooperation between the two foundries is not without precedent. According to Seoul Economic Daily, Samsung and TSMC first worked together about four years ago to ensure HBM compatibility with TSMC's GPU packaging process. The difference now is scope: the new arrangement extends their collaboration into the manufacturing of HBM components themselves, the first time the rival foundries have shared that layer of the stack.
Samsung, for its part, has not abandoned its own fabs. As ZDNet reported in September, the company is pursuing a dual-track strategy for custom HBM, tapping both its internal foundry and TSMC for base die production. The new report indicates that for Nvidia's NVHBM program specifically, the TSMC track is the one that matters.
A Truce Between Competitors — For Now
The arrangement is striking precisely because Samsung and TSMC remain head-to-head competitors everywhere else. The two foundries vie for the same marquee customers — Qualcomm and Tesla among them, as TrendForce notes — and Samsung has staked much of its foundry strategy on closing the technology gap that has left TSMC holding the majority of advanced-node orders.
That NVHBM pulls them onto a single die is a measure of how much leverage Nvidia's accelerator roadmap now exerts over the semiconductor supply chain. Memory makers whose HBM business once revolved around generic, interchangeable stacks are being asked to co-design custom silicon, and to build it at their rival's fab, to stay in Nvidia's good graces.
What It Means for the AI Memory Race
For Nvidia, standardizing the NVHBM base die on TSMC 3nm reduces a coordination problem to a manufacturing one: three memory suppliers with different process roadmaps can now deliver functionally identical base dies, simplifying qualification and smoothing supply for an accelerator roadmap whose appetite for bandwidth keeps growing.
For Samsung, the decision is a double-edged signal. Losing the base die business for a flagship memory program to its chief foundry rival stings — but participating in the program at all keeps Samsung's HBM business inside Nvidia's ecosystem, alongside SK hynix and Micron, rather than watching the entire package go elsewhere.
For TSMC, the deal extends an already dominant position: the foundry now sits not just under Nvidia's GPUs, but inside the memory that feeds them. As AI hardware makers push customization deeper into the memory stack, the base die is becoming the new battleground — and the latest report suggests TSMC starts that fight with a commanding lead.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
