Field report W-85 / P7 · sheet /operations
Operations & RestructuringFiled from the territory
NVHBM's 3 nm Base Die Could Redraw the HBM Supply Chain
NVIDIA's reported plan puts TSMC's 3 nm logic base die under Samsung, SK hynix and Micron, and raises the harder question of who qualifies the finished stack.
- Grid
- W-85 / P7
- Surveyed
- November 17, 2026
- Reading
- 5 min read
- Correspondent
- Priya Raman

Field notes
- NVIDIA asked Samsung, SK hynix and Micron to use TSMC's 3 nm process for NVHBM logic base dies, per Seoul Economic Daily, October 18, 2026.
- The same report says NVIDIA wants the DRAM dies standardized on the sixth-generation 1c process generation.
- Samsung reportedly accepted the requirements despite historically relying on its own foundry operations for HBM.
- SK hynix partnered with TSMC on HBM4 base dies in April 2024; Marvell unveiled a custom HBM architecture in December 2024.
- NVHBM moves the memory controller from the accelerator into the HBM base die, widening the split between logic and DRAM fabrication.
NVIDIA wants Samsung, SK hynix and Micron to build the logic base die for its custom NVHBM memory on TSMC's 3-nanometer process, and to standardize the DRAM dies beneath it on the sixth-generation 1c generation. Seoul Economic Daily reported the request on October 18, 2026, under the headline "Samsung to Tap TSMC for Nvidia's Custom HBM Base Die"; Reuters carried the same Seoul-based reporting the same day.
Samsung has reportedly accepted the requirements, despite a tradition of relying on its own foundry operations for HBM production. No company has publicly confirmed a manufacturing agreement, and the details remain industry reporting rather than a signed arrangement.
The request matters less as a process choice than as a division of labor. It gives logic-fabrication work to TSMC while memory makers keep DRAM and stacking, and it pulls NVIDIA's memory controller — normally hosted on the accelerator — down into the base die.
The framing question is blunt: "If the system designer defines the memory architecture, a foundry manufactures the logic base die, and memory suppliers fabricate and stack the DRAM, could HBM become a more modular manufacturing product?"
What does NVIDIA actually want changed?
Conventional HBM pairs vertically stacked DRAM dies with a logic base die. The DRAM supplies capacity; the base die handles supporting functions and the physical interface to the processor; the memory controller typically sits on the accelerator. NVHBM moves that controller into the base die and redesigns the memory-to-compute interface. NVIDIA's own NVLink Fusion announcement of August 26, 2026 introduced the custom architecture.
That single change turns the base die from a supporting part into a significant architectural block, and opens a seam between logic fabrication and DRAM fabrication. The reported 1c requirement pushes standardization one layer deeper. Instead of standardizing only the base die, NVIDIA appears to want consistency in the DRAM technology generation too. The same generation, though, is not the same process: two suppliers can build different device structures, flows and yields on one node.
Has this happened before?
Yes, in pieces. In April 2024, SK hynix announced a collaboration with TSMC to apply advanced foundry technology to HBM4 base dies. In December 2024, Marvell introduced a custom HBM architecture combining base-die logic, controller functionality, advanced die-to-die interfaces and packaging. NVIDIA's reported move pushes further, chasing one custom architecture across three memory suppliers at once.
Could HBM become a modular product?
HBM has always shipped as a tightly integrated product: DRAM fabrication, through-silicon via (TSV) formation, wafer thinning, stacking, bonding, testing and qualification have to move as one coordinated process. A standardized custom base die would formally split that chain into four roles:
- System designer: defines the controller, memory interface and system-level performance requirements.
- Logic foundry: manufactures the customized base die on an advanced logic process.
- Memory manufacturer: produces the DRAM dies and integrates them into a completed stack.
- Advanced packaging provider: integrates qualified memory stacks with the accelerator and package interconnect.
Where does the engineering difficulty land?
It moves, it does not disappear. A logic die built on one process must connect reliably to DRAM built on another, across differences in electrical characteristics, physical geometry, power and thermal behavior. Stacking yield depends on TSV formation, bonding structures, alignment accuracy and accumulated variation; a qualified base die plus individually tested DRAM does not guarantee a high-yield finished stack.
Moving the controller into the base die also relocates heat. Thermal design must now account for logic activity, heat transfer through the DRAM stack and the operating environment created by the neighbouring accelerator. Test access gets harder too: a package-level failure could sit in the DRAM array, the controller logic, the vertical interconnect, the bond, the power delivery or the surrounding package. That demands intermediate testing, diagnostic visibility and traceability across company lines, and the finished product must still prove bandwidth, latency, power, thermal performance and long-term reliability under representative conditions.
Could this change who builds HBM?
Potentially, though supplying components and delivering a qualified HBM stack remain different businesses. A foundry could fabricate controller-equipped base dies, a memory maker could concentrate on DRAM density and high-density stacking, and a specialist could perform integration and test. But entering that chain still requires competitive DRAM technology, qualified stacking and bonding processes, known-good-die strategies (pre-screening dies before assembly), thermal characterization, failure analysis and qualification capability.
Economics cuts both ways. Splitting the flow adds logistics, intermediate testing, traceability, yield ownership and coordination costs; the model works only if flexibility and specialization pay for them. Incumbents keep real differentiation in cell density, power efficiency, stacking capability, process control and yield.
What would prove the model works?
Not a shared base-die process announcement. The milestone is repeated delivery of complete stacks from multiple suppliers against common system requirements — measured on bandwidth, latency, power, signal integrity, thermal behavior, manufacturing variation and reliability, with test coverage before and after stacking and clear ownership of failure analysis when faults cross organizational boundaries.
Two suppliers on the same NVIDIA-defined architecture can both meet nominal bandwidth and still differ in operating margins once temperature, voltage, workload or ageing shifts. The closing maxim from the analysis is the test to remember: "Standardizing the design does not standardize the product."
That is the argument for NVHBM. Architectural ownership and manufacturing ownership need not sit in the same company. Whether the separation yields a genuinely modular HBM ecosystem, or merely a longer chain of interfaces to manage, will be settled by product-level results, not process roadmaps.
via semiwiki.com (Source)
More from Priya Raman
Show full bio
Senior reporter covering technology and product news in corporate strategy and chief-executive leadership at CEO Frontier.
460 articles
Adjacent sheets
- TSMC Will Ride Out the Jensen Huang Scandal, Analysts Say
- Memory shortage may tighten through mid-2027, delaying price relief
- U.S. AI Chip Proposal Covers Four Southeast Asian Nations
- Nvidia Loses $400 Billion in Two Days as Huang Scandal Meets China Chip Claims
- The $7.5 Billion Chip Route: How Megaspeed Moved Nvidia Silicon to China