MediumV Technology Geopolitics & AI19 October 2026, Monday
Nvidia funnels its custom memory base dies into one foundry
On a Seoul Economic Daily report of 18 October, Nvidia wants all three of its suppliers to make the base die of its custom memory on TSMC's 3 nanometre node. Samsung appears to have agreed by 19 October after an internal review.
NVHBM is the custom stacked memory Nvidia unveiled in August 2026, and the base die is the logic layer beneath the stack. Nvidia reports up to 30% more bandwidth, roughly 15% less power draw and a 60% thinner physical layer than standard HBM4E. The suppliers are Samsung, SK hynix and Micron, and all three will make the base die at TSMC rather than on their own lines.
The number behind the decision is yield. On industry estimates TSMC's 3 nanometre node runs above 90% yield, while Samsung's 2 nanometre SF2 line sits near 60%. Samsung continues to make the base die for standard HBM4 and HBM4E on its own 4 nanometre line and treats 2 nanometre as a fallback should TSMC capacity tighten. At TSMC's 15 October results presentation, 3 nanometre made up 33% of third-quarter wafer revenue.
The structural constraint appears here: once all three memory makers hang on a single node at a single foundry, the advantage of redundancy disappears. UBS estimates demand next year at roughly 61.5 billion gigabits, a rise of 91%, while new shipments stay at 54.6 billion gigabits, leaving a gap of around 10%. TSMC's new 3 nanometre fab in Tainan enters volume production in the first half of 2028.
Talay assessment
Bottom line
The bottleneck in AI hardware is shifting from the memory cell to base die fabrication. By gathering three suppliers on one node for yields above 90%, Nvidia lifts output in the near term and books a single point of failure into the supply chain in return. The most likely direction is bargaining over TSMC 3 nanometre allocation becoming decisive through 2027.
Likely effects
- Memory makers' bargaining powerNegative1–6 months
With the base die moving to TSMC, the cost advantage Samsung and SK hynix draw from vertical integration narrows and pricing power shifts towards Nvidia.
- Supply chain concentrationNegative6 months+
Tying three suppliers to one node creates the risk that a single capacity or production outage in Taiwan halts shipments from all three brands at once.
- Data centre build schedulesNegativeWeeks
With a memory gap of around 10% expected next year, accelerator delivery dates could slip and investors may pull back installation plans.
Possibilities, ranked
- 1TSMC settles in as sole source55%
All three suppliers make the NVHBM base die at 3 nanometre and Samsung's 2 nanometre fallback is never called on.
Watch: Samsung making no NVHBM base die qualification announcement on the SF2 line by the first quarter of 2027
- 2Capacity tightens, fallback opens30%
The 3 nanometre allocation proves insufficient and Samsung brings its 2 nanometre line in as a second source.
Watch: Samsung Foundry announcing SF2 yields well above 60%
- 3The plan bends to supplier resistance15%
The memory makers bargain to keep the base die in-house, and the decision stays confined to standard HBM4E.
Watch: SK hynix or Micron announcing that it is keeping the base die on its own line
Probabilities are calibrated judgement based on the sources, not measurement, and are revised as new information arrives. Not investment advice.
Market reaction
Indicators affected
- TSMC 3nm yield▲ above 90%
- Samsung SF2 yield▼ around 60%
- Projected memory gap▼ 10%
- TSMC 3nm wafer share▲ 33%