AI memory's control layer is shifting from Korea to Taiwan

Noise in the market, signal in the base die

The headline of the week was the selling wave. On the Taipei Times' count, foreigners sold $19.4 billion of Taiwanese equities in the week of 12-16 October, and the Philadelphia Semiconductor Index fell 15.7% between 7 and 16 October. That move prices valuation and rate expectations, not the structure of the supply chain. The quiet parameter arrived on 18 October. Seoul Economic Daily reports that Samsung will have the base die of Nvidia's custom NVHBM memory made on TSMC's 3-nanometre process. Nvidia put the same request to SK hynix and Micron, and none of the companies confirmed the report on the record.

The base die is the logic chip that sits beneath the stacked memory layers and runs them. According to The Next Web, NVHBM moves the memory controller down from the graphics processor to that die; the physical layer shrinks by 60% and bandwidth rises by up to 30%. Memory thus stops being mere storage and becomes part of the compute architecture. Whoever makes that piece holds the key to the memory chain.

Why it happened now

The constraint that made the decision possible is yield. Industry sources put TSMC's 3-nanometre yield above 90% and Samsung's 2-nanometre process at around 60%. Tech Monitor notes that the comparison is between different nodes and that the figures are unverified. Nvidia spares itself the burden of qualifying 3 suppliers separately. The timing is set by the supply gap: UBS expects HBM demand of 61.5 billion gigabits in 2027 against new supply of 54.6 billion. With a shortfall of roughly 10%, the buyer picks uniform, fast production over spreading its risk.

On the Korean side the price is paid by the turnkey model. Samsung makes the standard HBM4 base die on its own 4-nanometre line and will keep that model for customers such as AMD. At its largest customer, though, the logic layer goes to a rival. Counterpoint data cited by MoneyToday give SK hynix 50% of HBM revenue in the second quarter of 2026, Samsung 33% and Micron 18%. If all three head for the same foundry, the entire market is tied to TSMC at one link.

Bargaining power changes hands at this point. MoneyToday reports that J.P. Morgan expects SK hynix to supply 54% of Nvidia's HBM purchases this year and has raised its forecast for the fourth-quarter DRAM contract price increase from 3-8% to 10-15%. Prices are turning in the memory makers' favour while Nvidia sets the design and TSMC the critical logic layer. What Korea keeps is core DRAM production; Samsung is testing its own foundry only as a fallback for the moment TSMC's capacity runs short.

The structural constraint is deepening at one address

Focus Taiwan notes that TSMC alone accounts for more than 40% of the Taiwanese market's capitalisation. The Next Web writes that TSMC has added 2 new 3-nanometre lines at Fab 18 in Tainan, due to enter production in 2027 and the first half of 2028. No capacity or date has been given for the 3-nanometre plans in Arizona and Kumamoto. The NVHBM base die will therefore be made mostly in Taiwan until at least 2028. A blockade or quarantine scenario in the strait now severs not only the leading-edge processor but the control layer of AI memory.

The market is already pricing that concentration in the currency. Taiwan's central bank tied the outflow on 19 October to AI valuations and to financing risk in the sector. The NT dollar weakened to 32.57 intraday, while the Korean won and the Chinese yuan strengthened against the dollar in the same week. Capital is leaving a market tied to a single sector while the supply chain is being tied to that market more tightly. The two moves ran in opposite directions in the same week and they do not offset each other.

What would refute this reading

Our thesis is this. Over the next 12 months most of the NVHBM base die will be concentrated on TSMC's 3-nanometre lines in Taiwan, and the geographical concentration of AI memory supply will increase. We put 55% on that path.

First refutation: if SK hynix says at its third-quarter presentation on 29 October that it will make the NVHBM base die on its own or another process, the single-foundry thesis weakens. Second refutation: if Samsung's foundry shows a clear rise from the 60% yield on 2 nanometres and passes Nvidia's qualification, a second source appears. Third refutation: if TSMC gives a production date within 2027 for the 3-nanometre lines in Arizona or Kumamoto, the concentration risk spreads geographically.