The Stack Escaped Memory — The Week Samsung Took HBM's Geometry to CFET, Logic and Mobile

The one-line

This was the week Samsung Electronics escaped the memory-only narrative. At VLSI 2026 Samsung unveiled a 42nm gate-pitch CFET — 6nm tighter than the 48nm prior art from TSMC and imec at the same conference (TheElec). In the same week it disclosed a 3D logic stacking technology applying HBM-style vertical integration to non-memory chips, and DS head Jun Young-hyun publicly named mobile HBM packaging as the next front for on-device AI. SK hynix, meanwhile, shipped HBM4E 12-Hi samples ahead of guidance on 6th-gen 1c (10nm-class) 32Gb DRAM (TheElec).

The market treats these as separate stories. But the fact that they landed in the same week that KOSPI broke 9,000 for the first time and Korea was the only major equity market with net inflow from global funds tells a different story — the denominator on Korean semis is being re-cut. The word stacking, until now an HBM word, is bleeding sideways into logic, foundry and mobile simultaneously.

When stacking leaves memory

HBM is fundamentally a geometry: DRAM dies stacked vertically and connected through TSVs. This geometry just exited memory in three directions in a single week.

First, into logic. Samsung disclosed next-gen 3D stacking that vertically integrates non-memory logic dies using HBM-like methodology. Stacking logic dies lets NVIDIA's GB series escape the lateral area ceiling of CoWoS-L and lets AMD's MI series step up chiplet counts again. For Samsung this is the start of a unified packaging strategy that connects memory stacking (HBM) and logic stacking on the same line.

Second, into foundry. Samsung's VLSI 2026 CFET paper reported a 42nm gate pitch — tighter by 6nm than the 48nm prior art TSMC and imec showed at the same conference. CFET vertically stacks NMOS over PMOS — transistor-level stacking. The same week Intel disclosed trial production of 18A-P with a 9% performance gain (TheElec) — but on the gate-pitch number alone, Samsung walked into the conference with the tightest pitch on the floor.

Third, into mobile. Jun Young-hyun specifically calling out mobile HBM packaging as the on-device AI battle line is not a statement about LPDDR — it is a statement about porting HBM stacking know-how onto mobile AP packages. Note the timing: in the same week NVIDIA unveiled a PC product with 128GB onboard memory, the per-device memory content step-function is reaching down to mobile.