TSMC Says No to Memory. Intel Has Two Bets.

TSMC says it will stay in logic. Intel is already working on two ways to replace HBM: Z-Angle Memory with SoftBank, and a patented memory design called XBM.

By

·

Comic illustration of a tower of stacked memory chips rising from a chip factory
Comic illustration of a tower of stacked memory chips rising from a chip factory

Last week a report said TSMC was looking at making 3D DRAM, reportedly at Nvidia’s request. On November 7 TSMC answered. A spokesperson told Bloomberg that TSMC “always focuses on its core business of semiconductor foundry” and wants to remain a trusted provider to the logic industry.

So the world’s largest foundry is staying out of memory. Intel is not.

Why memory is the bottleneck

AI inference is limited by memory, not compute. Every token means streaming model weights from memory, and HBM is both rationed by three suppliers and running into physical limits: only so many stacks fit around a GPU, and only so many wires fit along its edge. Whoever fixes that gets paid.

Bet one: Z-Angle Memory

Intel is co-developing Z-Angle Memory with SoftBank’s Saimemory. The core technology comes from Intel’s Next Generation DRAM Bonding work under the US Department of Energy and NNSA Advanced Memory Technology program, run with Sandia, Lawrence Livermore and Los Alamos national laboratories.

  • Claimed 2 to 3 times the capacity of HBM, more bandwidth, and about half the power
  • Built from proven DRAM, stacked and bonded in a new way
  • Prototype around 2027, products around 2029

Bet two: XBM

A newly published Intel patent describes XBM, a stacked memory that drops HBM’s expensive silicon interposer. It connects over a 32 GT/s UCIe link instead of a very wide parallel bus, builds DRAM cells in the chip’s upper wiring layers, and includes built-in repair for defects. The goal is HBM4-class performance in a smaller, cheaper package, with memory sitting directly on CPUs and accelerators.

The people behind it

Lip-Bu Tan has called new memory architectures his pet project, saying memory is ripe for innovation and that stacking memory directly on a CPU could make a lot of sense. Intel has also reportedly recruited former SK hynix CEO Seok-Hee Lee.

The bottom line

TSMC would have to start memory from scratch, and doing so would put it in competition with its own HBM partners. Intel already has a stacked-memory program backed by US national labs, a second design on file, its own fabs, and EMIB-T and Foveros packaging to bond memory to compute. When the industry looks for a way past the HBM wall, Intel is the logic maker already building one.

Sources: Bloomberg, November 2026; HPCwire; Tom’s Hardware; TrendForce; Blocks and Files. Not investment advice.

Get the Intel bull case in your inbox

Not investment advice. The author may hold positions in the securities discussed.

Leave a Reply