Compute & silicon·Updated May 29, 2026

Silicon interposers

The passive silicon slab that wires HBM to logic inside CoWoS is reticle-limited and eats TSMC wafer capacity — so as AI packages balloon past reticle size, usable interposers per wafer collapse and packaging can't scale faster.

Tightness gaugeCritical

Tightness

71/ 100Critical

+4 this quarter — Interposers are made by TSMC inside CoWoS and are physically capped by the lithography reticle. As package sizes scale to 5.5x reticle and beyond, a 300mm wafer yields only a handful of large interposers, and panel/glass alternatives are years from volume — keeping this acutely tight.

Coverage35%
70
Lead time25%
70
Concentration20%
80
Momentum20%
65
CoWoS-S max interposer size
3.3x reticle (~2,700 mm²); larger needs CoWoS-L/R
as of 2026·TSMC 3DFabric
TSMC reticle-scaling roadmap
5.5x now → 9.5x 2027 → 14x 2028reticle multiples
as of 2026-04·TrendForce
Package size: Blackwell / Rubin / Rubin Ultra
~3.3x / ~4x / ~9xreticle (2,739 / 3,320 / 7,470 mm²)
as of 2026·TrendForce
Large-die yield per 300mm wafer (5.5x reticle)
4–7units per wafer
as of 2026-04·TrendForce
Panel/glass (CoPoS) volume production
2028–2029mass-production window
as of 2026-04·TrendForce
Why it gates the buildout

In CoWoS-S, the silicon interposer is the wafer-level layer that carries the dense interconnect between logic chiplets and the HBM stacked beside them. Because it is patterned by lithography it cannot exceed the reticle limit without stitching, and producing it consumes TSMC foundry capacity. As designs push past 3.3x reticle, only a few good interposers come off each round wafer — a hard cap on how many top-end accelerators can be packaged.

Who's exposed
CompanyRoleExposure
TSMCTSM
Fabricates the interposer within CoWoS; sole volume sourceDirect
NvidiaNVDA
Largest consumer; reportedly evaluating SiC interposer for RubinHigh
Glass-substrate/interposer material supplierModerate
Developing glass-interposer packagingWatch
Glass-substrate panel maker (US fab)Watch
Catalysts & timeline
NearMigration to CoWoS-L (LSI bridge) and CoWoS-R (organic RDL) eases the reticle ceiling for the largest packages.
2026TSMC CoPoS panel pilot line completes (~June 2026); early glass-interposer qualification.
2027–2029CoPoS/glass mass production and possible SiC-interposer adoption on Rubin-class parts break the round-wafer area limit.
What would loosen it

Shift to LSI-bridge (CoWoS-L), organic RDL (CoWoS-R) and panel/glass (CoPoS) packaging that escapes the circular-wafer area limit — but meaningful volume relief is a 2028+ event.

Latest developments
2026-04-13
TSMC CoPoS panel-level pilot line tooling underway (completion ~June); mass production eyed 2028–2029 as Rubin's 5.5x reticle yields as few as 4 units per wafer.
2026-01-01
TSMC pivots from monolithic CoWoS-S interposers toward CoWoS-L (LSI bridge) to exceed the reticle limit while doubling CoWoS capacity.
Linked bottlenecks
Leading-edge foundry (2nm / N2)CoWoS advanced packagingHBM (high-bandwidth memory)ABF substrates

Get the squeeze before the headlines.

A weekly read on the AI buildout's tightest constraints. Free, sourced.

Subscribe