Industry

Glass Substrates Enter Early Production to Replace Organic Chip Packaging

By Reporting from New York

Glass Substrates Enter Early Production to Replace Organic Chip Packaging
Photo via Wikimedia Commons

NEW YORK — What happened

Semiconductor packaging suppliers have begun pilot production lines for glass-core IC substrates, marking the start of a transition away from traditional organic substrates for multi-die AI processors.

Why it matters

Modern data center accelerators combine multiple compute chiplets and memory stacks onto a single substrate. Organic materials warp under extreme thermal stress and limit the density of microscopic electrical vias. Glass offers superior physical stability, lower thermal expansion, and flatter surfaces, allowing for vastly denser interconnects between chiplets.

Deep dive

As server chips grow past 100x100mm in total substrate area, organic substrates hit physical limits. Glass substrates feature ultra-low thermal expansion that matches silicon, allowing designers to etch Through-Glass Vias at pitches under 10 micrometers—nearly four times denser than organic alternatives. This enables faster data flow between memory and compute logic while drawing less power. However, cutting and drilling glass without creating micro-cracks requires specialized laser ablation tools. Supply chain reports indicate that initial glass substrate yield loss is primarily caused by edge chipping during panel singulation, keeping costs significantly higher than traditional packaging through 2026.

Report check

Reports suggest glass substrates will entirely replace organic substrates across all tech sectors by 2027. What is verified: Pilot lines are actively producing high-density test substrates for enterprise clients. Disproven claims: Organic substrates will remain the dominant material for mainstream CPUs and consumer devices for years due to lower production costs.

Open questions

Will the high tooling costs of laser-drilling Through-Glass Vias restrict glass substrates to ultra-high-end AI hardware for the foreseeable future?