MediaTek's Google TPU v9 faces potential delay over glass substrate yields, but 2026 orders remain strong

The glass substrate yield bottleneck on TPU v9 is less a packaging hiccup and more a structural signal: advanced packaging has become the new critical path in AI silicon. Technically, TPU v9's I/O density and die size have hit the signal-integrity ceiling of ABF organic substrates. Glass substrates—with lower dielectric loss and superior planarity—are the only viable path forward. Yet global production capacity is concentrated in a handful of players (Ibiden in Japan, Unimicron in Taiwan, China), and yield ramp cycles run 18 to 24 months. Any substrate-driven design revision cascades through RDL routing, thermal validation, and test, converting a "few weeks" slip into a full-quarter delay. Strategically, NVIDIA's GB300 and AMD's MI400 are locking in 2025 to 2026 cloud capex. If TPU v9 slides to H2 2026, Google faces near-term budget reallocation pressure between in-house silicon and purchased accelerators. Yet the strength of 2026 order flow signals the roadmap is intact—this is a timing issue, not a strategic pivot. MediaTek's role extends well beyond execution; it is the architectural anchor of Google's de-NVIDIA-ization strategy. Outlook (12 to 24 months): Glass substrates will shift from "premium upgrade" to "entry requirement" for AI-class dies. TSMC's CoWoS and Intel's EMIB both face substitution pressure. A confirmed delay at MediaTek's October earnings call will trigger capacity re-sequencing at ASE and Amkor, and accelerate capex cycles at Corning and AGC in upstream glass material supply.