Google TPU v9 Slightly Delayed, Marvell's Supply Chain Position Remains Solid
nashnova research
Google's next-gen TPU v9 has slipped roughly three weeks due to substrate yield issues, pushing mass-production shipments to Q1 2028; analysts say Marvell's medium-to-long-term role in Google's ASIC supply chain remains intact.
What caused the delay?
GF Securities analyst Jeff Pu reports that TPU v9's Core A — codenamed "Humufish" — fell behind by about three weeks, likely due to a redesign.
The bottleneck sits in EMIB substrates — embedded multi-die interconnect bridge, a packaging method that links multiple chips on one substrate via a silicon bridge.
This means → both the chip design and the packaging substrate hit problems at the same time, compounding the schedule slip.
How far did the timeline shift?
The sampling target still stands at November 2025, but tape-out — committing the design to silicon — is now pushed to January 2027.
Mass-production tape-out moves to Q2 2027; final shipments slip to Q1 2028.
In plain terms = from "design lock" to "chips in customers' hands," the entire chain shifted roughly one quarter.
Is the yield problem fixable?
The supply chain has identified the root cause of the EMIB substrate yield issue; current yield sits at roughly 45%–50%.
This means → with the root cause known, there is a foundation for faster yield ramp — but whether the ramp actually materializes is the key variable for any further delay.
Who takes the hit — and who doesn't?
Short-term pressure falls on MediaTek: the Core A redesign forces matching changes to MediaTek's Core B and I/O portions, disrupting its order cadence.
Marvell, as a core participant in Google's ASIC supply chain, sees no material change to its medium-to-long-term positioning.
This reflects a pattern: a modest delay hits the "design-execution layer" hardest, while the "architecture-partnership layer" absorbs it.
What else did TPU capacity constraints cause?
Demand for Google's TPU v7 is intense — from internal teams and external clients like Anthropic alike.
GF Securities believes this resource competition may partly explain why Gemini 4 launched later than expected: Google allocated heavy TPU v7 capacity to external clients in H1 2026, squeezing internal compute.
Google plans to raise internal TPU allocation in H2 2026. This means → training conditions for the next Gemini generation should improve, but the first-half "compute concession" already slowed the release cadence.
市场有风险，内容仅供研究参考，不构成投资建议。