Overview
- Multiple outlets on June 11 reported that Google is in early talks to have TSMC make the main compute die for its next‑generation TPU, codenamed Icefish, while Samsung would manufacture a separate memory input/output (I/O) die using its 2nm process.
- The Icefish design is reported to be in development with mass production targeted as soon as 2028, but no contracts have been signed and the companies cited in the reports declined to confirm the plans.
- Separately, earlier reporting this week said Google is in discussions with Intel to produce more than three million TPUs for 2028, reflecting a broader strategy to spread volume across multiple foundries.
- The proposed split lets TSMC focus on the high‑performance compute chip while Samsung would handle the I/O die that links the processor to high‑bandwidth memory, a role that depends on tight yields for advanced 2nm logic and on packaging choices for HBM integration.
- If confirmed, the moves could reduce Google’s reliance on TSMC, strengthen Samsung’s push into advanced contract manufacturing, bolster Intel’s foundry and packaging business, and affect cloud AI costs and supply resilience for customers.