Google is in discussions with Samsung Electronics to manufacture a critical component of its next-generation Tensor Processing Unit, a move that could reshape the competitive landscape of AI chip production and ripple through the broader tech supply chain.
The chip, codenamed Icefish, would use a split manufacturing approach. TSMC handles the primary computing die. Samsung would produce an interconnect module that links the processor to memory, leveraging its advanced 2-nanometer process technology.
Why Google is splitting the recipe
TSMC gets the main compute die. Samsung would handle the interconnect, the part that determines how quickly the processor can talk to memory.
Google has been designing its own TPUs since 2016, a decade-long effort to reduce dependence on third-party GPUs. The company uses these chips internally to power everything from search to its Gemini AI models.
