Key Takeaways
- Samsung is in discussions with Google to produce a memory input-output component for the upcoming TPU generation, dubbed “Icefish”
- The primary computational core of Icefish will be fabricated by TSMC utilizing its cutting-edge 1.4nm process technology
- Mass production timeline is tentatively set for 2028
- MediaTek has been brought on board for the design phase of the 8th-generation TPU series, including Icefish
- Reports from earlier in the week indicated Google is also discussing potential production of more than 3 million TPUs with Intel for 2028
According to Thursday’s report from The Information, Alphabet’s (GOOGL) Google is pursuing negotiations with Samsung for the production of a critical element in its forthcoming tensor processing unit.
Currently undergoing design development, the processor known by its codename “Icefish” involves a strategic split in manufacturing responsibilities. TSMC has been tapped to fabricate the primary computational component leveraging its advanced 1.4nm process node, whereas Samsung would manufacture a distinct memory input-output die using its 2nm fabrication technology.
These negotiations remain preliminary, with no formal contracts yet established. Representatives from Alphabet, Samsung, and TSMC have not provided immediate responses to inquiries regarding these discussions.
With 2028 as the projected timeframe for volume production of Icefish, this schedule remains subject to modification. Industry sources indicate MediaTek’s participation in the architectural design of this processor.
Expanding Google’s Custom Silicon Initiative
Google has been aggressively expanding its proprietary AI chip portfolio to provide cloud computing clients with alternatives to Nvidia’s dominant GPU offerings. Revenue from TPU services has emerged as an increasingly significant contributor to Alphabet’s cloud division performance.
The technology giant previously introduced its 8th-generation TPU lineup during the Google Cloud Next 2026 conference held in Las Vegas this past April — featuring the TPU 8t optimized for AI model training operations, alongside the TPU 8i configured for inference workloads.
Icefish signifies Google’s next evolutionary leap beyond these current offerings. The magnitude of this advancement is considerable — the decision to divide manufacturing between TSMC and Samsung suggests a substantially more sophisticated production approach than Google’s previous chip development efforts.
Intel Enters the Manufacturing Conversation
This week’s Samsung-related developments follow another significant report. Monday’s coverage from The Information revealed that Google is simultaneously negotiating with Intel regarding the potential fabrication of over three million TPU units scheduled for 2028 delivery.
This arrangement could represent a meaningful opportunity for Intel as the company continues its efforts to revitalize its foundry operations. Google has yet to publicly acknowledge these Intel discussions.
TSMC has traditionally served as Google’s principal semiconductor manufacturing collaborator and maintains its crucial role in the Icefish initiative. The semiconductor industry continues experiencing significant capacity constraints as AI-driven customer demand accelerates.
At this stage, Icefish remains in the design phase, with the 2028 production schedule representing an objective rather than a confirmed commitment.


