Google is in talks with Samsung Electronics to have part of its 10th-generation AI chip, the Tensor Processing Unit known internally as Icefish, manufactured by Samsung, according to a report by The Information. The chip is expected to enter mass production in 2028, and the move would give Google another manufacturing path as advanced capacity at TSMC remains tight.
The reported plan calls for Samsung to produce a memory input-output die for the next TPU using its 2nm process. This part connects the computing core to high-bandwidth memory, or HBM. The backdrop is straightforward: AI chip demand has surged, and companies such as NVIDIA have absorbed a large share of TSMC's leading-edge capacity, pushing Google to consider a broader supplier base.
Icefish would split production across two foundries
Under the design approach described in the report, TSMC would still handle the most critical compute engine for Icefish, using an advanced 1.4nm process, while Samsung would take on the I/O die with 2nm technology. The arrangement points to a divided manufacturing model rather than a full handoff, with each foundry responsible for different parts of the chip.
Google is also said to be working with MediaTek on ASIC architecture for the project. That indicates Icefish is being developed through a multi-party effort on the design side as well, instead of relying only on Google's in-house teams.
No final order has been signed yet
The talks are still at an early stage. The report says Google and Samsung have not signed a final order or a binding agreement, leaving time for the plan to change before the chip reaches production.
For Samsung, a deal of this kind would matter. High-end AI ASIC manufacturing has largely been concentrated at TSMC, so winning part of a future Google TPU program would be read as support for Samsung's progress in 2nm manufacturing. For Google, the rationale is more practical: secure optionality in a market where access to advanced chip production has become a constraint.