Sources: Google is in talks with Samsung to manufacture a memory input-output die for Icefish, its 10th-generation TPU, and TSMC will make the computing engine
Google is considering giving Samsung Electronics a crucial role in making a component of one of its most advanced future AI chips …
Context & Ripple Effects
Google’s TPU supply chain has been moving toward specialization: related reporting says it is discussing a Marvell-developed memory processing unit alongside TPUs, while the Icefish plan would separate the computing engine from a memory input-output die.
The reported arrangement also revisits Google’s manufacturing relationship with Samsung. Earlier coverage described Google moving its Pixel Tensor G5 production from Samsung’s foundry to TSMC; the Icefish talks would give Samsung a role in a different, advanced AI-chip component as demand pressures TSMC capacity.
First-order effects
- If talks result in an agreement, Google would split Icefish production between TSMC for compute and Samsung for the memory input-output die, rather than relying on one foundry for the full chip.
- Samsung would gain a potential advanced-chip manufacturing assignment from Google, while TSMC would retain the central compute-engine portion of the TPU program.
Second-order effects
- A split design and manufacturing flow would require tighter coordination across chip interfaces, packaging, yields, and production schedules, making the availability of the finished TPU dependent on both suppliers.
- The move could reduce Google’s exposure to a single foundry’s capacity for the full product and create a more credible opening for Samsung to win other AI-chip production work as TSMC capacity is strained.
Third-order effects
- If large cloud companies increasingly divide AI accelerators into separately sourced compute and memory-related components, foundry competition may shift from winning an entire chip to securing strategically important portions of heterogeneous designs.
- The pattern points to AI-chip supply chains becoming more multi-vendor and modular, though Samsung’s lasting position will depend on converting prospective component work into reliable high-volume execution.
The trend: AI infrastructure buyers are redesigning and sourcing custom accelerators across multiple specialized partners to balance performance needs with constrained leading-edge manufacturing capacity.