The discussions, revealed by sources familiar with the matter, center on the upcoming tenth-generation Tensor Processing Unit (TPU), internally codenamed “Icefish.”While Taiwan Semiconductor Manufacturing Company (TSMC) will remain the lead manufacturer for the chip’s primary computing engine, Samsung is poised to play a critical supporting role.

According to the current roadmap, which remains in the design stage, Samsung would leverage its cutting-edge 2-nanometer production technology to manufacture a specific silicon die known as the memory input/output (I/O) chip.  This small but essential component is the bridge between the main processor and the memory. For an AI accelerator running massive language models, this I/O chip dictates the speed of data flow; without a high-performance connection, even the most powerful compute engines risk stalling. 

 A Dual-Source Strategy

This potential partnership marks a significant departure for Google.Historically, the tech giant entrusted the entirety of its TPU production to TSMC.

Two major pressures are driving this shift. Capacity constraints at TSMC represent a growing industry risk. The Taiwanese giant is running at maximum capacity to satisfy insatiable demand for Nvidia’s AI chips, leaving less room for other hyperscalers.Furthermore, Google’s own TPU business is expanding rapidly. Once a purely internal tool, the TPU is increasingly a product sold to external cloud customers, forcing Google to secure a more resilient supply chain. 

Under the proposed “Icefish” blueprint, TSMC would keep the crown jewel of the project.The South Korean firm will manufacture the core compute engine using an ultra-advanced 1.4-nanometer process.This dual-track production allows Google to mitigate risk while maintaining bleeding-edge performance.

The Road to 2028

Industry observers note that the “Icefish” chip is still deep in the design phase, with Google collaborating closely with Taiwanese chip designer MediaTek on the architecture.Mass production is not expected to begin until 2028,leaving a long runway for design adjustments and validation. 

For Samsung, securing this order would represent a crucial validation of its foundry ambitions. Despite investing heavily in its contract manufacturing division since 2005, Samsung has struggled to close the yield and efficiency gap with TSMC.

A successful partnership on “Icefish” would place Samsung’s 2nm technology directly into the high-stakes AI arena. It joins a growing list of wins for the Korean giant, including orders from Tesla for its AI6 chip and a role in Nvidia’s upcoming Vera Rubin platform. 

Both Google and Samsung have declined to comment on the negotiations. However, if finalized, the deal would signal a fundamental realignment of the AI supply chain moving away from a single-source dependency toward a more diversified, albeit complex, manufacturing future.