In a major structural move to break away from Taiwan Semiconductor Manufacturing Co.’s (TSMC) near-monopoly on advanced foundry capacity, Alphabet’s Google is in advanced talks with Samsung Electronics to manufacture a critical portion of its tenth-generation custom artificial intelligence processor.

According to a report from The Information, the tech giant is pursuing a dual-foundry, split-manufacturing strategy for its upcoming Tensor Processing Unit (TPU), code-named “Icefish.” Under the developing roadmap, TSMC will continue to fabricate the chip’s core computing engine, while Samsung will take over production of the memory input-output (I/O) die using its cutting-edge 2-nanometer (nm) process node.

The “Icefish” processor, which is being co-designed alongside Taiwanese chip architect MediaTek, is targeted for mass production as early as 2028. The deal marks a major milestone for Google’s cloud infrastructure expansion and a hard-fought validation for Samsung’s contract manufacturing division.

The “Icefish” Split: Dividing the Silicon

To bypass the severe capacity constraints limiting advanced global chip foundries, Google’s hardware engineers are separating the structural components of the 10th-generation TPU into distinct silicon dies.

• The Computing Core (TSMC): The main computing engine, which requires intense calculation density to run complex generative models like Gemini, will be outsourced to TSMC. Google plans to utilize TSMC’s upcoming, ultra-advanced 1.4nm process technology for this central brain.
• The Memory Interconnect (Samsung): Samsung is poised to manufacture the memory I/O die. This component serves as the vital digital bridge connecting the main processor to High-Bandwidth Memory (HBM). Because AI hardware requires an immense, continuous flow of data to keep its computing cores from sitting idle, the efficiency of this specific interface is a critical bottleneck for system performance.

Google’s decision to tap Samsung for this specific component points directly to the South Korean giant’s unique status as a vertically integrated semiconductor company. Because Samsung manufactures its own advanced HBM, it possesses a deep, native understanding of memory architectures, packaging characteristics, and interface specifications—giving it a clear competitive edge over pure-play foundries when designing complex memory-routing silicon.

Breaking the TSMC Bottleneck

Google has historically relied almost exclusively on TSMC to manufacture its custom TPUs, which first began powering its search data centers back in 2016. However, the global explosion of generative AI and automated agent swarms has triggered unprecedented demand for server-grade silicon.

With companies like NVIDIA, Apple, and AMD aggressively booking up TSMC’s production lines years in advance, Google has been forced to aggressively diversify its supply network to guarantee it can support its scaling enterprise cloud clients.

This diversification isn’t limited to Samsung. Just days prior to the Icefish revelations, reports emerged that Google is also in preliminary talks with Intel Foundry Services to manufacture upwards of three million auxiliary TPUs by 2028. By spreading its infrastructure requirements across TSMC, Samsung, and Intel, Google is systematically insulating its AI data centers from single-point-of-failure vulnerabilities in global logistics.

A Massive Windfall for Samsung Foundry

Landing Google as a high-profile validation partner represents an immense strategic victory for Samsung’s foundry ecosystem, which has struggled to close the market-share gap with TSMC over the last decade.

The company is committing over $73 billion (110 trillion won) to aggressively expand its AI semiconductor manufacturing lines, and the investment is beginning to yield major institutional clients:

• The Tesla Blueprint: Samsung secured a multi-year contract to manufacture Tesla’s upcoming next-generation Full Self-Driving (FSD) AI6 chip architecture.
• The NVIDIA Integration: The firm has separately captured manufacturing allocations for automated Language Processing Units (LPUs) bound for NVIDIA’s next-generation “Vera Rubin” supercomputing platform.

Securing a piece of Google’s flagship cloud infrastructure proves that major tech giants are increasingly viewing Samsung’s 2nm node as a trusted alternative for mission-critical AI hardware. As these initial design pipelines mature into commercial scale over the next two years, the split-manufacturing blueprint pioneered by Google could easily become the standard playbook for enterprise platforms looking to survive the global computing shortage.