The AI chip supply chain is undergoing a quiet but significant realignment. Alphabet Inc (GOOGL) is in talks with Samsung Electronics Co., Ltd. (SSNLF) to have the South Korean firm manufacture some components for its next-generation Tensor Processing Unit (TPU), marking the latest move by a chipmaker to diversify its suppliers amid a persistent capacity crunch at Taiwan Semiconductor Manufacturing (TSM).
According to a report by The Information, two people with direct knowledge of the matter said that Google is negotiating with Samsung to produce the memory I/O chip for its tenth-generation TPU, codenamed "Icefish," using a 2-nanometer process. This independent silicon die is responsible for connecting the main processor to memory and is crucial for the sustained, efficient operation of AI chips. Both companies declined to comment.
Under the current plan, the chip's core computing engine would still be manufactured by TSMC using a more advanced 1.4-nanometer process, with TSMC handling the most demanding part of the production. Notably, Google has historically entrusted the entire manufacturing of its TPUs to TSMC. However, with a surge in orders for AI chips from companies like Nvidia, TSMC's advanced production capacity is under significant strain, prompting Google to seek supplementary manufacturing resources.
If the collaboration materializes, it would provide Samsung with a high-profile opportunity to prove its capabilities in one of the world's most-watched semiconductor arenas. It would also further strengthen South Korea's increasingly prominent strategic position within the AI chip supply chain.
Dual-Source Strategy: TSMC Leads, Samsung Steps In
The memory I/O chip plays a critical role in AI processor architecture. AI chips require a continuous data flow to keep their computing cores running at full capacity, and this chip serves as the high-speed conduit between the main processor and memory. Google's arrangement is for TSMC to manufacture the most advanced 1.4-nanometer computing engine, while outsourcing this relatively independent but equally crucial component to Samsung for production using a 2-nanometer process.
In chip manufacturing, a smaller process node number typically represents a newer generation of technology, implying higher transistor density and better performance or power efficiency, although the number is no longer a literal measure of transistor size.
The context for Google's move is not only TSMC's structural capacity shortage but also the expansion of its own chip business. As TPUs begin to attract more external customers, Google's need for mass production capacity has increased. The Information reported that capacity constraints are already pushing several chip designers to seek advanced packaging resources outside of TSMC. Advanced packaging technology is responsible for integrating computing chips with high-bandwidth memory in AI processors.
Icefish is still in the design phase, with Google collaborating with MediaTek on its development. Google brought MediaTek into the fold for some AI chip design work last year, having previously worked primarily with Broadcom on most TPU designs. According to sources, the chip could enter mass production as early as 2028, but the plans remain subject to change.
Samsung's Crucial Bid for Foundry Breakthrough
For Samsung, securing the Google order holds significant strategic importance. Samsung began laying the groundwork for its foundry business in 2005 and established an independent foundry division in 2017. However, its efforts to close the gap with TSMC in advanced processes have progressed slowly, with high client expectations for cost, production stability, and ramp-up capacity keeping the pressure on.
Recently, Samsung has made some progress in securing high-end AI chip orders. Electric vehicle maker Tesla commissioned Samsung last year to produce its next-generation AI6 chip. For Nvidia's upcoming Vera Rubin platform, Samsung is also set to produce the language processing unit (LPU) designed to enhance inference performance and efficiency. If Google's Icefish project proceeds, it would further serve as a market endorsement for Samsung's advanced process foundry capabilities and provide a comprehensive stress test for its 2-nanometer technology in a highly scrutinized sector.
South Korean Semiconductors: A Strategic New Pillar in the AI Supply Chain
A comprehensive shortage in the AI chip supply chain, from accelerators to memory chips, is driving the entire industry to look toward South Korea. As the home to two of the world's top three memory chip manufacturers, South Korea, through Samsung and SK Hynix, holds significant sway over the supply and cost of critical memory components.
This trend was further corroborated this week. Nvidia CEO Jensen Huang concluded a five-day visit to South Korea, his second trip to the country in seven months. During the visit, Nvidia announced a "multi-year technology collaboration" agreement with SK Hynix and signed cooperation deals with several South Korean companies.
As leading tech companies accelerate their strategies to diversify suppliers, South Korea's semiconductor industry is becoming an increasingly indispensable player in the global race for AI infrastructure.
Comments