Google Eyes Samsung 2nm for Next-Gen AI Chips

According to people familiar with the matter, Alphabet’s Google is in discussions with Samsung Electronics regarding the potential manufacturing of key components for its next-generation artificial intelligence (AI) processor.

Reports indicate that Google plans to assign TSMC the production of the main compute die for its upcoming “Icefish” Tensor Processing Unit (TPU), while Samsung Electronics could be responsible for fabricating chip-to-memory interconnect components using its advanced 2nm process technology. The TPU, still under development, is being co-designed with chip design partner MediaTek, with mass production potentially expected as early as 2028.

If finalized, the collaboration would represent a significant win for Samsung Electronics in expanding its foundry business, particularly in the highly competitive advanced-node segment. The 2nm process is expected to enable higher transistor density, improved performance, lower power consumption, and stronger AI computing efficiency.

In April, Samsung Electronics stated that it expects to secure more customers for its next-generation manufacturing technologies and is also evaluating the construction of a second fabrication plant in Texas to expand capacity. In July 2025, the company also signed a $16.5 billion agreement with Tesla to produce AI chips using its advanced process nodes.

The development also highlights Google’s ongoing efforts to diversify its semiconductor supply chain and reduce reliance on TSMC, which continues to face surging demand from the global AI boom and could become a potential bottleneck for the industry.

Earlier reports also suggested that Google has been in talks with Intel regarding the production of more than 3 million TPU units by 2028. Google’s in-house AI accelerators have increasingly emerged as strong competitors to Nvidia GPUs, with growing TPU adoption playing a key role in accelerating its cloud business expansion. In April, Google also introduced two new custom AI chips designed for training and inference workloads, further strengthening its AI hardware ecosystem.