Google is in talks with Samsung to build next-gen 'Icefish' TPU on 1.4nm and 2nm process split

Google is reportedly in talks with Samsung Electronics to manufacture part of its next-generation AI processor. According to The Information, citing people familiar with the matter, the chip in question is Google's 10th-generation Tensor Processing Unit, codenamed "Icefish", with mass production potentially beginning as early as 2028.

The arrangement under discussion would split manufacturing between two foundries. TSMC would handle the main compute die on its 1.4nm process, while Samsung would produce the memory I/O die using its 2nm technology. Google is co-developing Icefish with MediaTek, and the chip is still under development.

The Information noted that Samsung's deep understanding of memory characteristics, including HBM, appears to have been a key factor in Google's consideration, and Samsung's ability to handle memory production, foundry work, and advanced packaging under one roof gives it a compelling integrated pitch.

3

Samsung already supplies the HBM used in Google's existing TPUs, meaning that a broader role in Icefish could involve the company's memory division, foundry business, and packaging unit in a single project. Samsung's Taylor fab in Texas is reportedly being discussed as a potential production site.

AI demand has pushed TSMC's production to its limits, prompting major chip designers to look for alternatives. Earlier this week, a separate report revealed that Google is also in talks with Intel to manufacture more than three million TPUs in 2028, and NVIDIA has been evaluating Intel for future multi-die GPU designs.

3

Samsung's potential role in Icefish would sit alongside that arrangement rather than replace it, with Google effectively spreading its manufacturing footprint across three foundries for its next-generation AI chips. Samsung has also been building momentum on the foundry side. The company secured a $16.5 billion contract last year to manufacture Tesla's next-generation AI6 chip using its 2nm process, and is also producing components for NVIDIA's Vera Rubin platform.