The Information: Google plans to commission Samsung to produce the 10th generation AI chip "Icefish," diversifying TSMC's supply chain to mitigate shortages.

According to today's (11th) exclusive report by The Information, tech giant Google is engaging in in-depth negotiations with Samsung Electronics, planning to have some components of its upcoming tenth-generation AI chip — the Tensor Processing Unit (TPU, codenamed: Icefish), scheduled for mass production in 2028 — produced using Samsung's advanced 2-nanometer (2nm) process for the first time. This move aims to break the chip shortage crisis by diversifying the supply chain amid NVIDIA's aggressive sweep of TSMC's (Taiwan Semiconductor Manufacturing Company) capacity.
(Background summary: Google’s new open-source DiffusionGemma model: generates nearly 4 times faster, but lags behind Gemma 4 in quality)
(Additional background: Note — the photos you take, the words you say, Google now wants to store them for AI training (how to turn off training))

The global artificial intelligence (AI) computing power arms race is facing its most severe chip "supply cutoff" test. According to the latest exclusive report by authoritative tech media The Information, in response to the extreme capacity shortages caused by TSMC’s current surge in demand for AI chips from NVIDIA and others, Google is secretly planning a new "dual-track" supply chain strategy, actively negotiating with Samsung Electronics, TSMC’s top competitor.

It is understood that Google plans to incorporate Samsung’s advanced process in its next-generation AI chip project, using its 2-nanometer (2nm) technology to produce one of the core components of the upcoming "Tenth Generation Tensor Processing Unit (TPU)" — the "memory input/output die."

Codenamed "Icefish"! Google teams up with MediaTek to develop the next-generation TPU

This highly anticipated tenth-generation TPU accelerator, internally known as "Icefish," is expected to enter mass production as early as 2028. Notably, in this large-scale AI chip project, Google is not working alone but has once again enlisted Taiwanese chip design giant MediaTek for in-depth ASIC (custom chip) architecture technology collaboration.

In the currently planned hybrid design scheme, Google adopts a highly pragmatic "mixed-chip" division of labor:

• TSMC: Will still be responsible for manufacturing the most critical and technically challenging "computing engine" of the chip, expected to use TSMC’s most advanced 1.4-nanometer (1.4nm) process.
• Samsung: Will be responsible for the "I/O die" connecting the computing core and high-bandwidth memory (HBM), utilizing its full-force development of the 2-nanometer process.

Samsung’s 2nm process faces a turning point, Wall Street awaits the final decision

Semiconductor chip analysts point out that this heavyweight negotiation is undoubtedly a morale booster for Samsung Electronics, which has struggled significantly in recent years to catch up in advanced process technology. For a long time, high-end AI ASIC chips have been almost exclusively dominated by TSMC, and Google’s proactive diversification of orders indicates that Samsung’s yield and efforts in the 2nm process are gradually gaining tangible recognition from tech giants.

However, since the project is still years away from mass production in 2028, negotiations remain in the early stages, and Google and Samsung have not yet signed final purchase orders or legally binding agreements. Nonetheless, Google’s move to diversify supply chains and reduce over-reliance on a single foundry has already sent a shockwave through Silicon Valley and the Taiwanese semiconductor chain, adding more variables to the future global AI computing power race.