Samsung Electronics could win a role in producing Google's next-generation AI chip, codenamed "Icefish," under a split-manufacturing plan reported by The Information. According to the June 11 report, Google is in talks to have Samsung's foundry produce the input/output die for its 10th-generation Tensor Processing Unit on a 2-nanometer process, while TSMC manufactures the main compute die at 1.4nm. The talks are preliminary, with no agreement finalized, and the plan could change.
The detail that makes this more than a routine supplier note is the word "part." Google has historically had TSMC build its TPUs entirely. Splitting one chip across two foundries — and two different process nodes — is a specific engineering choice, and it is the choice that opens a door for Samsung.
Why You Can Split a Chip in Two
Modern AI chips are no longer single slabs of silicon. They are assembled from multiple smaller dies, or chiplets, each handling a different job and later bonded together in an advanced package. The compute die does the math; the input/output die shuttles data to high-bandwidth memory — a function that, in an AI chip moving enormous volumes of data continuously, is as much a bottleneck as raw compute.
Crucially, those two dies do not need the same manufacturing process. The compute die benefits from the densest, most advanced node available, which is why Google would keep it on TSMC's leading-edge 1.4nm. The I/O die does not need the bleeding edge to do its job well, so Samsung's 2nm process is enough — and that gap is precisely what creates the opening. Samsung has built its credibility for exactly this kind of work through its sixth-generation HBM4 memory, where its System LSI division designs a logic base die that its foundry manufactures. The I/O die is adjacent territory, and there is even a reported possibility that Samsung could supply the HBM, make the I/O die, and handle the packaging that assembles it with TSMC's compute die.
TPUs are Google's in-house AI chips, used to train and run its Gemini models and increasingly sold to external cloud customers. Earlier TPUs were designed with Broadcom; Icefish is reportedly being developed with Taiwan's MediaTek, with mass production targeted as early as 2028. Google unveiled its eighth-generation TPUs — the training-focused 8t and inference-focused 8i — at Cloud Next in April; Icefish is the step beyond.
Why Google Would Reach Beyond TSMC Now
The Information attributes the shift to two converging pressures. TSMC's advanced-node capacity is stretched thin by surging AI orders from Nvidia and others, while Google's own TPU volumes are climbing as external sales grow. Splitting the chip lets Google tap Samsung's 2nm capacity for the part that doesn't need TSMC's most advanced node, while keeping the performance-critical compute logic where it wants it. For Samsung — which has struggled for years to close the gap with TSMC in advanced logic — landing even a component role on a hyperscaler's flagship AI chip would be a notable validation of its 2nm process. Alphabet shares fell about 1.9% after the report.
Read more: Chip Supply Shifts From TSMC: Google's Reported Intel TPU Order Lifts Foundry, but Doubts Remain
A Pattern of Hedging Against One Foundry
The potential deal fits a broader pattern of big tech de-risking from TSMC's capacity limits by turning to Samsung. Bloomberg reported last month that Apple executives visited Samsung's fab in Texas to discuss foundry cooperation, as TSMC's capacity falls short of Apple's targets. Samsung has also secured a roughly $16.5 billion contract to build Tesla's next-generation AI6 chip at its Taylor, Texas plant, and is already manufacturing components for Nvidia. Each of these is preliminary or narrow on its own; together they describe a market actively looking for a credible second source.
Samsung's chip chief detailed the Nvidia work this week. After meeting Nvidia CEO Jensen Huang on June 8, Jun Young-hyun, vice chairman and head of Samsung's Device Solutions division, said Samsung is producing Nvidia's autonomous-driving and Groq chips — the Drive AGX Thor processor for self-driving vehicles, and the language processing unit for U.S. inference startup Groq — on 4nm and 8nm processes, and that the two sides discussed next-generation products as well.
None of this makes Samsung TSMC's equal — its foundry still holds a single-digit share of the market and has trailed on yield for years. But the Icefish talks, if they hold, would mark something new: a flagship AI chip from one of the world's largest buyers, deliberately designed so that Samsung makes a piece of it. In a market where TSMC has been the only option for the leading edge, even a slice is a foot in the door.
Frequently Asked Questions
What is Google's Icefish TPU?
Icefish is the codename for Google's 10th-generation Tensor Processing Unit, its in-house AI chip used to train and run models like Gemini and increasingly sold to cloud customers. It is still in development, reportedly designed with MediaTek, with mass production targeted as early as 2028.
What is an I/O die?
An input/output die is the part of a chip that moves data between the main compute logic and the chip's high-bandwidth memory. In AI processors that shuttle enormous volumes of data continuously, it is a critical component — and because it does not require the most advanced manufacturing node, it can be built on a slightly less cutting-edge process than the compute die.
Why would Samsung make only part of Google's chip?
Modern AI chips are built from multiple dies bonded together, and they don't all need the same manufacturing process. Under the reported plan, TSMC would build the performance-critical compute die on its 1.4nm node while Samsung makes the memory I/O die on 2nm. Splitting the work lets Google ease pressure on TSMC's strained advanced-node capacity while keeping the most demanding part with TSMC.
Why does this matter for Samsung?
Samsung's foundry has trailed TSMC for years and holds less than 10% of the market. Winning even a component role on a major hyperscaler's flagship AI chip would validate its 2nm process and add a high-profile customer, joining recent foundry wins with Tesla and Nvidia as Samsung tries to rebuild credibility in advanced chipmaking.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
