Samsung Considers Outsourcing Google TPU Backend Design as 2nm Demand Tightens
Alina Collins
Samsung is considering outsourcing back-end design work on an I/O chip for Google's tenth-generation TPU, as surging 2nm foundry orders stretch its engineering headcount to the limit — a sign that Samsung's capacity bottleneck is shifting from wafers to people.
What exactly is Samsung looking to outsource — and why?
The scope covers only back-end design — turning the chip's logic blueprint into a manufacturing-ready physical layout, including place-and-route and final sign-off. Samsung will still fabricate the I/O chip on its own 2nm process.
This means → Samsung is short on design engineers, not fab capacity. The 2nm order book has outgrown the in-house team.
Samsung handled back-end design in-house for Tesla's 2nm autonomous-driving chip, but with more projects piling up the staffing crunch is now visible.
Where does this chip come from, and what does it do?
The project is codenamed Icefish — Google's tenth-gen TPU, co-developed with MediaTek, with mass production possible as early as 2028.
Samsung's part is the I/O chip that links processor to memory, built on 2nm. The main compute die goes to TSMC on a 1.4nm process.
In plain terms = one TPU, two chips, two foundries. Google gave the "brain" to TSMC and the "data courier" to Samsung.
Who might take the work — and do they even want it?
ADTechnology and Gaonchips are the leading candidates; both already work on Samsung 2nm projects. Yet neither is reportedly eager — ADTechnology is focused on its own ADP620 2nm CPU, aiming to push annual revenue past ₩1 trillion by 2028–2029. Gaonchips is preparing for South Korea's roughly ₩800 billion K-On-Device AI programme and co-developing a 5nm ADAS chip with Hyundai Motor.
Back-end design contracts typically run in the low billions of won — far below full-flow ASIC deals that can reach hundreds of billions to trillions of won. That limits the financial appeal for larger firms.
Alphachips is the outlier, actively pursuing the work as a growth opportunity. This reflects a dynamic where smaller design houses value the résumé line of a 2nm Google project more than the contract size alone.
How full is Samsung's 2nm order book?
Samsung's 2nm node (internal codename SF2) entered mass production in 2025, targeting mobile, HPC, AI and automotive.
Confirmed clients include Tesla (a deal worth $16.5 billion) and Qualcomm, whose CEO Cristiano Amon publicly mentioned 2nm foundry talks with Samsung. The Elec has also reported Anthropic and DeepX as 2nm customers, though that has not been independently verified.
This means → Samsung's 2nm line is absorbing orders across autonomous driving, mobile and AI simultaneously. The engineering strain is structural, not a one-off spike.
What is Google's own supply-chain play?
Google is diversifying fabrication: Reuters reported that Google has placed orders with Intel to manufacture more than 3 million TPUs starting in 2028.
Add Samsung for the I/O chip and TSMC for the main compute die, and Google's TPU supply chain will span three foundries.
In plain terms = Google refuses to put all its eggs in one basket — locking in TSMC, Samsung and Intel in parallel to secure production capacity for its custom silicon.
Content is for reference only, not financial advice.