Samsung Foundry is poised to land a marquee win against TSMC. Google is in advanced talks to hand Samsung Electronics the memory I/O die for its 10th-generation Tensor Processing Unit — code-named Icefish — manufactured on Samsung's 2-nanometer gate-all-around process, according to reports from Seoul Economic Daily on June 12. Mass production of Icefish is targeted for 2028.
Breaking The Single-Source AI Supply Chain
The memory I/O die sits between the compute tile and high-bandwidth memory stacks and is one of the most critical pieces of any modern AI accelerator. By splitting the design across two foundries, Google is taking direct aim at the supply-chain bottleneck that has emerged as NVIDIA's Vera Rubin ramp strained TSMC's leading-edge capacity and as hyperscalers including Apple and Anthropic crowded onto sub-3nm nodes.
Samsung Foundry's Comeback Story
Samsung Electronics' long-struggling foundry business is showing signs of life. Improving yields on its 2nm node, rising utilisation and a string of high-profile AI projects have analysts forecasting a return to operating profit as soon as the third quarter. TSMC was 11 times larger than Samsung's foundry business in the first quarter of 2026 — $35.8 billion versus $3.2 billion in revenue — but the Google win would mark Samsung's biggest leading-edge customer in years.
What Icefish Buys Google
Custom TPUs are central to Google's pitch against NVIDIA's H- and B-series accelerators in cloud training and inference. The Icefish generation is expected to power both internal Gemini workloads and Google Cloud's external TPU offering. Pairing a 2nm I/O die with TSMC compute tiles also gives Google a hedge against single-vendor pricing — a calculus that has driven NVIDIA's $150 billion Taiwan investment and prompted the recent NVIDIA-Doosan and NVIDIA-LG partnerships in Korea.
Reporting based on coverage from Seoul Economic Daily, Bloomberg and TrendForce.