Will Samsung Make New Google ‘Icefish’ TPU?

Rumors are swirling that Google’s next-generation Tensor Processing Unit, codenamed 'Icefish,' may be manufactured by Samsung—a move that would challenge TSMC’s near-total grip on AI chip production. The speculation, first reported by analysts tracking Google’s custom silicon roadmap, underscores a fragile truth: despite years of geopolitical pressure and billions in investment, semiconductor diversification remains agonizingly slow. If Samsung secures the Icefish contract, it would mark the first time Google has tapped a second foundry for its most advanced TPU, potentially reshaping the balance of power in the $500 billion chip industry.

Google has relied exclusively on TSMC for its TPU lineup since the first iteration in 2016. The cloud giant’s custom AI accelerators power everything from search ranking to Gemini model training, making them a strategic linchpin for Alphabet. Yet TSMC’s concentration in Taiwan—a flashpoint for U.S.-China tensions—has forced buyers like Google, Apple, and Nvidia to explore alternatives. Samsung’s foundry division, Samsung Electronics Co. Ltd., has long pitched itself as the backup, but yield rate issues and process node gaps have kept it from winning marquee AI chip orders.

The Icefish rumor, which surfaced in June 2026 via industry insiders and leaked supply chain documents, suggests Google is testing Samsung’s 3nm Gate-All-Around (GAA) process for a portion of Icefish production. Icefish is expected to succeed the current Trillium TPU, which Google launched in 2024 with a claimed 4.7x performance boost over the previous generation. If Samsung can hit the required performance and power-efficiency targets, the move would validate years of investment in its GAA transistor architecture—a technology both Intel and TSMC are also racing to commercialize.

Samsung’s foundry business has struggled to close the gap with TSMC, which commands roughly 90% of the leading-edge AI chip market. In 2025, Samsung posted its fourth consecutive quarterly loss in its foundry segment, partly due to low yields on its 3nm node. By contrast, TSMC’s 3nm N3E process achieved near-100% yield for Nvidia’s Blackwell GPUs. The Icefish contract, even as a secondary supplier, would be a vital win for Samsung’s credibility and financial health.

Yet the benefits of diversification extend beyond corporate balance sheets. Policymakers in Washington, Brussels, and Tokyo have poured tens of billions into domestic chip fabrication, largely driven by a desire to reduce reliance on a single Taiwanese supplier. The U.S. CHIPS Act alone has allocated $52.7 billion in subsidies, but constructing advanced fabs takes 3–5 years, and no non-TSMC facility has yet mass-produced a 3nm AI chip at scale. Analysts at Bernstein caution that even if Samsung wins Icefish, TSMC will likely retain the majority of wafer starts, leaving systemic risk largely intact.

Looking ahead, the Icefish decision could catalyze a broader shift. Google’s willingness to split production signals to other hyperscalers—Microsoft, Amazon, Meta—that dual-sourcing is feasible. Intel foundry, meanwhile, is also making a play for Google’s future chips with its 18A process. If Samsung proves it can deliver Icefish on time and within spec, the competitive dynamics of the AI silicon market could finally loosen TSMC’s stranglehold. For now, the industry watches Seoul and Silicon Valley with anticipation—and a dose of caution.