Google's AI Chip Strategy Shift
Google in Talks with Samsung Foundry to Produce Next-Gen Icefish AI Chip Component
Key Points
Samsung is pursuing the foundry/AI business with large-scale funding, including plans to invest “over 110 trillion won, roughly $73 billion, through 2026” in AI semiconductor efforts, and it has already secured high-profile AI-chip work such as “a $16.5 billion agreement with Tesla to deliver AI chips.”
The Information’s reporting context highlights the pressure on TSMC: due to “high demand for AI chips, particularly from Nvidia,” TSMC “produces more than 90 percent of the world’s advanced AI chips,” squeezing available capacity for other buyers like Google.
Samsung already supplies key memory for Google’s TPU ecosystem—one article notes Samsung provides “over 60% of the High Bandwidth Memory used in Google’s seventh-generation Ironwood TPUs”—and the companies’ relationship is expected to deepen with “HBM4…Samsung’s next-generation memory technology set to be unveiled in 2026.”
Talks are still early: while Google and Samsung Foundry discussed producing components for Icefish, an article says no “formal agreement has yet been finalized,” and another report (Reuters) frames the item as an Information report based on “two people familiar with the matter.”
Google is in talks with Samsung to make part of its next AI chip, codenamed Icefish, according to The Information. Samsung would use its 2-nanometer process to build a memory input-output die, while TSMC would still make the core computing engine. Mass production is targeted for 2028.
The move signals a major shift in how Google builds its AI hardware. TSMC now makes more than 90% of the world's advanced AI chips, per Reuters. That dominance has left little room for companies like Google, pushing them to find backup suppliers.
Google Splits Icefish Production Between Two Foundries
The Icefish chip is designed in two parts. TSMC will make the compute engine — the brain of the chip — using its most advanced process technology, reportedly 1.4nm. Samsung would handle a separate memory I/O die using its 2nm process, according to MarketScreener. No formal deal has been signed yet.
Google is also working with MediaTek on the Icefish project, The Information reported. That collaboration helps Google move away from its long reliance on Broadcom. Broadcom has historically added a 15% to 20% markup on components it sources on Google's behalf, according to analysts at Counterpoint Research.
TSMC Capacity Crunch Forces Google to Diversify
Nvidia's soaring chip demand has locked up most of TSMC's advanced production lines. TSMC is reportedly fully booked through 2028 on its 2nm nodes. That leaves companies like Google scrambling for alternatives. Google executives visited Samsung's fabrication plant in Taylor, Texas in early 2024 to evaluate its capabilities, according to Reuters.
Google recently placed an order with Intel Foundry for over 3 million TPU units for 2028 delivery, per MoneyCheck. That makes Intel a third pillar in Google's manufacturing strategy. By spreading production across TSMC, Samsung, and Intel, Google reduces its dependence on any single supplier.
Samsung Bets $73 Billion to Win AI Chip Business
Samsung is spending heavily to compete. Co-CEO Jun Young-hyun announced an investment of 110 trillion won — roughly $73 billion — for 2026 alone to "secure leadership in the AI semiconductor era." The company recently signed a $16.5 billion deal with Tesla to manufacture AI chips, giving it a high-profile win to show potential customers, according to Parameter.
Samsung already supplies over 60% of the High Bandwidth Memory used in Google's seventh-generation Ironwood TPUs, according to MoneyCheck. Its next-generation memory technology, HBM4, is expected to launch in 2026. Landing the Icefish contract would be a major boost for Samsung's foundry division, which has posted losses for five straight quarters.
Samsung's Yield Gap Remains the Biggest Hurdle
Samsung's 2nm yields have improved from around 20% in late 2024 to roughly 55% today. But TSMC still runs at 80% to 90% yields on advanced nodes, according to analysts cited by MoneyCheck. Experts say Samsung needs to hit at least 60% stable yields before it can win top-tier logic chip orders.
That gap explains why Google is only asking Samsung to make the I/O die — not the main compute engine. The I/O die is less performance-sensitive, making it a lower-risk test for Samsung's process. If Samsung proves its reliability on Icefish, it could open the door to larger orders down the road.
Sentiment
Mostly PositivePublishers
21
Articles
76
Reach
97
