Google's "Icefish" TPU Turns to Samsung: Realities and Limitations of Semiconductor Supply Chain Diversification

Event Overview

According to a report by *The Information* on June 14, Google is in talks with Samsung to outsource part of the production of its next-generation TPU (Tensor Processing Unit) "Icefish" to Samsung's foundry. However, further reports from TechZine indicate that TSMC will still be responsible for the chip's core computing engine, manufactured using a 1.4nm process, while Samsung will only produce components that connect to memory. This division of labor clearly reveals the current landscape of the global AI chip supply chain: TSMC is almost irreplaceable in advanced logic processes, while Samsung can only secure limited orders for peripheral circuits.

Background: TSMC's Monopoly and Diversification Efforts

TSMC has long dominated the global advanced process foundry market. From 7nm to 5nm and then to 3nm, TSMC has outshone competitors in each generation of process technology and has become the sole choice for leading AI chip customers such as NVIDIA, AMD, Apple, and Broadcom. According to industry estimates, TSMC holds a 90% share of the high-end AI chip foundry market and approximately 72% of the overall global wafer foundry market.

Although regions such as the United States, Europe, and Japan have attempted to break TSMC's monopoly through policy subsidies (e.g., the U.S. CHIPS Act) and investments from local companies (e.g., Intel Foundry Services), progress has been slow. Intel's massive investment project in Ohio, USA, has faced repeated delays, and TSMC's own U.S. factory is struggling with talent shortages and cost issues. Other regions, such as Europe and Southeast Asia, have yet to produce competitive alternative capacity.

In-depth Analysis

Technological Impact: 1.4nm Remains TSMC's Moat

The computing engine of the "Icefish" TPU adopts TSMC's 1.4nm process, marking TSMC's continued lead in advanced nodes. Although Samsung is also advancing its 2nm and 1.4nm processes, its yield and performance have not yet been verified by major AI chip customers. Google's decision to keep the most complex part with TSMC and only assign peripheral interface circuits to Samsung reflects limited trust in Samsung's advanced processes. This division of labor reduces Samsung's strategic value at the technological level, positioning Samsung more as a capacity supplement rather than a substitute.

Supply Chain Impact: Who Benefits? Who Faces Risks?

• TSMC: Although some orders are transferred, it retains the production of the highest-value computing engine, and the 1.- TSMC: Although some orders have been transferred, it has retained the production of the highest-value computing engines, and 1.4nm capacity remains tight. Demand from major customers like Nvidia has already pushed TSMC to full capacity; Google's move can be seen as a tactical measure to ease its own capacity pressure, rather than a divergence from TSMC. In the long term, TSMC will still dominate advanced AI chip manufacturing processes.
• Samsung: Securing Google's order is a significant breakthrough for Samsung's foundry business, but it is limited to non-core components, with limited profit margins and strategic implications. Samsung needs to prove the competitiveness of its 2nm/1.4nm processes to win more orders.
• Nvidia: As one of TSMC's largest customers, Nvidia benefits from TSMC's continued capacity tightness—competitors (such as Google) find it difficult to obtain sufficient capacity, further solidifying Nvidia's market position.
• Other potential customers: AMD, Apple, Broadcom, etc. may also consider similar multi-sourcing strategies, but it will be difficult to fundamentally change their reliance on TSMC in the short term.

Competitive Landscape: TSMC's Position Unshakable

In the current global wafer foundry market, TSMC holds an absolute lead with a 72% share, followed by Samsung at about 12% and UMC at about 6%. In advanced processes (below 7nm), TSMC's share exceeds 90%. Although Intel's foundry is catching up, it has yet to receive large-scale orders from mainstream AI customers. Google's cooperation with Samsung is more of a tactical capacity supplement than a strategic replacement. TSMC's technological leadership and scale effects create strong competitive barriers.

Regional Impact: Slow Diversification

• United States: Although the CHIPS Act has spurred domestic fab construction, TSMC's and Intel's factories in the US are facing construction delays and cost overruns, making it difficult to produce advanced process chips in the short term. Google's Icefish being partly produced in Korea instead strengthens Korea's role in the semiconductor supply chain.
• Taiwan, China: TSMC continues to keep the most advanced manufacturing in Taiwan, maintaining its core position. Google outsourcing non-core components actually highlights Taiwan's irreplaceability in advanced manufacturing.
• South Korea: Samsung's winning of Google's order helps raise its foundry business profile, but if it cannot win core computing components, its industrial status will improve only marginally.
• China: Affected by export controls, China cannot obtain advanced process equipment and technology. It is not directly involved in this cooperation, but it further realizes the urgency of self-developed chip manufacturing.

Investment Perspective: Long-Term Value Centers on TSMC

The capital market has fully recognized TSMC's moat. Despite geopolitical risks, TSMC's technological leadership and pricing power make it a core target for AI infrastructure investment. The valuation increase of Samsung's foundry business still needs to be proven by more major customer orders. This round of Google cooperation will not significantly change investors' long-term judgment of the two companies.

Long-Term OutlookIn the next 3-5 years, TSMC will still hold the vast majority of advanced process share for AI chips. Samsung and Intel need at least 2-3 generations of stable performance to win the trust of major customers. However, the trend of supply chain diversification is irreversible. Leading customers (such as Google, Amazon, Microsoft) may continue to place small batch trial orders with other foundries, gradually forming a dual-source or multi-source supply model with TSMC as the primary and others as secondary. Over a 10-year horizon, if competitors can achieve breakthroughs at nodes below 1nm, the landscape may be reshaped, but TSMC will still likely maintain its lead.