Google May Join Hands with Samsung: Supply Chain Differentiation Signal for Next-Generation TPU "Icefish"

Event Overview

On June 11, 2026, U.S. tech media The Information reported that Google is discussing with Samsung Electronics to outsource part of the I/O Die for its next-generation AI chip "Icefish" (tenth-generation Tensor Processing Unit, TPU) to Samsung for manufacturing, expected to use Samsung's 2nm process. The main compute processor will still be manufactured by TSMC using its 1.4nm process. The chip is currently being co-developed with MediaTek and is expected to enter mass production as early as 2028.

Notably, Google's previous nine generations of TPUs were entirely reliant on TSMC. This shift to Samsung marks the first time it has introduced a second supplier. After the news broke, Samsung Electronics' stock price rose slightly, while TSMC's ADR dipped slightly, reflecting the market's sensitivity to changes in the foundry landscape.

Industry Background

Strategic Position of Google TPU Google TPU is an application-specific integrated circuit (ASIC) designed for AI training and inference, used to support the Gemini large model and for external sales. With the expansion of cloud business and AI services, Google's procurement of TPUs has surged, and demand from external customers (such as enterprises and research institutions) continues to grow.

TSMC Capacity Bottleneck TSMC's advanced process (3nm/2nm/1.4nm) capacity has long been booked by major clients such as NVIDIA, AMD, and Apple. Since 2025, demand for AI chips has exploded, with TSMC's 3nm/5nm capacity utilization exceeding 100%, while 2nm capacity has not yet been fully ramped up, and 1.4nm is still in development. Google, as an emerging major client, faces extended lead times and capacity allocation pressure.

Samsung's Catch-up Opportunity Samsung has invested heavily in advanced processes (3nm GAA, 2nm GAA), but customer adoption has been slow. Securing Google's I/O Die order is an important validation project for Samsung's 2nm process, which will help improve yield and customer confidence.

In-depth Analysis### Technology Impact - Technology Roadmap: Google's choice to outsource the I/O die to Samsung while keeping the compute die at TSMC reflects a "heterogeneous integration" strategy. The I/O die uses a mature 2nm process that does not involve cutting-edge logic scaling, reducing risk; the compute die pursues maximum performance density. - Technology Barriers: Whether Samsung's 2nm GAA process can meet Google's requirements in yield, performance, and power consumption is the key. Google's I/O die needs to be integrated with the main die through advanced packaging (e.g., 3D packaging or silicon bridges), requiring coordination between Samsung's I/O die and TSMC's compute die in interconnect standards (e.g., UCIe) and thermal management. - Advanced Packaging: Google's TPUs have historically used advanced packaging (e.g., CoWoS or InFO). If Samsung manufactures the I/O die, the packaging stage may still be dominated by TSMC, but Samsung also has packaging capabilities (I-Cube, X-Cube). In the future, mixed packaging across foundries could emerge, posing challenges for packaging providers.

Supply Chain Impact - Upstream: Samsung's 2nm process depends on ASML's High-NA EUV lithography machines, which it has already ordered multiple units of, but equipment supply is tight. In terms of materials, Samsung requires stable supplies of photoresists, gases, etc., which may promote the growth of local material companies in Korea. - Midstream: Although TSMC loses some I/O die orders, the compute die orders remain secure and overall demand expands. After securing the order, Samsung's advanced process capacity utilization will increase, but it must bear yield risks. MediaTek, as a design partner, may get involved in advanced manufacturing for its own chips. - Downstream: Google reduces supply chain risk through supplier diversification while enhancing its bargaining power. NVIDIA has previously outsourced some chips to Samsung (4nm/8nm), and Google's follow-up further validates Samsung's viability as a second supplier.### Competitive Landscape - Foundry market: TSMC still holds absolute dominance in AI chip manufacturing, but Samsung begins to penetrate high-value-added orders. Intel Foundry is not involved yet, but may attract similar demand in the future with advanced packaging. Mature process players like UMC and GlobalFoundries are unaffected. - AI chip competition: Google's TPU competes with NVIDIA GPU, AMD MI series, Amazon Trainium, etc. Google, by developing its own chips and selling externally, is shifting from "self-use" to "semi-open". Supply chain diversification helps improve delivery stability and support its market share. - Design partners: MediaTek participates in AI ASIC design for the first time, marking its expansion from mobile chips to AI infrastructure, potentially challenging Broadcom's position in custom chips.

Regional Implications - United States: As a US company, Google's push for a supply chain not dependent on a single region aligns with geopolitical security orientation. Samsung has a factory in Texas, which could become the basis for further cooperation. Apple previously discussed foundry with Samsung, showing US tech giants' concerns about concentration of capacity in Taiwan. - Taiwan: TSMC remains irreplaceable in AI chip manufacturing, but the outflow of I/O Die indicates Taiwan faces competition in low-value-added segments. However, the 1.4nm process for Compute Die brings higher added value, and Taiwan remains core. - South Korea: Samsung winning Google orders is a major boost for South Korea's semiconductor industry, helping solidify its foundry roadmap and enhance the global competitiveness of its ecosystem (design services, IP, equipment). - China: Mainland Chinese fabs (SMIC, Hua Hong) are restricted in advanced nodes and may be excluded from AI chip supply chains in the long term. Google's cooperation with Samsung may accelerate technology regionalization, and China needs to rely more on self-developed alternatives. - Japan/Europe: Japan benefits in materials and equipment; Europe focuses on automotive chips through Infineon, STMicroelectronics, etc. AI chip orders have limited impact on Europe.

Investment Perspective - Samsung Electronics: If successful in mass production, it will boost market confidence in its foundry capabilities, shifting valuation logic from memory to diversification. However, attention should be paid to the cost of 2nm yield ramp-up, which may pressure margins in the short term. - TSMC: Orders are diversified but not harming the core; capital intensity remains high. In the long run, demand for advanced nodes continues, and diversified customers actually strengthen its bargaining power. - MediaTek: Entering the AI ASIC market may enhance growth-based valuation, but it must bear design failure risks. - Packaging/Testing: Cross-foundry integration will drive demand for advanced packaging; ASE, Amkor, etc. may benefit.### Long-Term Outlook (3-10 years) - 3 years (2026-2028): Samsung's 2nm process has passed Google's verification, and it begins to take more I/O Die orders. TSMC's 1.4nm enters mass production, but capacity is fully booked by major customers, so Google and other customers continue to seek a second source. - 5 years (2028-2030): AI chips show a trend of separation into "compute die + memory die + interconnect die", with different dies possibly manufactured by different foundries, making advanced packaging a competitive focus. Investments in self-owned fabs in South Korea, the US, and Europe come to fruition, and the supply chain gradually localizes. - 10 years (2030-2036): The post-Moore era arrives, 3D heterogeneous integration becomes mainstream, and the boundaries between foundries and packaging houses blur. Under geopolitical pressure, the manufacturing model that completely relies on a single region (Taiwan) is deconstructed, forming a multipolar supply chain.

Industry Chain Analysis

| Segment | Upstream (Materials/Equipment) | Midstream (Design/Manufacturing) | Downstream (Application/Packaging) | |---------|-------------------------------|-----------------------------------|--------------------------------------| | Beneficiaries | ASML (Samsung EUV orders), South Korean materials companies (photoresist, gases) | Samsung Foundry (capacity utilization), MediaTek (design capability) | ASE/Amkor (advanced packaging demand), Google (supply chain security) | | Risks | Some Japanese materials companies dependent on TSMC (if orders shift) | TSMC (loss of I/O Die orders), Broadcom (loss of TPU design) | Traditional packaging houses (technology upgrade pressure) | | Technical Points | High-NA EUV supply limited; 2nm GAA requires entirely new materials | Heterogeneous integration design, UCIe standard unification; hybrid bonding process | Cross-foundry packaging interconnection, thermal and test solutions |

Conclusion

This potential cooperation between Google and Samsung is essentially a microcosm of the AI chip supply chain shifting from "single-point dependence" to "multi-source coexistence." Under the combined factors of TSMC's tight capacity, heightened geopolitical risks, and the explosion of AI demand, tech giants have begun to proactively build backup capacity. Samsung's 2nm process has gained endorsement from a key customer, which will accelerate its foundry business catch-up. Meanwhile, although TSMC temporarily cedes some I/O Die orders, the value of compute dies is higher, and customer diversification actually reduces risk.In the next 5-10 years, AI chip manufacturing will not see a "de-TSMC-ization," but it will form a hybrid supply system of "TSMC + Samsung + in-house + regional foundries." Advanced packaging and chip interconnect technologies will become new competitive high grounds, and the decoupling of design and manufacturing will give rise to new business models. For investors, focusing on Samsung Foundry's yield ramp-up, TSMC's capacity allocation, and the AI transformation of design companies like MediaTek will be the key to grasping the next semiconductor cycle.