Google is considering splitting production of its next-generation AI accelerator between Taiwan's TSMC and Samsung Electronics, a strategic move driven by TSMC's capacity crisis and reflecting mounting pressure across the industry to diversify away from a single supplier. The arrangement would mark a significant shift in how major tech companies approach semiconductor manufacturing for artificial intelligence chips.

Why Is TSMC Becoming a Bottleneck for AI Chips?

TSMC's dominance in semiconductor manufacturing has created a dangerous single point of failure for the entire AI industry. The Taiwanese foundry controls 73% of the advanced foundry market, far ahead of Samsung Electronics at 7%. However, this concentration has become a critical vulnerability. As demand for cutting-edge AI chips has exploded, TSMC's production capacity has hit a wall. TSMC Chairman Wei Che-chia acknowledged the strain at a shareholders meeting, stating that "customer demand is so high that there is a limit to handling it".

The bottleneck is particularly acute for chips using TSMC's most advanced processes below 3 nanometers, where Nvidia, Apple, and other giants have secured priority access. This leaves companies like Google scrambling to find alternative suppliers for non-critical but essential components. The situation has become urgent enough that major tech firms are now adopting multi-foundry strategies to secure stable, long-term supply chains.

What Is Google's Dual-Foundry Strategy for Its Next AI Chip?

Google has approached Samsung Electronics about producing a key component for its 10th-generation Tensor Processing Unit (TPU), code-named Ice Fish, targeted for mass production in 2028. The component in question is the memory input/output die, or I/O die, a critical connection plate that manages data flow between the computing unit and high-bandwidth memory (HBM), preventing congestion that would slow down AI inference.

Under the preliminary arrangement being discussed, TSMC would handle the core computing chip using its advanced 1.4-nanometer process, while Samsung Electronics would produce the I/O die using its 2-nanometer process. This dual configuration allows Google to work around TSMC's capacity constraints while leveraging Samsung's strengths in memory-related components. From Google's perspective, the calculation is that by allocating I/O die volumes to Samsung, it will gain priority in future negotiations for next-generation HBM volumes to be installed in the 10th-generation TPU.

Samsung already supplies more than 60% of the HBM used in Google's TPUs, making this partnership a natural extension of their existing relationship. The two companies have a history of collaboration, having jointly produced Tensor chips for Pixel phones from 2021 to 2024.

How Is Samsung Positioning Itself as an Alternative Foundry?

Samsung's entry into logic die production, a domain TSMC had effectively monopolized, represents a significant shift in the foundry landscape. Samsung developed this capability during the creation of its sixth-generation HBM (HBM4) memory, where its System LSI division designed the chip and its foundry division produced it using a 4-nanometer process. This internal expertise is now being offered to external customers like Google, proving that Samsung can compete in advanced logic chip manufacturing.

The company's integrated approach is attractive to major tech firms. Samsung offers what analysts call an "integrated device manufacturer turnkey solution," providing foundry fine processing, HBM, and advanced packaging all in one package. This vertical integration means Google can work with a single partner for both memory and logic components, simplifying supply chain management and potentially securing priority access to next-generation HBM volumes.

Samsung

How Are Tech Companies Adapting to Foundry Constraints?

• Multi-Foundry Strategies: Big Tech companies with their own semiconductor design capabilities have begun adopting multi-foundry strategies to secure stable supply chains, moving away from reliance on TSMC alone.
• Alternative Partnerships: Apple has signed a preliminary agreement to produce some iPhone chips at Intel, while Samsung has secured orders from Tesla for autonomous driving chips and from Nvidia for language processing units.
• Supply Chain Resilience: By splitting production across multiple foundries, companies gain negotiating leverage with memory suppliers, earlier access to next-generation components, and protection against single-supplier delays.

Samsung Electronics stands to benefit significantly from this shift. The company's non-memory business, which includes foundry and System LSI divisions, has struggled with weak earnings in recent quarters. However, analysts project a dramatic turnaround if Google's order materializes. Korea Investment & Securities forecasts that Samsung's non-memory division is expected to record operating profit of 154 billion won in the fourth quarter of 2026, after posting losses in previous quarters, eventually reaching annual earnings in the 2 trillion won range by 2027.

The broader implication is that TSMC's stranglehold on advanced chip manufacturing is finally being challenged. Samsung's proven ability to deliver competitive 2-nanometer logic dies and integrated HBM solutions positions it as a credible alternative for companies willing to split their supply chains. This creates a more competitive foundry landscape where specialized AI chip makers have more options for bringing their designs to market without facing years-long delays at a single supplier.

For the AI industry, this diversification is essential. As inference workloads continue to grow and companies push the boundaries of what's possible in fast, efficient AI processing, having multiple manufacturing partners ensures that innovation isn't constrained by a single supplier's capacity. The next few years will reveal whether Samsung can execute at scale and whether other foundries can challenge TSMC's dominance in truly advanced processes.