TSMC, Samsung, and Intel Foundry Competition: Will AI Chip Customers Diversify Their Orders?

AI chip customers will try to diversify foundry orders, but it will be difficult for them to move away from TSMC at scale in the short term. The reason is that when high-end GPUs, AI ASICs, and high-speed networking chips choose a foundry, customers do not only look at 2nm or 3nm node names. They also look at yield, advanced packaging, HBM coordination, design ecosystem, and delivery certainty. Samsung and Intel both have opportunities to enter the market, but their opportunities are more likely to begin with secondary chips, packaging, regional backup capacity, and future-node pilot production rather than immediately replacing TSMC’s role in high-end main-die production.

Key Takeaways

• AI chip customers will diversify orders, but TSMC remains the main choice in the short term.
• Foundry selection depends not only on nodes, but also on yield, packaging, and ecosystem.
• Samsung’s strengths are GAA, memory, packaging, and turnkey solutions.
• Intel is betting on 18A, 14A, PowerVia, and U.S. manufacturing.
• Order diversification will likely begin with packaging, I/O dies, and secondary chips.
• Investment analysis should focus on customer adoption, production execution, and valuation gaps.

Why AI Chip Customers Want to Diversify Foundry Orders

AI chip customers want to diversify foundry orders mainly to reduce supply chain concentration, geopolitical risk, advanced capacity queues, and cost negotiation risk. High-end GPUs, AI ASICs, CPUs, and networking chips all require advanced process technology and advanced packaging. If customers rely too heavily on a single foundry, any capacity shortage, packaging constraint, node delay, or regional risk can affect chip launches, server deliveries, and cloud computing deployment.

Order diversification is not simply about lowering wafer prices. For Nvidia, AMD, Broadcom, Marvell, or cloud providers’ in-house ASICs, the real priority is supply chain optionality. Customers can evaluate Samsung, Intel, or other packaging supply chains to gain future negotiating leverage and build backup manufacturing capacity in specific regions. As AI data center investment scales rapidly, any single point of supply risk becomes magnified.

However, AI chips are harder to diversify than ordinary chips. Large compute dies are highly sensitive to yield, while HBM, 2.5D packaging, high-speed interconnects, and power integrity all require long-term coordination among foundries, packaging providers, memory suppliers, and design teams. Different foundries’ PDKs, design rules, IP, and EDA flows are not fully interchangeable. Migrating a design is expensive and can also affect validation timelines and product launch schedules.

A more realistic diversification path is usually not “moving flagship GPUs to another foundry overnight.” It is more likely to begin with peripheral chips, I/O dies, mature-node chips, some packaging capacity, regional manufacturing, and future-node pilot runs. Customers will separate products by risk level: the most critical main compute die continues to use the most reliable supplier, while secondary modules are used to test a second source.

Summary: AI chip customers do have reasons to diversify foundry orders, but motivation does not equal immediate large-scale order transfers. Advanced AI chips are not standard components. Foundry selection involves yield, packaging, HBM, validation, delivery cycles, and customer ecosystems. A more reasonable view is that customers will gradually build second-source options while avoiding excessive risk in their most critical products. Samsung and Intel’s opportunities will likely appear first in backup orders, future-node validation, regional capacity, and specific customer projects, rather than fully replacing TSMC in the short term.

Why TSMC Remains the Main Foundry Hub for AI Chips

TSMC remains the main foundry hub for AI chips because it has built a combined advantage across advanced-node mass production, customer ecosystem, capacity scale, advanced packaging, and delivery certainty. AI customers choose TSMC not only because its 3nm and 2nm nodes are leading, but because TSMC can connect advanced logic, CoWoS packaging, the HBM supply chain, and large-scale production schedules.

From a revenue structure perspective, TSMC’s advanced process advantage has already translated into financial results. TSMC disclosed in 2025 that it defines advanced technologies as 7nm and below, and that 3nm accounted for 24% of full-year wafer revenue, showing that 3nm has moved from early adoption to a major revenue pillar. In the second quarter of 2026, TSMC’s revenue rose 36% year over year to NT$1.27 trillion, with market reports attributing the growth mainly to AI application demand.

Beyond advanced process technology, advanced packaging is an even harder part of TSMC’s position to replace. TSMC’s CoWoS-S targets AI and ultra-high-performance computing applications, integrating logic chiplets and HBM on a silicon interposer. For AI accelerators, advanced packaging is not an add-on step. It is a core capability that determines whether GPUs, HBM, I/O, and interposers can be connected efficiently. Reuters also reported that TSMC will add two advanced packaging plants in Chiayi Science Park, showing that packaging capacity has become one of the key supply bottlenecks for AI chips.

TSMC is not free from competitive pressure. Customers worry about regional concentration, pricing power, advanced packaging queues, and capacity priority. But for high-end AI main dies, customers care first about certainty. If a flagship GPU or AI ASIC is delayed, the loss is not limited to one batch of chip revenue. It can also affect cloud provider data center deployment, server OEM delivery, model training schedules, and customer contracts.

Summary: TSMC remains the main advanced foundry center for AI chips in the short term because its moat is not a single node name. It is a combination of production experience, advanced packaging, customer trust, yield ramp-up, and supply chain coordination. AI chip customers will want to diversify risk, but for the most critical high-end compute dies, reliable delivery matters more than short-term price. Unless Samsung and Intel continue proving themselves in advanced-node yield, packaging capacity, design ecosystem, and large-customer production cases, TSMC’s leadership will be difficult to challenge in the near term.

Can Samsung Win AI Customers Through 2nm GAA and Integrated Packaging?

Samsung has a chance to win some AI chip customers, but to challenge TSMC at scale, it still needs to solve yield, customer trust, and ecosystem scale issues. Samsung’s advantages include its early push into GAA, DRAM and HBM resources, advanced packaging capability, and the ability to offer an integrated foundry, memory, and advanced packaging solution. For some cloud providers’ in-house ASICs and regional backup customers, this turnkey model can be attractive.

At Samsung Foundry Forum 2024, Samsung emphasized that the AI era requires high-performance, low-power semiconductors and presented AI/HPC-oriented technology roadmaps such as SF2 and SF2Z. Samsung Foundry’s SF2 entered mass production in 2025. It is a second-generation MBCFET GAA node targeting mobile, HPC, AI, and automotive applications. SF2P targets HPC performance-per-watt needs and is scheduled for mass production by the end of 2026.

Samsung’s most representative AI customer case is Preferred Networks, a Japanese AI company, for which Samsung will provide 2nm GAA and 2.5D I-Cube S packaging. Samsung announced that it would use 2nm GAA and I-Cube S to support Preferred Networks in developing an AI accelerator. Reuters also described the order as Samsung’s first publicly disclosed order for its advanced 2nm technology. This shows Samsung is trying to prove itself through “advanced process + advanced packaging + AI customer.”

Samsung’s competition strategy will not simply copy TSMC. It is more likely to package GAA, HBM, 2.5D/3D packaging, and turnkey services to win customers. For cloud providers, AI ASIC companies, and regionally diversified customers that want to establish a second source, Samsung can be an important alternative. But whether customers are willing to hand their most critical AI chips to Samsung at scale still depends on yield, capacity, delivery record, and ecosystem maturity.

Summary: Samsung is one of the most realistic second advanced foundry options, especially in AI ASICs, regional backup, integrated packaging, and some HPC projects. The Preferred Networks case shows that Samsung has secured a public AI order, but a single order does not equal a broad replacement of TSMC. What Samsung really needs to prove is sustained production capability, yield stability, customer expansion, and delivery certainty. In the short term, Samsung is more likely to win partial orders, backup orders, and specific customer projects rather than quickly taking over large-scale production of high-end GPU main dies.

Can Intel Foundry Enter High-End Customers Through 18A and U.S. Manufacturing?

Intel Foundry has a chance to enter some high-end customers through 18A, PowerVia, RibbonFET, advanced packaging, and U.S. manufacturing, but it is still in the proof stage in the short term. Customers are willing to evaluate Intel because advanced foundry supply needs a second source, and U.S. manufacturing has strategic value. Customers remain cautious because Intel still needs to prove its external customer service capability, yield stability, on-time delivery, and ecosystem maturity.

Intel Foundry positions itself as a systems foundry for the AI era, emphasizing process technology, packaging, IP, EDA, and system-level collaboration. Intel announced that Microsoft selected a chip design planned for production on Intel 18A and disclosed an expected lifetime deal value of more than $15 billion across wafers and advanced packaging. This case is important for Intel Foundry confidence, but it does not mean all cloud providers will immediately shift to Intel.

On the technology roadmap, Intel Foundry uses 18A as a core node. In its 2026 VLSI Symposium update, Intel said Intel 18A entered production in 2025 and combines RibbonFET with PowerVia. 18A-P entered risk production as a performance-enhanced version of the 18A family. For AI chip customers, backside power delivery, GAA, advanced packaging, and U.S. manufacturing together form Intel’s differentiated narrative.

Intel’s biggest opportunity is strategic value. U.S. cloud providers, defense-related customers, and enterprises that need regional manufacturing backup may be willing to pay some cost for supply chain resilience. Compared with Samsung, Intel also has the combined advantage of U.S.-based manufacturing and advanced packaging. But advanced foundry is not a business that can be completed through policy support and roadmaps alone. Customers ultimately need to see tape-out success rate, yield, delivery, cost, and ecosystem support.

If you follow AI supply chain names such as TSM, INTC, NVDA, AMD, ASML, AVGO, and MRVL, you also need to pay attention to actual trading costs alongside share price movements. U.S. stock trading costs often include more than commissions; they may also include platform fees, external institution fees, and trading activity fees. For example, Biya charges $0 commission for U.S. stock trading, while platform fees, external institution fees, and other charges are subject to the U.S. stock trading fee structure and the order page. Service availability depends on the user’s location, identity verification result, platform rules, and applicable laws and regulations.

Summary: Intel is the foundry competitor with the largest potential variable. If 18A, 18A-P, and later 14A execute as planned and win more cloud providers, chip design companies, and advanced packaging customers, AI chip foundry competition could become more diversified. But before real production cases, external customer service capability, cost structure, and yield stability are fully proven, Intel is more of a strategic backup and future option than a near-term primary replacement for TSMC.

Where Will AI Chip Customers Diversify Orders First?

AI chip customers are most likely to diversify orders first in non-core dies, mature-node chips, advanced packaging, regional capacity, and future-node pilot production, rather than immediately moving their most critical GPUs or AI ASICs elsewhere. Main dies require the highest yield and packaging coordination, so customers are usually more conservative. Peripheral chips, I/O, networking chips, chiplets, packaging, and next-node validation are better suited for multi-supplier experimentation.

The easiest place to diversify is mature-node peripheral chips. Power management, control chips, some I/O dies, certain networking components, and specific chiplets can use different processes and do not necessarily depend entirely on the most advanced nodes. Customers can assign these areas to different foundries to reduce supply chain concentration and test different suppliers’ service capabilities.

Advanced packaging is also an important entry point for order diversification. AI accelerator delivery bottlenecks are not only in wafers, but also in CoWoS, I-Cube, Foveros, EMIB, and other packaging capacity. Customers may evaluate different regions and packaging solutions without moving the main compute die. For Samsung and Intel, packaging capability is both a differentiated entry point and a long-term battlefield against TSMC.

Cloud providers’ in-house ASICs are a more likely entry point for Samsung and Intel. Compared with Nvidia’s flagship GPUs, in-house cloud chips are more focused on internal workload optimization and may be more willing to try multi-supplier strategies in second- or third-generation products. Even then, customers will control risk and usually begin with small batches, specific applications, or future-node validation.

Summary: Customers will diversify orders, but they will not easily move high-end main chips. The more realistic path begins with peripheral chips, I/O dies, chiplets, packaging, regional capacity, and future-node pilot production. Samsung and Intel’s opportunities are also more likely to appear in these areas before gradually moving toward core compute dies. For AI chip customers, the real goal is not to completely replace TSMC, but to build more supply chain options without sacrificing delivery certainty.

How Foundry Competition Will Affect the Supply Chain and Investment Expectations

Foundry competition among TSMC, Samsung, and Intel will affect AI chip companies, semiconductor equipment makers, advanced packaging, memory, and investor expectations. If order diversification succeeds, customer supply chain resilience improves, and the valuation narratives for Samsung and Intel strengthen. If diversification progresses slowly, TSMC will continue to maintain stronger pricing power in advanced process technology. The key is not who releases the most aggressive roadmap, but who can turn customer tape-outs into stable production and financial revenue.

For AI chip companies and cloud providers, multi-foundry relationships can increase negotiating power and supply chain resilience. Nvidia, AMD, Broadcom, Marvell, Amazon, Google, Microsoft, and others will continue evaluating supply stability, but core products will still lean toward the most reliable supplier. Cloud providers’ in-house ASICs are more likely to become entry points for Samsung and Intel because they require customization and create stronger incentives to establish a second source.

For the equipment, packaging, and memory chain, foundry competition may increase global capital expenditure. If Samsung and Intel expand advanced nodes, equipment companies such as ASML, Applied Materials, Lam Research, KLA, and Tokyo Electron may benefit. Advanced packaging will also become a competitive focus, and coordination between HBM suppliers and foundries will become more important. However, multi-region expansion can also create duplicated investment, higher costs, and capacity-cycle risk.

For investors, there are five indicators to judge whether foundry competition is truly changing: real customer tape-outs, yield and capacity, advanced packaging expansion, financial execution, and ecosystem development. Roadmaps and market rumors should not be used as standalone trading evidence. When a customer announces adoption of a node, it only means validation or cooperation has begun. Only when it reaches stable production and appears in revenue, gross margin, and orders does it signal a real change in the competitive landscape.

You can use U.S. stock market information to track AI supply chain names such as TSM, INTC, NVDA, AMD, AVGO, MRVL, and ASML, then combine earnings, customer orders, capital expenditure, and valuation to judge expectation gaps. Biya is a global multi-asset trading wallet that supports U.S. stocks, Hong Kong stocks, and crypto trading, while also covering multi-asset market data and trading scenarios. Public market information can support research, but it does not constitute investment advice.

Summary: The outcome of foundry competition will not be an extreme scenario where “TSMC loses all orders” or “Samsung and Intel immediately stage a comeback.” A multi-layered division of labor is more likely: TSMC continues to control production of the most critical high-end main chips, Samsung competes for turnkey AI and some 2nm/GAA customers, and Intel competes for U.S. manufacturing and 18A/14A strategic customers. Investors should focus on real customers, real production, and real financial execution instead of only looking at roadmaps and market rumors.

If you follow AI chip foundry competition, you can divide your research framework into three layers. The first layer is the advanced nodes and yields of TSMC, Samsung, and Intel. The second layer is CoWoS, I-Cube, Foveros, HBM, and the equipment chain. The third layer is earnings guidance, valuation, orders, and trading costs. For users who meet the applicable service conditions, Biya can be used to observe and trade multiple asset classes, including U.S. stocks, Hong Kong stocks, and crypto. Biya charges $0 commission for U.S. stock trading, while platform fees, external institution fees, and other charges are subject to the fee center and the order page. Before trading, users should still confirm order types, fee structures, market volatility, and local regulatory requirements.

FAQ

Why Do AI Chip Customers Want to Diversify Foundry Orders?

AI chip customers want to diversify foundry orders to reduce supply chain concentration, geopolitical risk, capacity queues, and cost negotiation risk. But order diversification cannot be judged only by node names. Yield, advanced packaging, HBM coordination, design migration cost, and delivery stability all matter.

Will TSMC Lose Major AI Customers Such as Nvidia?

TSMC is unlikely to lose its core AI customers at scale in the short term. It still has clear advantages in advanced process technology, CoWoS packaging, yield, and customer ecosystem. However, major customers such as Nvidia will continue evaluating Samsung, Intel, and regional backup options.

Why Is Samsung 2nm GAA Attractive to AI Customers?

Samsung 2nm GAA is attractive because it offers potential performance and power improvements, GAA experience, memory resources, and integrated advanced packaging. Whether it can truly win major customers still depends on yield, capacity, delivery record, customer trust, and ecosystem maturity.

Can Intel 18A Change the Advanced Foundry Landscape?

Intel 18A has the potential to change the advanced foundry landscape, but it needs more external customer production validation. PowerVia, RibbonFET, U.S. manufacturing, and advanced packaging are advantages, while yield, cost, on-time delivery, and customer service are the key tests.

Where Will AI Chip Order Diversification Happen First?

AI chip order diversification will most likely begin with I/O dies, mature-node peripheral chips, advanced packaging, cloud providers’ in-house ASICs, and future-node pilot production. High-end GPU main dies are harder to move at scale in the short term because of yield, packaging, and delivery risks.

How Can Investors Judge Whether Foundry Competition Is Truly Changing?

Investors should look at real customer tape-outs, production progress, yield, capital expenditure, packaging capacity, and financial revenue. Roadmaps and rumors should not be used as standalone trading evidence. When trading, investors should also consider valuation, fee structure, and their own risk tolerance.