How Do Foundries Participate in the AI Storage Chain? Mature Nodes, Embedded Memory, and Industry Division of Labor

Foundries participate in the AI storage chain not by directly producing DRAM, NAND, or HBM, but by providing manufacturing platforms for AI accelerators, storage controllers, interface chips, power management ICs, embedded memory, and advanced packaging. If you are following the AI storage industry, you should not look only at memory manufacturers. You also need to understand who manufactures GPUs, ASICs, and SoCs, and who provides CoWoS, silicon interposers, eNVM, BCD, PMICs, and MCUs. The value of foundries lies more in enabling AI computing chips and storage systems to work together efficiently.

Key Takeaways

• Foundries are not HBM manufacturers, but they are deeply involved in AI chip manufacturing.
• The AI storage chain includes memory, logic, packaging, interface chips, and power management.
• Advanced nodes and CoWoS determine how efficiently AI accelerators integrate with HBM.
• Mature nodes still support demand for PMICs, MCUs, interface chips, and eNVM.
• To assess foundry exposure, look at revenue structure, process platforms, and customer quality.

What Role Do Foundries Really Play in the AI Storage Chain?

The role of foundries in the AI storage chain is first and foremost that of a manufacturing platform, not a memory brand. You can think of them as the production layer behind AI chips, controller chips, interface chips, and some embedded memory chips. AI servers need HBM, DRAM, and NAND, but they also need GPUs, AI ASICs, networking chips, PMICs, and high-speed interface chips to work together. Memory manufacturers provide high-bandwidth memory and large-capacity storage, while foundries manufacture many of the logic and system-support chips and connect computing and storage through advanced packaging.

In the semiconductor industry, companies are commonly divided into four types: memory manufacturers, fabless chip designers, foundries, and packaging and testing providers. TSMC positions itself as a customer-focused foundry model, which is different from memory manufacturers such as Samsung, SK hynix, and Micron. The former mainly manufactures chips based on customer designs, while the latter mainly design and produce standardized memory products such as DRAM, NAND, and HBM.

The most common misunderstanding when looking at the AI storage chain is focusing only on HBM. HBM is indeed a key memory product in AI training and inference, but it can only unlock its bandwidth value when packaged together with AI accelerators. Around an AI GPU or AI ASIC, there are also power management, network interconnect, storage control, signal conditioning, and system management chips. Many of these links are not completed by memory manufacturers, but by foundries and specialty-process manufacturers.

Therefore, a foundry’s AI storage exposure is usually “indirect but important.” It may not sell a single HBM chip, but it may manufacture the logic chips, controller chips, and supporting chips around HBM. Especially in AI servers, computing, storage, networking, power, and cooling are all tied to overall system efficiency. If any manufacturing link becomes constrained, server delivery may be affected.

Summary: Foundries are not memory manufacturers in the traditional sense, but they are an important manufacturing foundation in the AI storage chain. When judging whether a foundry benefits from AI storage, you should not simply ask whether it produces DRAM, NAND, or HBM. Instead, you should look at whether it participates in AI accelerators, advanced packaging, storage controllers, PMICs, MCUs, interface chips, and embedded memory platforms. AI storage is not a single-chip competition; it is a coordinated competition among computing chips, memory chips, and system-level packaging. The value of foundries lies precisely in this coordinated manufacturing capability.

How Do Advanced Nodes and Advanced Packaging Connect HBM with AI Accelerators?

Advanced nodes and advanced packaging are the most important entry points through which foundries participate in the AI storage chain. AI accelerators need advanced logic chips for computation, while HBM provides high-bandwidth memory. The two must be combined through 2.5D/3D packaging, silicon interposers, and high-density interconnects. You can think of HBM as a “high-speed memory pool,” advanced-node chips as the “computing core,” and CoWoS-like packaging platforms as the bridge that connects the two efficiently.

In its 2025 annual report, TSMC disclosed that advanced nodes at 7nm and below accounted for a high proportion of wafer revenue, showing that AI and high-performance computing demand is pushing more value toward advanced nodes. At the same time, TSMC’s CoWoS-S is clearly positioned for high-performance applications such as AI and supercomputing, enabling logic chips and HBM stacks to be connected on a silicon interposer.

Why has advanced packaging become one of the bottlenecks in the AI supply chain? The reason is that AI chip competition is no longer only about smaller process nodes. Large-model training requires higher memory bandwidth, lower latency, and better energy efficiency, and a single chip can hardly complete all tasks alone. TrendForce has noted that since 2023, AI demand has driven bottlenecks in 3nm–2nm wafers and 2.5D/3D advanced packaging, with shortages extending from CoWoS to equipment, substrates, and packaging materials.

This also explains why you need to look at both HBM and packaging capacity when analyzing the AI storage chain. Even if memory manufacturers can supply HBM, AI accelerators cannot complete system integration smoothly if advanced packaging capacity is insufficient. TrendForce later estimated that TSMC’s CoWoS capacity is still expanding rapidly, reflecting strong demand from AI chip customers for high-end packaging.

From the perspective of industry division of labor, HBM is led by memory manufacturers, advanced logic is manufactured by foundries, and system integration is supported by advanced packaging platforms. You should not simply define TSMC as a “memory company,” but through advanced nodes and CoWoS, it connects AI accelerators with HBM and becomes a critical manufacturing link in AI storage systems.

Summary: Advanced nodes solve the compute-density problem for AI accelerators, while advanced packaging solves the connection problem between AI accelerators and HBM. Foundries are highly relevant to the AI storage chain not because they directly produce HBM, but because they manufacture computing cores and provide the packaging platforms that connect them with HBM. When analyzing the AI storage industry, you should not look only at HBM shipments. You should also look at CoWoS, silicon interposers, substrates, testing yield, and advanced-node capacity. What limits AI chip delivery is often the combined constraint of computing, memory, and packaging.

Why Are Mature Nodes Still Affected by AI Storage Demand?

Mature nodes do not directly manufacture the most advanced HBM and are usually not the main process for AI GPUs, but they are still affected by AI storage demand. The reason is straightforward: in addition to high-end computing chips, AI servers and edge AI devices need large quantities of PMICs, power management chips, interface chips, MCUs, sensors, analog chips, and control chips. These products care more about reliability, cost, voltage tolerance, low power consumption, and long-term supply. They do not necessarily require 3nm or 5nm, and often sit on mature nodes such as 28nm, 40nm, 55nm, 90nm, or 0.18µm.

Mature nodes are not “outdated capacity.” They are important carriers of specialty processes. For example, BCD processes are suitable for power management and high-voltage driving; eFlash or eNVM is suitable for MCUs, smart cards, and industrial control; RF and analog processes are suitable for communications and sensing. UMC once worked with partners to develop secure embedded flash solutions on a 55nm embedded flash platform, and later disclosed cooperation with SST to improve ESF4 eNVM performance. These technologies are more related to system control and secure storage than to large-capacity memory.

AI demand is usually transmitted to mature nodes in an indirect way. The first layer is AI data center expansion, which drives upgrades in server power, networking, cooling, and storage systems. The second layer is the rise of edge inference in AI PCs, smart cars, robots, and industrial devices. The third layer is growing demand for embedded memory in controllers, sensors, and low-power MCUs. Hua Hong’s Hua Hong Grace emphasizes specialty processes including embedded/standalone NVM, power devices, analog, and power management, which is a typical way mature nodes participate in the AI peripheral chain.

However, you should also avoid overstating the relationship between mature nodes and AI. Not all mature-node capacity directly benefits from AI, and not all eNVM or PMIC orders come from AI servers. To judge the real impact, you need to look at end customers, product platforms, capacity utilization, average selling prices, and financial disclosures—not just whether the company mentions AI.

Summary: Mature nodes participate in the AI storage chain mainly through system-support chips, not by directly manufacturing HBM. AI servers need high-end computing chips, but they also require a large number of mature-process products such as power management, control, interface, analog, and secure storage chips. When analyzing mature-node foundries, you should focus on whether PMICs, MCUs, eNVM, BCD, analog, and power devices enter AI servers, edge AI, smart cars, or industrial-control supply chains. The value of mature nodes lies in reliability, low cost, long-term supply, and specialty processes, rather than simply pursuing the smallest process node.

How Is Embedded Memory eNVM Different from Traditional DRAM and NAND?

Embedded memory, or eNVM, is on-chip memory integrated inside logic chips. It is not the same as standalone DRAM, NAND, or HBM. You can think of HBM as high-bandwidth memory next to an AI accelerator, NAND as large-capacity data storage, DRAM as system working memory, and eNVM as memory used inside MCUs, SoCs, controllers, and edge AI chips to store code, configuration, security keys, firmware, and small amounts of model parameters. Its focus is not maximum capacity, but integration with logic, low power, reliable retention, and process compatibility.

Common forms of eNVM include eFlash, RRAM, MRAM, OTP, and EEPROM. As process nodes advance, traditional embedded flash faces cost and scalability challenges at more advanced nodes, which is why new embedded non-volatile memory technologies such as RRAM and MRAM are gaining attention. TSMC’s introduction of eFlash and RRAM notes that embedded RRAM can be used to replace embedded flash and support smart applications.

The role of eNVM in edge AI is particularly important. Edge devices usually require low power consumption, fast wake-up, local inference, and long-term stable operation. They cannot always rely on the cloud or external storage. eNVM inside MCUs or SoCs can store firmware, device configuration, calibration parameters, security keys, and some lightweight model data. For smart cars, industrial control, wearables, and IoT nodes, this on-chip retention capability is highly valuable.

This is also where specialty-process foundries have opportunities. eNVM is not just about the memory cell itself. It also involves compatibility with logic processes, reliability validation, IP support, automotive or industrial certification, and customer design-in cycles. GlobalFoundries’ Embedded Memory portfolio includes MRAM and emphasizes its use in data and code storage. The company has also launched AutoPro 150 eMRAM for automotive applications, highlighting high-temperature, endurance, and automotive SoC scenarios. Samsung also began commercial shipments of 28nm eMRAM relatively early, showing that embedded MRAM has already entered mass-production applications.

When assessing eNVM-related companies, you should not focus only on the word “memory.” Standalone memory chips are more like standardized products and are heavily affected by pricing cycles, capacity, and bit supply-demand dynamics. eNVM is more like a process-platform capability, usually tied to MCUs, SoCs, smart cards, industrial control, automotive chips, and edge AI designs. Both belong to memory technology, but their business models, demand elasticity, and investment logic are different.

Summary: The biggest difference between eNVM and HBM, DRAM, or NAND is that eNVM is embedded inside logic chips and serves code, configuration, security, and low-power retention, rather than pursuing maximum capacity or the highest external bandwidth. Foundries participate in the embedded memory chain through process platforms such as eFlash, RRAM, and MRAM, especially in MCUs, SoCs, edge AI, industrial control, and automotive chips. When analyzing eNVM, you should view it as a specialty-process capability, not simply as part of the traditional DRAM/NAND cycle of memory manufacturers.

How Do TSMC, GlobalFoundries, UMC, Hua Hong, and Other Foundries Differ in the AI Storage Chain?

Different foundries participate in the AI storage chain in different ways. TSMC is closer to the main chain of advanced nodes, AI accelerators, and advanced packaging. GlobalFoundries, UMC, Hua Hong, and similar companies are more focused on mature nodes, specialty processes, eNVM, PMICs, MCUs, analog chips, and power devices. You should not evaluate all foundries with the same AI storage framework. For advanced foundries, look at customers, nodes, and packaging. For specialty-process foundries, look at platforms, end applications, and capacity utilization.

TSMC’s strength lies in advanced logic and advanced packaging. AI GPUs, AI ASICs, networking chips, and some high-performance computing chips often rely on advanced-node manufacturing and then connect with HBM through platforms such as CoWoS. Its AI storage exposure is closer to “computing chip + HBM packaging integration.”

GlobalFoundries follows a different logic. It emphasizes differentiated processes, FDX, RF, low power, and embedded memory. GF’s eMRAM is more suitable for automotive, IoT, industrial, and edge-intelligence devices. It does not directly compete in HBM, but participates in the upgrade of edge AI and embedded systems.

UMC mainly focuses on mature nodes and specialty platforms. Its ongoing cooperation with partners such as eMemory and SST in embedded flash, PUF-based secure storage, and eNVM shows that its AI storage exposure is more related to controllers, security, MCUs, and low-power systems.

Hua Hong Semiconductor’s positioning is also closer to specialty-process foundry services. DBS describes Hua Hong Semiconductor as having strengths in specialty applications and China’s power-device capacity. These capabilities have indirect links with AI data center power chains, edge AI control chips, eNVM platforms, and analog power management, but they should not be equated with the HBM or DRAM main line.

If you observe these companies from an investment perspective, you need to distinguish between the “main chain” and the “supporting chain.” Advanced-node foundries are more affected by AI accelerators, ASICs, and advanced packaging capacity. Mature-node foundries are more affected by power devices, MCUs, PMICs, eNVM, automotive electronics, and industrial demand. Both may be related to AI storage, but their sources of elasticity and valuation logic are different.

Summary: Foundries occupy very different positions in the AI storage chain. TSMC is more like an advanced manufacturing platform connecting AI computing with HBM, while GlobalFoundries, UMC, Hua Hong, and others are more like mature-node and specialty-process platforms. You should not see “foundry + AI + storage” and immediately draw the same conclusion for every company. Instead, break down the analysis by node, packaging, customers, eNVM, PMICs, MCUs, power devices, and end applications. The valuable question is not whether a company is an “AI storage concept stock,” but where it has verifiable revenue and capacity advantages in the manufacturing chain.

How Can Retail Investors Assess a Foundry’s AI Storage Exposure?

Retail investors should assess a foundry’s AI storage exposure by focusing on verifiable indicators rather than company narratives. The core dimensions include advanced-node revenue share, advanced packaging capacity, eNVM platforms, mature-node product mix, AI customer quality, capacity utilization, capital expenditure, and gross margin trends. You need to distinguish real supply-chain benefits from concept-driven exposure. The former will show up in orders, capacity, revenue structure, and margins; the latter often stays at the level of press releases or industry labels.

You can use the following checklist for an initial screen:

Strong signals usually include advanced packaging expansion, more AI ASIC customers, high CoWoS utilization, eNVM platforms entering mass production, improving PMIC or automotive MCU orders, and clear demand sources disclosed by management in earnings reports. Weak signals include companies only mentioning AI in general terms, revenue still mainly depending on consumer electronics, no improvement in mature-node utilization, or no clear process platform and customer validation.

Trading costs also matter. If you follow semiconductor companies such as TSMC, GlobalFoundries, UMC, Hua Hong, SMIC, Samsung Electronics, Micron, or SK hynix, you should pay attention not only to industry logic, but also to actual trading fees, liquidity, and market risk. The cost of trading U.S. stocks, Hong Kong stocks, or ADRs usually includes more than commission. It may also include platform fees, external institutional fees, trading activity fees, and other charges. Biya charges 0 USD commission for U.S. stock trading, while platform fees, external institutional fees, and other charges are subject to the U.S. stock trading fees and order display. Service availability depends on the user’s location, identity verification results, platform rules, and applicable laws and regulations.

If you need to track semiconductor companies, you can also use U.S. stock information search to check basic information on U.S.-listed stocks, then combine it with company filings, order disclosures, industry reports, and fee structures for a more complete judgment. Industry-chain analysis cannot replace investment decision-making, especially because AI storage, foundries, and advanced packaging all involve cyclicality, capital expenditure pressure, and technology-route risks.

Summary: To judge a foundry’s AI storage exposure, return to verifiable indicators. You should focus on advanced nodes, advanced packaging, eNVM, PMICs, MCUs, customer structure, capacity utilization, and financial performance, rather than relying only on concept labels. The upside for advanced-node foundries comes more from AI accelerators and CoWoS, while mature-node foundries benefit more from system-support chips, edge AI, automotive electronics, and specialty processes. On the trading side, you should also pay attention to fees, liquidity, and volatility risk. Public information can help build an analytical framework, but it does not constitute investment advice.

After understanding the relationship between foundries and the AI storage chain, the next step is to continuously track company earnings, process-platform progress, advanced packaging capacity, HBM supply-demand changes, and mature-node utilization. You can break the chain into several lines: advanced logic, HBM, packaging, eNVM, PMICs, MCUs, and equipment materials, then observe whether related companies are actually seeing improvements in revenue and profit. If relevant services are available in your region, you can use Biya to follow U.S. stocks, Hong Kong stocks, crypto assets, and other multi-asset markets, or download App to manage your watchlist, understand order fees, and assess trading risks. Public market information, trading rules, and fee structures should be evaluated together with platform disclosures and local regulatory requirements. Any semiconductor industry-chain analysis should not be interpreted as a return guarantee.

FAQ

Are Foundries Considered AI Storage Stocks?

Foundries should not be simply classified as AI storage stocks. They usually do not directly produce HBM, DRAM, or NAND, but they may participate in the AI storage chain through AI accelerator manufacturing, CoWoS advanced packaging, eNVM, PMICs, MCUs, and interface chips. The key is to examine revenue sources and process platforms, not just the concept label.

Why Are Mature Nodes Affected by AI Server Demand?

Mature nodes are affected by AI server demand mainly because servers require large numbers of power management, analog, interface, control, and security chips. These chips often care more about reliability, cost, and long-term supply than advanced nodes. The actual impact depends on capacity utilization, customer structure, and product pricing trends.

Is Embedded Memory eNVM the Same as HBM?

Embedded memory eNVM is not the same as HBM. eNVM is on-chip non-volatile memory integrated inside MCUs, SoCs, or controllers, commonly used for code, configuration, security data, and low-power retention. HBM is high-bandwidth memory for AI accelerators, emphasizing bandwidth, capacity, and packaging integration.

How Is Hua Hong Semiconductor Related to the AI Storage Chain?

Hua Hong Semiconductor has a more indirect relationship with the AI storage chain. It is not a DRAM, NAND, or HBM manufacturer, but a specialty-process foundry. Its relevance mainly comes from eNVM, power devices, analog and power management, MCUs, and mature-node supporting chips, which should be verified through earnings reports and customer demand.

How Does TSMC Participate in the AI Storage Chain?

TSMC mainly participates in the AI storage chain through advanced logic manufacturing and advanced packaging. AI GPUs and AI ASICs require advanced-node manufacturing and then connect with HBM through CoWoS, silicon interposers, and related packaging technologies. HBM itself is mainly supplied by memory manufacturers, while TSMC plays more of a system-integration manufacturing role.

What AI Storage Metrics Matter When Investing in Foundry Stocks?

When investing in foundry stocks, key AI storage metrics include advanced-node revenue, CoWoS or 2.5D/3D packaging capacity, eNVM platforms, PMIC and MCU demand, capacity utilization, gross margin, and capital expenditure. These judgments should be based on company filings, order disclosures, and platform fee rules, and do not constitute investment advice.