Why Are Memory Manufacturers Shifting Capacity to HBM? The Impact on Regular DRAM and NAND Prices

Memory manufacturers are shifting capacity to HBM mainly because AI servers are driving rapid demand for high-bandwidth memory. HBM offers stronger pricing, margins, customer lock-in, and long-term order visibility than regular DRAM. The issue is that HBM is not produced out of thin air. It consumes advanced DRAM wafers, TSV, stacking, advanced packaging, testing, and engineering resources. As a result, supply for regular DDR5, DDR4, LPDDR, and server DRAM becomes tighter. NAND is not directly displaced by HBM production, but it can still rise in price as AI data centers increase demand for enterprise SSDs.

Key Takeaways

• HBM is critical memory for AI accelerators, not a simple upgrade of regular DRAM.
• HBM consumes advanced DRAM wafers, packaging, and testing resources, tightening conventional supply.
• Memory makers prioritize HBM because margins, customer lock-in, and order visibility are stronger.
• Regular DRAM price increases come from both HBM crowding-out and stronger server demand.
• NAND is not directly replaced by HBM, but AI server eSSD demand can push NAND higher.
• Investment analysis should track HBM qualification, ASPs, inventory, capex, and end demand.

What Is HBM? Why Do AI Servers Need High-Bandwidth Memory?

HBM is a high-bandwidth memory technology that vertically stacks multiple DRAM dies and places them close to GPUs, AI ASICs, or other accelerators through TSV and advanced packaging. Its value is not in replacing all regular DRAM, but in solving the “memory bandwidth wall” in AI computing: compute chips may be extremely fast, but if data cannot be delivered to the compute units quickly enough, GPUs or ASICs sit idle and overall efficiency falls.

Structurally, HBM is very different from regular DDR5, LPDDR, and GDDR. Rambus’s explanation of HBM notes that HBM is a 3D memory module formed by stacked DRAM layers connected through TSVs. In a 2.5D architecture, HBM sits next to GPUs or CPUs on a silicon interposer, enabling shorter-distance and higher-bandwidth data transfer. Lam Research’s explanation of HBM also emphasizes that HBM provides faster data access and lower energy use than traditional memory.

AI servers need HBM because large-model training, inference, long-context workloads, multimodal models, and AI ASICs all require extremely high data throughput. Model parameters, activations, and intermediate computation results must move rapidly between the chip and memory. Ordinary system memory alone cannot meet this level of bandwidth requirement. Samsung HBM also highlights high throughput, TSV stacking, and AI/HPC workloads as core selling points, while positioning HBM4 for next-generation AI infrastructure.

HBM also carries deeper strategic meaning for memory suppliers. Regular DRAM is closer to commodity memory, with prices fluctuating quickly based on supply, demand, and inventory. HBM requires a longer qualification cycle, more complex manufacturing, and closer customer collaboration. For Samsung, SK hynix, and Micron, HBM is not just a new product. It is an opportunity to move from “selling standardized bits” to “binding with AI platforms.” Micron’s financial materials estimate that the HBM total addressable market will grow sharply from a 2024 level of $16 billion and exceed $100 billion by 2030, which explains why memory makers are willing to prioritize capacity, R&D, and customer qualification resources for HBM.

Summary: HBM’s core value comes from “high bandwidth + proximity to compute + high-end customer binding.” It is not simply an enhanced version of regular DDR5, but a key component in the packaging architecture of AI accelerators. To understand HBM, you need to place it inside the full AI server: the GPU or ASIC performs computation, HBM delivers data quickly to the compute units, DDR5 provides system memory, and NAND/eSSD stores data. Because HBM sits at the bottleneck of AI compute efficiency, memory makers are prioritizing capacity, R&D, and customer qualification resources for it.

Why Are Memory Manufacturers Shifting Capacity to HBM?

Samsung, SK hynix, and Micron are shifting resources to HBM because AI customer demand is strong, product margins are high, supply is structurally tight, and order visibility is better. HBM is deeply tied to the roadmaps of AI chipmakers and cloud providers such as NVIDIA, AMD, Google, Amazon, and Microsoft. Once a supplier passes customer qualification, the supply relationship is often more stable. Compared with regular DRAM, HBM is more like a high-end strategic product than a standardized memory chip.

Memory manufacturers prioritize HBM for six main reasons:

• AI server demand is clear, and cloud providers and chipmakers lock in supply in advance.
• HBM has higher unit value and usually better margins than regular DRAM.
• Customer qualification cycles are long, creating stronger customer stickiness once approved.
• HBM3E and HBM4 iterate quickly, and technology leadership can translate into pricing power.
• High-end customers are more willing to sign long-term agreements, improving revenue visibility.
• HBM drives upgrades in advanced process nodes, packaging, and testing capabilities, creating barriers.

However, HBM capacity cannot be expanded by simply flipping a switch. It requires advanced DRAM dies, TSV, stacking, packaging, testing, substrates, and yield management. Even if wafer output is sufficient, limitations in packaging, testing, or yield can still restrict final HBM shipments. This is why HBM supply constraints are not only a wafer issue, but also an advanced-packaging and system-qualification issue.

TrendForce’s estimate of HBM supply shows that HBM wafer input as a share of total DRAM wafer input among the three major suppliers is expected to reach about 18%, 22%, and 30% by the end of 2025, 2026, and 2027, respectively. However, HBM bit supply as a share of total DRAM bit supply is expected to be only about 8%, 9%, and 13% during the same period. This is critical: HBM consumes a large amount of wafer and engineering resources, but it contributes a much smaller share of total bit supply, meaning it significantly changes the supply-demand structure of regular DRAM.

The reluctance to expand regular DRAM aggressively also reflects the cyclical history of the memory industry. In past DRAM/NAND downturns, overexpansion led to price collapses and high inventories. After experiencing multiple cycles, original manufacturers are more inclined to allocate capex toward high-margin products, process upgrades, and high-end customers, rather than expanding standard DRAM indiscriminately. SK hynix’s 2026 market outlook also identifies HBM3E, HBM4, and server DDR5 as key pillars of the DRAM market in 2026.

Summary: Memory manufacturers are shifting capacity to HBM because of a combination of business returns, technology barriers, and customer structure. HBM demand comes from AI chipmakers and cloud providers. It offers higher pricing, deeper qualification, and more stable orders. Regular DRAM is still large in volume, but it is more exposed to cyclical price competition. The problem is that HBM consumes advanced DRAM wafers, packaging, and testing resources, reducing the supply available for regular DRAM. When analyzing the memory industry, you should not look only at HBM revenue growth. You also need to watch its crowding-out effect on DDR5, LPDDR, PC DRAM, and traditional server DRAM.

How Does the HBM Capacity Shift Affect Regular DRAM Prices?

The shift of capacity toward HBM can push regular DRAM prices higher because advanced process wafers, equipment, packaging, and engineering resources are being absorbed by HBM, server DDR5, GDDR7, and other high-end products. Available DRAM capacity for PCs, smartphones, consumer electronics, and some traditional servers becomes smaller. Even if end demand is not particularly strong, prices may still rise due to supply constraints. In other words, regular DRAM price increases are not only about demand recovery, but also about supply reallocation caused by high-end AI demand.

HBM crowds out regular DRAM through three main paths. First, HBM itself uses advanced DRAM dies and cannot be fully separated from DRAM wafer capacity. Second, HBM’s packaging, testing, and yield requirements absorb engineering and equipment resources. Third, when capacity is tight, suppliers prioritize higher-priced and more certain HBM and server customers, reducing allocation to PCs, consumer devices, and some mobile products.

TrendForce’s January 2026 survey shows that DRAM suppliers continue to shift advanced process nodes and new capacity toward server and HBM products, significantly restricting supply to other markets. The firm expected conventional DRAM contract prices to rise about 55%–60% quarter over quarter in 1Q26. By 2Q26, TrendForce’s memory price survey expected conventional DRAM contract prices to rise another 58%–63% quarter over quarter, suggesting that price increases were no longer just a single-quarter inventory recovery, but a supply-structure shift.

Contract prices and spot prices may not move in sync. Contract prices mainly reflect procurement cycles of large customers, OEMs, and cloud providers. Spot prices are more sensitive and are easily affected by channel inventory, short-term shortages, and module-maker expectations. Server and HBM-related prices often move first, then the effects spread to PC memory modules, consumer electronics, and low-end devices. When you see memory prices rising at the end-user level, several stages may have already happened behind the scenes: cloud providers locked in supply, original manufacturers tightened allocation, and channels began rebuilding inventory.

Summary: Regular DRAM price increases are not caused by a single factor. They result from “high-end AI demand + HBM crowding-out + server DDR5 growth + cautious supplier expansion + customers locking in supply early.” HBM consumes a large amount of advanced resources but does not add total bit supply in the same proportion, making regular DRAM supply tighter. Even when PC and smartphone demand is weak, prices can still rise because supply is being redirected toward higher-margin products. To judge DRAM price trends, you need to track HBM wafer-input share, server orders, regular DRAM inventory, and the gap between spot and contract prices.

How Does the HBM Capacity Shift Affect NAND and SSD Prices?

HBM does not directly use NAND wafer lines because DRAM and NAND are different product lines. However, HBM and AI server demand can indirectly push NAND and SSD prices higher. AI data centers need not only HBM and DDR5, but also enterprise SSDs, QLC SSDs, training-data storage, inference logs, model snapshots, and data lake capacity. When cloud providers increase AI infrastructure spending, high-end NAND demand tightens as well.

You can understand the roles of HBM, DRAM, and NAND in AI servers as follows:

NAND is not a substitute for HBM. HBM handles high-speed data movement during computation, while NAND handles long-term or semi-long-term storage. Before AI model training, raw datasets, cleaned data, and labeled data must be stored. After training, model versions, checkpoints, and evaluation datasets must be stored. During inference, systems also generate logs, vectorized data, user feedback, and audit records. All of these require enterprise SSDs or other storage systems.

AI data centers make enterprise SSDs a key variable for NAND pricing. When cloud providers procure high-capacity eSSDs, original manufacturers prioritize higher-quality NAND resources for enterprise products. As a result, consumer SSDs, smartphone storage, and PC SSDs can rise in price or see fewer promotions even when consumer demand is not strong. TrendForce’s 2Q26 data noted that more NAND capacity was being allocated to enterprise SSDs, while consumer applications were being reduced due to cost pressure. The firm expected overall NAND Flash contract prices to rise 70%–75% quarter over quarter.

NAND price transmission is more complex than DRAM. DRAM is more directly affected by HBM and server memory, while NAND depends on enterprise SSDs, client SSDs, smartphone storage, inventory, and supplier production discipline. If AI data center eSSD demand remains strong while consumer demand also recovers, NAND price increases will become more visible. If end demand weakens, price increases may remain more concentrated in enterprise SSDs and high-capacity QLC products.

Summary: HBM does not directly take NAND production lines, but AI data centers tighten high-end resources across the entire memory and storage supply chain. HBM solves AI compute bandwidth, DDR5 supports server system memory, and NAND/eSSD handles model, data, and log storage. The real driver of NAND price increases is demand for enterprise SSDs and high-capacity QLC SSDs, not HBM itself. To judge NAND prices, you should watch cloud provider eSSD procurement, NAND supplier inventories, consumer SSD channel pricing, and AI data center capex—not just HBM capacity.

Which End Products Are Affected by Rising Regular DRAM and NAND Prices?

Rising regular DRAM and NAND prices gradually pass through to servers, AI data centers, PCs, smartphones, SSDs, memory modules, cloud services, and some consumer electronics. The speed of transmission depends on supplier inventory, contract cycles, channel spot markets, end demand, and brand bargaining power. Servers and AI infrastructure usually feel the cost pressure first, while consumer products may absorb it through price increases, lower specifications, delayed promotions, or fewer low-cost SKUs.

The main transmission paths can be divided into six categories:

• AI servers: HBM, DDR5, and eSSDs account for a higher share of BOM cost, increasing system cost.
• General-purpose servers: Higher server DDR5 and eSSD prices raise cloud provider and enterprise procurement costs.
• PCs and memory modules: Once OEM allocation tightens, channel modules and DIY memory prices rise.
• Smartphones: Higher LPDDR and NAND costs may affect storage configurations and retail prices.
• Consumer SSDs: Prioritized enterprise SSD allocation reduces supply for consumer SSDs.
• Cloud services and enterprise IT: Higher hardware costs may affect rental, storage, and AI inference budgets.

Reuters’ reporting on the global memory shortage noted that memory shortages driven by the AI boom could delay AI projects, that SK hynix expected shortages to last until the second half of 2027, and that smartphone makers warned rising memory costs could create pricing pressure. This shows that memory price increases do not stop at chip suppliers’ financial statements. They affect downstream devices and services.

Cloud providers locking in long-term supply also changes the bargaining power of smaller customers. Large CSPs can sign long-term agreements, prepay, or secure capacity, while smaller server vendors, PC OEMs, module makers, and consumer electronics brands often have to accept higher prices or lower allocation. Reuters’ report on CXMT and Tencent mentioned that Tencent signed a multi-year server DRAM supply agreement with CXMT, reflecting that large internet companies are more willing to secure long-term supply during a memory shortage.

End consumers may feel the impact with a delay. You may not immediately see all smartphones, laptops, and SSDs rise in price, but you may see higher prices for premium configurations, fewer low-cost promotions, higher upgrade costs for storage capacity, or specification cuts in low-end products. For enterprises, server procurement, cloud storage, AI inference, and data analytics costs may also rise, especially for workloads that require large amounts of memory and SSD capacity.

Summary: Memory price increases do not remain limited to HBM or DRAM chips. They move through servers, cloud services, end devices, and enterprise IT budgets layer by layer. AI servers are the first to absorb higher costs for high-end memory and eSSDs, followed by general-purpose servers, PCs, smartphones, SSDs, and consumer electronics. Weak end demand does not necessarily prevent price increases because supply allocation is shifting toward HBM, server DRAM, and enterprise SSDs. To observe the transmission effect, you should track inventory cycles, contract prices, spot prices, brand pricing, and product-configuration changes.

HBM Capacity Shift from an Industry and Investment Perspective: Opportunities, Risks, and Key Metrics

If you look at HBM capacity shift from an industry or investment perspective, the key is not only “who can produce HBM.” You need to track HBM customer qualification, wafer-input share, advanced packaging capacity, regular DRAM supply gaps, NAND/eSSD prices, and long-term supply agreements. HBM may extend the memory upcycle, but it can also amplify industry cyclicality and valuation volatility. Price increases should not be treated as guaranteed returns.

The beneficiaries are not limited to memory manufacturers. SK hynix, Samsung, and Micron are the most direct beneficiaries of HBM and server DRAM. Advanced packaging, TSV, testing equipment, substrates, and packaging materials may also benefit. AI GPUs, AI ASICs, cloud providers, and the enterprise SSD supply chain together form the broader AI infrastructure ecosystem. In contrast, PCs, smartphones, consumer SSDs, and low-end electronics may face cost pressure.

The risks should not be ignored. First, HBM qualification and yield are difficult; technical leadership does not automatically translate into large-scale shipments. Second, if AI capex slows, high HBM expectations may be repriced. Third, excessive increases in regular DRAM and NAND prices may suppress PC, smartphone, and consumer electronics demand. Fourth, if new capacity comes online in a concentrated wave, the memory industry may still return to a downcycle. Fifth, antitrust issues, export controls, and supply-chain geopolitics can affect customers and valuations.

If you follow memory, AI semiconductor, and advanced-packaging companies in U.S. or Hong Kong markets, you also need to consider actual trading costs in addition to share-price volatility. U.S. stock trading costs often include not only commissions, but also platform fees, external agency fees, transaction activity fees, and other charges. Through Biya, you can track Micron, Western Digital, Seagate, SanDisk, AI chips, cloud computing, and semiconductor-equipment-related names while placing fee structure and company research into the same decision framework. Biya charges $0 commission for U.S. stock trading, while platform fees, external agency fees, and other charges are subject to the U.S. stock trading fees and the order page. Public market information, trading rules, and fee structures do not constitute investment advice. Service availability depends on the user’s location, identity verification result, platform rules, and applicable laws and regulations.

Summary: The industry logic behind the HBM capacity shift is clear: AI servers drive demand for high-end memory, memory suppliers allocate resources to higher-margin and more stable HBM and server DRAM, and regular DRAM and NAND prices rise because supply is reallocated. But from an investment perspective, price increases alone are not enough. You need to watch customer qualification, capacity execution, gross margin, inventory, AI capex, and valuation levels. HBM may extend the memory upcycle, but overheated expectations can also lead to sharp pullbacks. A more balanced approach is to view HBM within the broader AI infrastructure and memory-cycle framework.

If you want to keep tracking HBM capacity shifts, DRAM/NAND prices, and the memory supply chain, you can place Micron, SK hynix, Samsung Electronics, Western Digital, Seagate, SanDisk, AI chip companies, advanced packaging firms, semiconductor equipment names, and cloud computing ETFs into the same watchlist. You can use U.S. stock information search to check basic information on related stocks, then cross-check HBM revenue, DRAM ASPs, NAND/eSSD prices, inventory, and cloud customer orders in earnings reports. If the relevant services are available in your region, you can also download the app to learn more about multi-asset market data, fee structures, and trading rules. Keep in mind that rising memory prices and strong AI demand do not mean related stocks will necessarily rise. Before any trade, you should confirm order type, fee details, market volatility, and local regulatory requirements.

FAQ

Why does shifting capacity to HBM push up regular DRAM prices?

Shifting capacity to HBM pushes up regular DRAM prices because HBM consumes advanced DRAM wafers, packaging, testing, and engineering resources. Once less capacity is available for regular DRAM such as DDR5, DDR4, and LPDDR, prices may rise even if PC or smartphone demand is not strong. Server demand and cloud provider supply lock-ins can further amplify pricing pressure.

Does HBM directly cause NAND Flash shortages?

HBM does not directly cause NAND Flash shortages because HBM belongs to the DRAM product line, while NAND is a different type of memory chip. However, AI data centers also drive demand for enterprise SSDs, high-capacity QLC NAND, training-data storage, and log storage. As a result, NAND prices can be indirectly pushed higher, especially in the enterprise SSD supply chain.

Will ordinary consumers be affected when buying memory or SSDs?

Ordinary consumers may be indirectly affected when buying memory and SSDs. When memory manufacturers prioritize HBM, server DRAM, and enterprise SSDs, consumer memory modules, PC SSDs, and smartphone storage may see price increases, fewer promotions, or higher upgrade costs. The scale of the impact depends on channel inventory, end demand, and brand bargaining power.

Will HBM capacity expansion quickly bring DRAM prices down?

HBM capacity expansion does not necessarily bring DRAM prices down quickly. HBM expansion cycles are long, and new resources often continue to prioritize AI customers. HBM itself also consumes large amounts of advanced DRAM resources. Price pressure may ease only when regular DRAM bit supply recovers, inventories rise, and end demand slows at the same time.

How should ordinary investors evaluate risks in HBM-related stocks?

Ordinary investors should evaluate HBM-related stock risks by looking at customer qualification, yield, capacity execution, gross margins, inventory, AI capex, valuation, and competition. Strong HBM demand does not mean related stocks are free from cyclical risk. If prices already reflect high growth expectations, any disappointment in orders, technology roadmap, or pricing can trigger a pullback.