AI chip demand triggers new global supply chain crisis: HBM capacity crowds out traditional memory, chain reactions in the industry chain emerge.

Introduction

In the fourth quarter of 2025, the global semiconductor supply chain sounded the alarm once again. Unlike the comprehensive chip shortage caused by the pandemic from 2020 to 2022, the root cause of this crisis is more concentrated: the insatiable demand for high-bandwidth memory (HBM) from AI data centers is forcing memory manufacturers such as Samsung, SK Hynix, and Micron to shift a large amount of production capacity from traditional DRAM and NAND Flash to HBM. According to Reuters, inventories of traditional memory chips have plummeted from a healthy high double-digit week level at the end of 2024 to just a few weeks of supply for most of 2025. Prices have begun to rise, some electronic products have seen purchase limits at retail in Japan, and Chinese smartphone manufacturers have issued early warnings of price increases.

This is not only a structural imbalance in the memory industry, but could also evolve into a systemic risk affecting global electronics manufacturing, logistics, and even the macroeconomy. This article will analyze the supply chain crisis driven by AI chip demand from the perspectives of technology, industry chain, competitive landscape, and regional impact.

Background: The Rise of HBM and the "Squeeze" on Traditional Memory

HBM achieves bandwidth far exceeding traditional DRAM through 3D stacking and through-silicon via (TSV) technology, and is a key component for AI training and inference server GPUs (such as NVIDIA H100/B200, AMD MI300X). As the AI arms race intensifies, cloud service providers and AI companies have seen exponential growth in HBM procurement. According to TrendForce estimates, HBM bit demand in 2025 will grow by more than 150% year-over-year, with its share of total DRAM bits rising from less than 2% in 2023 to nearly 10% in 2025.

However, HBM production is not on independent lines. HBM is essentially a stacked package based on advanced DRAM dies, and its core DRAM wafers share the same manufacturing processes (such as 1α, 1β nm) as traditional DRAM. To fulfill HBM orders, manufacturers must allocate more wafer capacity to HBM-dedicated DRAM—which means capacity for standard DDR5, LPDDR5, and even server DRAM is significantly reduced. At the same time, HBM also requires additional advanced packaging capacity (such as TSV, micro bumps, hybrid bonding), further occupying resources in the packaging segment.

In-depth Analysis

Technology Impact: HBM Technology Roadmap and BarriersHBM technology is evolving from HBM3 to HBM3E and HBM4, with stacking layers advancing from 12 to 16, enabling single-stack capacities exceeding 36 GB and bandwidths over 1.6 TB/s. The technical barriers are concentrated in: - DRAM process: Requires the most advanced 1β nm or smaller process nodes to reduce power consumption and increase density. - TSV microfabrication: HBM requires precise through-silicon vias and micro-bump arrays, making yield control extremely demanding. - Hybrid Bonding: A bump-less connection method for higher stacking layers, key to future generations like HBM4.

The progression of these technical routes means that HBM capacity expansion depends not only on DRAM fabs but also on advanced packaging equipment (e.g., ASML’s TSV lithography tools, Applied Materials’ deposition equipment, and KLA’s inspection tools).

Supply Chain Impact: A Chain Reaction from Wafers to End-Products

Upstream: The demand structure for materials such as silicon wafers, photoresists, and specialty gases is shifting. HBM uses thicker stacked silicon wafers with higher quality requirements, but the consumption volume is lower than that of traditional DRAM (since each DRAM die is larger, but total wafer input decreases).

Midstream: Memory manufacturers are converting DRAM fab capacity to HBM, leading to reduced output of standard DRAM wafers. Meanwhile, although NAND Flash is not directly used in HBM, manufacturers may also repurpose some NAND production lines for CMOS image sensors or other logic chips to balance capital expenditures, further tightening flash memory supply-demand dynamics. Among the three major manufacturers, SK Hynix and Samsung lead in HBM, with Micron catching up; but all three face a common challenge: how to meet HBM customer demands while maintaining basic supply for traditional memory customers.

Downstream: The direct victims are PC, smartphone, server OEMs, and industrial equipment manufacturers. DRAM and NAND are core components of nearly all electronic products. Price increases and supply shortages will lead to delayed product launches, reduced configurations, or higher prices. The logistics sector faces order fluctuations, increased emergency airfreight, and abnormal inventory levels in warehousing.

Competitive Landscape: The Big Three Lead, Chinese Manufacturers in a Delicate Position- SK Hynix: Thanks to its deep ties with NVIDIA, it holds about a 50% share in the HBM3E market, but its conventional DRAM business share is declining. Its capital expenditure is almost entirely focused on HBM and advanced packaging. - Samsung: It has the most complete memory-logic-packaging ecosystem and is accelerating HBM3E mass production, while facing the risk of customer churn due to insufficient conventional DRAM capacity. - Micron: It started late in HBM, but its HBM3E volume began to ramp up in 2025. In the short term, conventional DRAM still accounts for a relatively high proportion, so this shortage has the greatest impact on its downstream customers. - Chinese memory manufacturers: ChangXin Memory Technologies (DRAM) and YMTC (NAND) are still catching up with international advanced processes. However, HBM technology is constrained by equipment export controls (e.g., ASML high-end immersion lithography tools, HBM packaging equipment), making it difficult for them to enter this market in the short term. Nonetheless, the shortage of standard DRAM presents an opportunity for Chinese manufacturers to fill market gaps, but they must be wary of further U.S. sanctions.- Short-term profitability for storage manufacturers: HBM prices are high (3-5 times that of equivalent-capacity DDR5) and supply is tight, boosting manufacturer profit margins. However, traditional DRAM production cuts may lead to a trade-off in revenue, so overall revenue changes need to be monitored. - Equipment makers benefit: HBM capacity expansion drives orders for TSV equipment and advanced packaging equipment, benefiting Applied Materials, KLA, ASML (TSV lithography), etc. - Downstream risks: PC, smartphone, and automotive chip distributors face inventory impairment risks, and caution is needed for the possibility that traditional DRAM price rebounds may fall short of expectations after HBM capacity is released in 2026.

Long-Term Outlook: Structural imbalance may reshape the storage industry

Over the next three years, HBM demand will continue to grow rapidly (CAGR of approximately 50%), while traditional DRAM demand growth will be modest (about 15%). Storage manufacturers may gradually divide DRAM capacity into "HBM-dedicated" and "traditional-dedicated" lines, or even build independent HBM fabs. At the same time, the advanced packaging segment will become a new bottleneck, benefiting packaging and testing companies like TSMC, ASE Technology, and Amkor.

By 2028, with the mass production of HBM4 and the maturity of new interconnect technologies such as CXL, traditional DRAM may be partially replaced by HBM or near-memory computing. The supply chain structure will become more differentiated: the manufacturing chains for AI storage and general-purpose storage will gradually separate, and geopolitical factors (such as US-based HBM manufacturing) may drive regionalized layouts.

Industry Chain Analysis: Full industry chain impact map

| Segment | Impact | Beneficiaries | Risk Parties | |---|---|---|---| | Upstream silicon wafers | Increased demand for HBM wafers, but total demand slightly declines due to reduced wafer input | Shin-Etsu Chemical, SUMCO | Ordinary polished wafer suppliers | | Upstream equipment | Surge in orders for TSV, hybrid bonding, and inspection equipment | Applied Materials, KLA, Disco | Slower growth in traditional lithography equipment | | Midstream storage fabs | Capacity shift from traditional DRAM to HBM DRAM, increased back-end packaging | SK Hynix, Samsung, Micron | Chasing companies like CXMT face widening technology gap | | Midstream packaging | HBM advanced packaging capacity is tight | TSMC (CoWoS), ASE Technology, Amkor | Traditional packaging houses | | Downstream electronic products | Standard DRAM/NAND price hikes and shortages, product delays | HBM suppliers | Apple, Lenovo, Xiaomi, automotive Tier 1 | | Logistics | Order volatility, increased emergency shipping | Air freight, courier companies | Stable sea freight volumes decline |

ConclusionThis AI chip demand-induced storage supply chain crisis is essentially a "capacity squeeze" of dedicated high-performance memory on general-purpose memory. It reveals that in the AI era, the industry chain structure is shifting from a "general-purpose platform" to "dedicated tracks," rendering the traditional supply-demand balance model invalid. The core industry judgments are as follows:

1. HBM will long-term squeeze traditional DRAM. Over the next 2-3 years, supply of standard memory will remain tight, and prices will tend to rise rather than fall. 2. Memory manufacturers face a strategic dilemma: continuing to bet on HBM will lose traditional customers, but giving up HBM may mean missing the future. Leading companies are likely to adopt a "dual production line" strategy, but capital expenditure pressure will be enormous. 3. Regional supply chain risks are intensifying: HBM manufacturing is highly concentrated in South Korea and Taiwan, China. Both the US and China are seeking localization, but replacement in the short term is difficult. Geographic concentration has actually increased due to the demand surge. 4. Investors need to distinguish between short-term and long-term: in the short term, bullish on the HBM chain (equipment, packaging and testing), but be wary of price fluctuations after the concentrated release of HBM capacity in 2026-2027. In the long term, the storage industry will enter an era of "AI storage" and "general-purpose storage" running on dual tracks.

This crisis did not happen overnight; it stems from the insatiable demand for bandwidth in AI training and inference. The semiconductor industry needs to rethink capacity allocation: should dedicated AI memory fabs be built? Will traditional memory be forced to accept a "second-tier" status? The answer will be revealed in the next three years.

*Source: Reuters report (2025.12.3)*