Will DRAM Crash Again? Why the Memory Cycle Is Changing—but Has Not Disappeared

The short answer: DRAM can fall again, but one market-wide collapse is less automatic #

DRAM remains cyclical. Prices can still fall sharply when customers reduce inventory, end demand weakens or too much new capacity arrives. What has changed is the structure of the cycle.

In the 1990s and 2000s, many suppliers expanded similar commodity capacity into the same PC-led demand wave. When supply outran demand, price declines forced weaker producers out of the market. Today, Samsung, SK hynix and Micron account for approximately 89% of global DRAM revenue, based on Counterpoint Research's first-quarter 2026 figures. Products and demand sources are also more diverse: HBM, server DDR5, mobile LPDDR, PC DRAM, automotive memory and legacy industrial DRAM do not necessarily move together.

The likely outcome is therefore a differentiated cycle, not the end of the memory cycle. HBM and high-capacity server memory can remain tight while PC or mobile DRAM weakens. DDR4 can rise because suppliers retire capacity even when its end market is mature. One DRAM price index can hide several opposing markets.

Why the old “chicken game” became less destructive #

Commodity memory once encouraged every supplier to keep spending. A company that slowed expansion risked losing scale; if everyone expanded, the industry created its own oversupply. The eventual consolidation removed many players and left three global leaders with most of the market.

Counterpoint's Q1 2026 revenue shares were Samsung 38%, SK hynix 29%, Micron 22%, CXMT 8%, Nanya 2% and others around 1%, subject to rounding. Concentration gives leading suppliers more visibility into industry conditions, but it does not create perfect control. They still compete for technology leadership, customer qualification and market share, while CXMT's growth adds another important supply variable.

Suppliers also respond to downturn signals more actively than in earlier cycles. They can adjust wafer starts, slow node transitions, prioritise higher- margin products or manage shipments. That can reduce the duration of extreme loss-making periods. It cannot guarantee a price floor if demand falls faster than supply can be adjusted.

HBM changes the capacity equation #

HBM is made from DRAM, but it consumes more wafer area per delivered bit than conventional products and requires complex stacking, testing and packaging. Moving resources toward HBM can therefore tighten the conventional DRAM market even before a new fab reaches production.

TrendForce estimates that HBM wafer input across the top three suppliers could represent approximately 18% of total DRAM wafer input in 2025, 22% in 2026 and 30% in 2027. The corresponding HBM share of DRAM bit supply is lower—roughly 8%, 9% and 13%—because HBM consumes more wafer capacity per bit.

This is more useful than the broad claim that “30%–40% of all DRAM production has already moved to HBM.” Allocation depends on whether the metric is wafer input, bit output, revenue or capacity under construction.

HBM also creates a second bottleneck: qualification. A supplier needs working DRAM dies, stacking capability, packaging capacity and approval on the target accelerator platform. Wafer capacity cannot instantly become interchangeable qualified HBM supply.

Conventional DRAM can be more profitable than HBM at points in the cycle #

The relationship is not simply “HBM always earns more, so suppliers abandon everything else.” TrendForce reported that sharp conventional-DRAM price increases pushed the per-wafer revenue and profitability of 64 GB DDR5 RDIMM above HBM in early 2026, while HBM annual contracts adjusted more slowly.

Suppliers therefore optimise a changing product mix. If server DDR5 becomes more profitable, it competes for the same cleanroom and engineering resources. If HBM4 contract prices rise, allocation can shift again. This feedback makes the market less like one commodity pool, but it can still transmit shortages from one category to another.

For component buyers, an HBM shortage can matter even when they never purchase HBM. Capacity and capital redirected toward AI products can reduce flexibility for DDR4, LPDDR4X or other long-life parts used in industrial, networking and automotive systems.

Long-term agreements reduce volatility—but move risk elsewhere #

Historically, memory transactions had strong exposure to quarterly contracts and spot prices. In 2026, major customers are committing to multi-year supply agreements with volume and pricing mechanisms.

Micron said in June that 14 of its 16 signed strategic customer agreements had approximately $100 billion of cumulative revenue at contractual minimum prices over their remaining terms, with $22 billion of projected deposits and related financial commitments. These agreements span DRAM—including HBM where appropriate—and NAND across data-centre, consumer, automotive and industrial customers.

Such contracts can give suppliers better demand visibility and customers more supply certainty. They may also establish price bands or floors that reduce the speed at which contract prices collapse.

But risk is not eliminated:

• Demand becomes more concentrated in a small number of hyperscalers and AI

platform companies.

• Contract renegotiation and customer credit become more important.
• Committed capacity can reduce supply available to smaller buyers.
• A large customer's platform delay can affect an entire product generation.

The memory business may gain steadier cash flow while becoming more dependent on fewer purchasing decisions.

Capacity is expanding—just not quickly enough to remove near-term tightness #

It is inaccurate to say the leading suppliers no longer build fabs. Samsung, SK hynix and Micron continue investing in fabs, cleanrooms, advanced nodes and HBM capacity. The relevant point is timing.

New cleanroom space and equipment take years to plan, build and qualify. Node transitions can increase bits per wafer, but advanced DRAM scaling is harder, and HBM's larger die and stacking requirements consume part of those gains. Micron's March 2026 outlook described cleanroom constraints, long construction lead times, higher HBM mix and weaker bit-per-wafer improvement as limits on DRAM supply growth.

TrendForce likewise expects HBM's crowding-out effect on conventional DRAM to intensify into 2027. That supports the near-term market, but forecasts must not be treated as guaranteed supply outcomes. A faster yield ramp, delayed AI platform or weaker macroeconomic demand can change the balance.

What could still cause a serious DRAM downturn? #

1. Hyperscaler investment slows #

AI infrastructure is now a large incremental demand source. If several cloud providers cut capital expenditure together, HBM and high-capacity server DRAM orders could weaken faster than suppliers can reallocate capacity.

2. Customer inventory becomes excessive #

Shortage fears encourage double ordering and early purchases. Once delivery improves, customers can consume inventory rather than place new orders, making apparent demand fall abruptly.

3. New capacity and yield improvements arrive together #

A single new fab may not crash the market. Several expansions, higher yields and node-driven bit growth arriving during weak demand can.

4. AI architecture reduces memory growth #

Compression, lower precision, better caching, model efficiency or changes in training/inference architecture could reduce memory required per unit of compute. Today, models are becoming more memory intensive, but that trend is not a physical law.

5. Geopolitics or export controls split the market #

Restrictions can tighten supply in one region while creating excess inventory in another. A globally balanced headline may conceal severe local divergence.

6. Legacy demand disappears faster than suppliers exit #

DDR4 and DDR3 shortages can reverse if industrial customers complete redesigns or end products reach end of life. Scarcity does not make a mature technology's underlying demand permanent.

A crash in share prices is not the same as a crash in DRAM prices #

Memory-company shares can fall because investors believe peak earnings are already priced in, even while physical supply remains tight. Conversely, shares can rise before contract prices recover because markets anticipate the next cycle.

Procurement teams should therefore watch physical indicators:

• Contract and spot-price direction by memory type and density.
• Supplier and channel inventory, not only manufacturer share prices.
• Wafer-start and bit-supply guidance.
• HBM, server DDR5 and legacy-capacity allocation.
• Quote validity, lead time and confirmed lot availability.
• Customer redesigns and approved-vendor changes.

Financial-market sentiment is context, not an executable RFQ signal.

What buyers should do in a differentiated cycle #

The worst response is one purchasing policy for every memory product. Separate AI/server, mainstream and legacy parts by risk.

Memory RFQ checklist #

• Full manufacturer part number and orderable suffix.
• DRAM generation, density, organisation, voltage and speed grade.
• Package, temperature grade and applicable qualification.
• Date-code range, single-lot requirement and packing condition.
• Original labels, moisture protection and storage history.
• Approved manufacturer list and whether substitution requires requalification.
• Current quantity, quote validity, delivery window and shipment destination.
• Product lifecycle, PCN/EOL status and expected programme duration.

Current stock, price, date code and lot condition must be confirmed before purchase. A broad statement that “DRAM is tight” does not prove that one exact part number is unavailable—or that a nearby suffix is compatible.

Conclusion: the cycle is segmented, not abolished #

DRAM can crash again. A global demand shock, inventory correction or mistimed capacity expansion could still produce sharp declines. What is less likely is that every DRAM category follows the same path at the same time and for the same reason.

Supplier consolidation, HBM's wafer intensity, long construction lead times, advanced packaging constraints and multi-year customer agreements can reduce the depth of an industry-wide collapse. At the same time, dependence on a few AI customers creates a new concentration risk.

The most realistic 2026–2027 view is a differentiated cycle: HBM and selected server products may remain structurally strong, mainstream consumer memory can soften, and legacy DRAM can experience shortage and decline in sequence. Engineers and buyers should manage the exact technology, suffix and programme lifecycle—not trade on a single memory-market headline.

Primary market references #

• Counterpoint Research: Global DRAM and HBM market share
• TrendForce: HBM wafer input and 2027 supply outlook
• TrendForce: DRAM capacity and legacy-product pressure
• Micron Q3 FY2026 strategic customer agreement presentation
• Micron Q2 FY2026 market outlook