Supply and Demand: Why the Memory Market Will Not Ease on Its Own
The price development of the first half of 2026 is no coincidence. It is the result of several factors acting simultaneously on the supply and demand side – and they cannot be resolved in the short term.
On the demand side, AI is the most obvious driver, but by no means the only one. Traditional enterprise workloads remain high in volume and continue to grow. Memory demand is therefore rising not only at the top of the market, but across the entire breadth of data center infrastructure.
On the supply side, what matters is not nominal bit production but effectively available, qualified product. The single most important factor behind this is the wafer penalty of HBM production – complemented by four structural factors that act independently of it.
Context
Purpose and context of the paper DRAM and Flash Prices 2026 – Market Assessment
This paper is aimed at system integrators, IT solution providers, industrial enterprises, as well as operators of local and regional data centers who seek to understand why prices, availability, and market mechanisms in the memory and infrastructure market are currently developing the way they are.
The objective of this document is not to forecast short-term price movements or to provide concrete procurement recommendations. Instead, we classify the current status quo and explain which structural factors on the supply and demand side are contributing to shortages and rising prices. We deliberately avoid alarmism or oversimplified narratives and focus on transparent causes and interdependencies.
The current market situation is the result of multiple developments that have been building up over years and are now taking effect simultaneously – ranging from changes in manufacturers’ production and capex strategies to new demand drivers and shifting priorities in supply allocation. Individual observations such as empty spot markets or sharply rising prices are therefore less root causes than symptoms of these shifts.
Based on available market data, external sources, and our own market observations, we also provide an objective assessment of possible developments in 2026. This should not be understood as a forecast in the narrow sense, but rather as an interpretation of what appears plausible under the current conditions – and which assumptions can be considered robust from today’s perspective.
Guiding question of the paper: How will DRAM and Flash prices develop in 2026?
The paper consists of three interrelated articles that approach this topic step by step – through a retrospective, an assessment of the current status quo, and our objective outlook for 2026.
Demand Side: AI Is the Accelerator, but Not the Only Driver
The most obvious demand impulse continues to come from the AI environment. Training, inference, large GPU clusters, HBM-centric system architectures and, increasingly, memory-intensive workloads around agentic AI and retrieval-augmented generation are raising the need for high-performance compute and memory infrastructure.
The current situation, however, is not attributable to this demand impulse from the AI environment alone. BCG expects global demand for data center capacity to grow by around 16 percent per year from 2023 to 2028, reaching approximately 130 gigawatts by 2028. According to BCG, this growth is around 33 percent above the momentum of the 2020 to 2023 period. At the same time, traditional enterprise workloads will remain a substantial block of demand in 2028 as well, accounting for around 55 percent of total data center demand (cf. Lee et al., 2025).
This distinction is central to the memory market. GenAI is the fastest-growing driver, but not the only one. File storage, databases, virtualization, transaction systems, cloud migration, backup, analytics and traditional enterprise applications remain high in volume. Memory demand is therefore growing not only at the top of the market, but also across the breadth of data center infrastructure.
Micron describes this development in similar terms. The company expects bit shipments for data center DRAM and data center NAND in calendar year 2026 to be more than twice as high as two years earlier. Micron also sees additional demand not only in traditional accelerator racks, but in CPU racks and storage racks as well, as agentic AI applications require more context memory and higher-performing storage tiers (cf. Micron, 2026a).
What this means for enterprise and data center customers
The current shortage affects more than just HBM or specialized AI components. It increasingly extends into product classes that are relevant to traditional server and storage infrastructures.
Supply Side: What Matters Is Not Nominal Capacity, but Effectively Available Product
On the supply side, the decisive question remains not how many bits the industry produces in total, but which product classes are actually available to specific customer groups.
Our previous analysis already pointed out that the memory industry is prioritizing its capacity more strictly and that freely available volumes for the commodity or spot market are correspondingly limited.
The IDC source linked below supports this view: for 2026, limited bit supply growth of around 16 percent for DRAM and around 17 percent for NAND was expected. At the same time, the shortage was described not as conventional misplanning but as a strategic reallocation of wafer capacity toward HBM, data center DRAM, AI-related products and higher-value memory segments (cf. Jeronimo et al., 2025).
As of mid-2026, current manufacturer statements show that even higher bit growth does not automatically bring relief. For calendar year 2026, Micron expects industry-wide DRAM bit shipments in the low- to mid-20-percent range and NAND bit shipments of around 20 percent. At the same time, the company emphasizes that demand for DRAM and NAND significantly exceeds supply and that the supply-demand situation is likely to remain tight beyond calendar year 2027 (cf. Micron, 2026a).
TrendForce also points to structural supply limits. The DRAM industry has only limited cleanroom capacity. Samsung and SK hynix can expand their production lines only moderately, while Micron's new ID1 fab in the US is not expected to become relevant until 2027. Additional CapEx therefore has only a limited effect on bit supply growth in 2026. For NAND flash, the investment focus is moreover on process upgrades and hybrid bonding rather than on an aggressive expansion of volume capacity; TrendForce expects tight supply throughout 2026 (cf. TrendForce, 2026a).
The single most important factor behind this reallocation is the so-called wafer penalty of HBM production. Due to die stacking, TSV processes and significantly lower yields, HBM requires roughly three times the wafer capacity per bit compared with conventional DRAM.
TrendForce quantifies the consequence precisely: at the end of 2025, 2026 and 2027, HBM is expected to account for approximately 18, 22 and 30 percent of the DRAM wafer input of the three major manufacturers – while delivering only around 8, 9 and 13 percent of bit supply. In 2026, HBM thus ties up a good fifth of the wafers for less than a tenth of the bits; that difference is missing from the market as conventional DRAM. With the larger die sizes of coming HBM generations, this crowding-out effect is likely to intensify further in 2027 (cf. TrendForce, 2026b).
Die stacking refers to layering multiple memory chips on top of one another to form a single component. TSVs (through-silicon vias) are the vertical interconnects that electrically link these layers.
Both increase capacity per component but reduce the yield of functioning chips per wafer – hence the wafer penalty.
Four structural factors that act independently of HBM
- Technology transitions are becoming more capital-intensive. 1b, 1c and 1-gamma DRAM as well as 200+-layer NAND deliver diminishing bit gains per wafer – additional investment translates into less additional supply than in earlier cycles. How much complex products weigh on throughput can be seen at SK hynix: NAND shipments fell by around ten percent in the first quarter, attributed among other things to longer production cycle times (cf. SK hynix, 2026).
- Manufacturers are consolidating their portfolios upward, that is, toward higher-value, higher-margin products: DDR4 production cuts, the discontinuation of MLC NAND and the removal of individual densities such as 96GB DDR5 RDIMM are shrinking the deliverable range from the bottom up.
- Investment discipline remains high despite record margins. For 2026, TrendForce expects only around 14 percent more DRAM CapEx and roughly 5 percent more NAND CapEx – the lesson from the 2023 downturn is preventing the supply glut that followed price peaks in earlier cycles (cf. TrendForce, 2025).
- New capacity arrives late and is already committed. SK hynix is directing the bulk of its rising CapEx into infrastructure (Yongin cluster, M15X) that will not take effect until 2027; Samsung is converting existing lines to the 1c process for HBM4 rather than adding volume. The fact that Samsung achieved its record memory division result in the first quarter explicitly despite limited supply availability names the limiting factor unmistakably: it is supply, not demand (cf. Samsung, 2026a).
Is CXL a capacity drain?
CXL is occasionally cited in this context as another capacity drain alongside HBM. That is a misconception: CXL is an interconnect technology, not a memory type. CXL memory expanders consist of standard DRAM components plus a controller – with no wafer penalty and no stacking-related yield losses.
CXL therefore acts on the demand side, not the supply side; Micron itself classifies the technology as net positive for DRAM bit demand growth (cf. Micron, 2023). For the open market, the topic is barely relevant in 2026 anyway, as available modules are allocated to hyperscalers almost entirely through bilateral contracts.
In the medium term, CXL could even provide slight relief to the market: memory pooling reduces overprovisioning, and expanders can put older DRAM generations to continued use.
More investment does not automatically mean more freely available product in the short term. New fabs, cleanroom space, packaging capacity and process conversions take effect with a considerable time lag. In addition, additional capacity flows preferentially into strategically prioritized product lines – not necessarily into conventional server modules or freely tradable NAND components.
Flash and Enterprise SSDs: The Shortage Reaches Deeper Into the Storage Market
While the public market picture long focused primarily on DRAM and HBM, the flash and enterprise SSD market has also tightened considerably as of mid-2026. Back in November 2025, Heise had already pointed out that flash memory for 2026 was in part almost sold out, while DRAM prices were rising sharply and HDDs showed long lead times in some cases. The report also described how the price of certain 1 Tbit TLC NAND components had more than doubled since 2025 (cf. Labs, 2025).
This development is particularly relevant for data center and enterprise customers. Flash is no longer merely an alternative to HDDs or a performance upgrade for selected workloads. In modern data center architectures, flash is increasingly becoming a strategically allocated resource in its own right. AI and HPC infrastructures require not only compute power but also fast data paths, high-performance checkpointing, training data access, vector and object storage, and storage tiers for growing context and inference workloads.
In its third fiscal quarter of 2026, Micron reported revenue of more than USD 5 billion from data center SSDs, more than double the prior quarter. At the same time, the company points out that AI context memory storage and the displacement of HDDs are expanding the addressable SSD market (cf. Micron, 2026b).
Samsung is likewise gearing its outlook explicitly toward enterprise SSDs. The company names PCIe Gen6 eSSD products and KV cache storage demand as strategic priorities for the second half of 2026 (cf. Samsung, 2026a).
The shortage is thus shifting deeper into the storage market. Not only individual flash components, but also qualified enterprise SSD product lines, controller-level platforms and high-capacity data center products are being prioritized more strictly and planned further ahead.