Techtalk

Latest

RAMpocalypse Now – The 2026 Memory Crisis in Embedded Systems

DDR & NAND Memory Under Stress Testing: What Matters Now. (Last updated in October 2026)

[Translate to Englisch:]

key points at a glance

  • The memory crisis will hit the embedded world in 2026: DDR3, DDR4, NAND, and eMMC will become scarce as manufacturers shift their production capacity to DDR5 and HBM memory for servers and AI applications.
  • Prices are rising even in existing contracts, not just on the spot market. TrendForce expects increases of +55–60% for DRAM and +33–38% for NAND in Q1 2026.
  • No relief in sight in the short term: Analysts expect the effects to last until 2027. New NAND capacity (Micron, Singapore) won’t be available until H2 2028 at the earliest.
  • Memory from China is an option, but only with extended testing (thermal cycling, aging, margining) and verified compliance.
  • The most effective measures against the memory crisis: Reduce memory demand, secure at least two approved sources, and establish framework or allocation agreements covering 6–18 months.

Storage bottleneck

What was long considered purely a PC or server issue will hit the embedded world with full force in 2026. DRAM and NAND manufacturers are shifting their capacity on a massive scale toward servers, AI, and HBM. For embedded systems, this means one thing above all else: rising prices, longer lead times—and growing risks for ongoing production runs.

The latest market figures clearly underscore this trend: In the first quarter of 2026 alone, contract prices for conventional DRAM rose by about 90 percent. Further price increases—some of them drastic—followed in subsequent quarters as well. This affects not only new designs but, in particular, already-qualified mass-production products, for which redesigns are expensive or hardly feasible.

causes of the storage shortage

  • HBM requires many times more wafer area.
    HBM (High Bandwidth Memory) consists of multiple DRAM dies stacked on top of one another. These dies are larger than those used in standard DRAM. In addition, there are complex process steps such as through-silicon vias (TSVs) and thinning, as well as yield losses during stacking. One gigabyte of HBM therefore requires about three times as much wafer area as one gigabyte of DDR5. Every wafer used in HBM production is thus missing from standard DRAM production multiple times over.
  • The fabs had previously been scaled back.
    In 2022 and 2023, memory manufacturers struggled with full inventories and falling prices. As a result, they reduced their wafer starts, lowered utilization rates, and cut back on investments. Micron, for example, scaled back wafer starts for DRAM and NAND by about 20%, while Kioxia reduced its NAND utilization by 30%. That capacity is missing today. It takes months to ramp up production again, and new fabs take years to build. The capacity that does return is primarily allocated to HBM and server memory.

 

This leaves little for embedded applications. The memory shortage is therefore not a short-term anomaly, but a consequence of structural decisions made by manufacturers.

How are storage prices trending?

  • Memory prices rose in every quarter of 2026. Since the summer, they have been rising more slowly, but have not fallen. TrendForce’s forecasts for contract prices compared to the previous quarter:

  • Q1/2026: DRAM +90–95%, NAND +55–60%
  • Q2/2026: DRAM +58–63%, NAND +70–75%
  • Q3/2026: DRAM +13–18%, NAND +10–15%
  • Q4/2026: DRAM +10–15%, NAND +15–20%
    Source: TrendForce. As of October 2026.

 

No all-clear for embedded series:
The price increases are building on an already significantly elevated baseline. Based on average figures, DRAM prices at the end of 2026 will be roughly four times the level seen at the end of 2025 (our own rough estimate). TrendForce continues to expect a tight DRAM market in 2027.

Embedded memory is likely to be more severely affected:
DDR3, DDR4, and industrial eMMC are not tracked separately in the averages. Due to low volume and low priority among manufacturers, their prices may exceed the market average.

What to do:
Regularly review cost estimates, consolidate demand, and secure supplies through framework agreements. This is particularly important in cases where a component change would trigger new testing and approvals.

DDR3 remains critical in embedded systems, DDR4 and DDR5 overshadowed by the HBM boom

Many established processor platforms that have been in stable mass production for years still require DDR3 or DDR3L. It is precisely this segment that is becoming increasingly scarce, especially at higher memory densities. The reason for this is less a sudden technological shift than a gradual transition: Legacy DRAM is losing manufacturing capacity to DDR5 and HBM. This will not lead to an abrupt disappearance of DDR3, but it will have noticeable effects (rising prices, limited selection, and an increased risk of end-of-life [EOL]). For long-lasting embedded products, this is a serious issue that should be addressed early on.

Who still supplies DDR3 memory?

Samsung and SK hynix sent out EOL notices for 2- and 4-Gbit DDR3 chips about two years ago. The market is shifting toward specialized suppliers such as Nanya and Winbond from Taiwan. But even these sources have their limits: Winbond is already fully booked for 2026 and 2027, and both manufacturers’ expansion plans are primarily focused on DDR4. At the same time, prices are rising; Nanya reported significant price increases for DDR3 and DDR4 contracts in Q1 2026.

Actively manage EOL risk

For product lines using DDR3 memory, this means: consistently tracking PCN and EOL notices, evaluating last-time-buy options in a timely manner, and qualifying at least one secondary source. For new developments, the long-term availability of memory should already be a criterion when selecting a platform.

A similar picture is emerging for DDR4 as well. Production lines are increasingly shifting toward server DRAM and HBM. DDR4 remains technically established but is becoming significantly more volatile from a procurement perspective. For embedded teams, this means that those who allocate resources early and secure them contractually gain a clear advantage. Spot procurement is becoming increasingly expensive, especially when EMC testing, approvals, and validations have already been completed and a last-minute switch is virtually impossible. TrendForce expects prices to continue rising through the end of 2026.

NAND & Flash - Embedded Systems Take a Back Seat

A similar pattern is evident on the NAND side. Enterprise SSDs are being prioritized, while client SSDs, eMMC, and industrial NAND components for embedded boards are increasingly being phased out. In practice, this leads to rising contract prices, longer lead times, and increased risks for mass production, especially in the industrial sector.

Samsung is phasing out MLC NAND. MLC NAND, on which many industrial flash and eMMC solutions are based, is particularly affected. Samsung, previously the largest supplier, announced in March 2025 that it would discontinue its MLC products. The final shipments were scheduled for June 2026. Kioxia, SK hynix, and Micron now manufacture MLC primarily for existing customers. Kioxia has also announced that it will phase out certain older NAND generations by 2028.

How long will the storage crisis last?

The situation is not expected to ease until 2028 at the earliest, and even then only gradually. The major manufacturers have repeatedly pushed back their forecasts for 2026. Micron and Samsung expect shortages to persist until at least 2028.

New fabs will do little to provide quick relief. Micron will begin operations in Idaho in mid-2027, Samsung will launch P5 in Pyeongtaek in 2028, and Micron’s NAND fab in Singapore will start up in H2/2028. It will take several more months before full production volumes are reached. In addition, a large portion of the new capacity will be allocated to HBM and server memory. DDR3, DDR4, MLC-NAND, and eMMC will see little benefit from this.

Recommendation: Plan 24–30 months in advance. For embedded production runs, this means: Forecast demand for this period, align framework agreements accordingly, and qualify alternatives now.

Option China – Opportunity with conditions

Against this backdrop, Chinese memory manufacturers are coming more into focus. The most relevant players here are CXMT in the DRAM sector and YMTC for NAND, supplemented by suppliers such as Longsys and FORESEE, among others for DDR3. China now holds around 5% of the global DRAM market (2024) – and the trend is rising.

Technically, DDR3L, DDR4L, and NAND components from China are generally available and qualifiable. However, we recommend a significantly increased level of testing, for example through thermal cycling, aging tests, and margining. At the same time, market reports show that CXMT could focus more strongly on DDR5 in the future. For long-life designs, this increases the roadmap and EOL risk, which is why multi-source strategies are becoming imperative.

On the NAND side, YMTC delivers modern 3D NAND generations on par with established manufacturers. Nevertheless, compliance and export requirements must be carefully examined depending on the industry. In short, Chinese memory is an option—if qualification, compliance, and alternatives are professionally set up. For regulated applications such as medical technology, redundant strategies are indispensable.

What embedded project teams can do now

A key strategy is to actively reduce memory requirements. Lean images, targeted logs, and clean update paths ensure that less RAM and flash capacity are needed. This expands the selection of available components and reduces dependence on the market.

Embedded distributions such as GELin help keep the software footprint small—without unnecessary bloat. At the same time, memory options should be expanded and qualified early on. For DDR3 and DDR4, having at least two approved sources is advisable; for NAND, eMMC, and SSD, support for multiple types should be planned into the design from the outset.

Another important point is contracts. Framework and allocation agreements spanning six to 18 months with clear PCN and call-off provisions currently offer the most stable path through the market. On the flash side, it’s also worth consciously managing the product lifecycle.

Continuously monitor the lifecycle and EOL. Product discontinuations are currently happening more quickly and often come with short last-time-buy deadlines. Those who systematically track PCN and EOL notifications for all memory chips in the bill of materials (BOM) gain time to make decisions. It makes sense to regularly compare the BOM with availability, lifecycle status, and price trends.

Plan redesigns early. For platforms using DDR3 or MLC NAND, a redesign is often only a matter of time. Planning early allows for a smooth implementation: with sufficient time for testing, certification, and customer approvals. A redesign carried out under pressure following a product discontinuation is usually more expensive and risky. Last-time-buys can bridge the gap until the new design is ready, but they are not a substitute for it.

Besprechung einer Broschüre bei Ginzinger

How Ginzinger supports you

Ginzinger electronic systems supports its customers and partners throughout the entire supply chain. From BOM stress tests for DDR3, DDR4, and NAND, to design and update strategies with power-fail-safe memory layouts, to procurement and life cycle management. The goal is to keep series production running smoothly, even when the market remains volatile. With the RAMpocalypse BOM check, you will receive a prioritized list of measures for your current design in 45 minutes. Together, we will secure your memory roadmap for 2026/27.

Jetzt BoM-Check anfragen

FAQ

Older standards with low production volumes are the hardest hit. These include DDR3 and DDR3L—especially at higher memory densities—DDR4, as well as MLC NAND and eMMC. Major manufacturers are scaling back or discontinuing these products as they shift their capacity to HBM and server memory. For embedded products with long lifecycles, this means rising prices, longer lead times, and an increased risk of end-of-life (EOL).

Generally not. DDR3, DDR4, and DDR5 are neither pin-compatible nor electrically compatible. The memory controller in the processor determines which standard is supported. If it only supports DDR3, an upgrade isn’t possible without a new processor. Even if the processor supports multiple standards, a new layout, software adjustments, and new testing and approvals are required. That’s why it’s worth planning a redesign early on, rather than waiting until a product is discontinued to react.

That depends on the specific case. A lifetime buy can ensure the availability of a product line until the end of its life cycle or until a redesign. On the other hand, there are costs associated with tied-up capital, inventory, and the risk of inaccurate demand forecasts. In the case of long-term storage, storage conditions and the workability of the components must also be ensured. Furthermore, those who buy now are purchasing at high price levels. A lifetime buy therefore makes sense primarily when a redesign is not economically feasible or when a time gap needs to be bridged. Making an informed decision requires weighing demand, costs, and the effort involved in a redesign.

Long-term care with Ginzinger

Completely carefree

Sources

(As of october 2026)