RAM shortage 2026: what memory costs embedded products

The 2026 RAM shortage is not a PC-gamer problem that will pass on its own: for companies that build electronic products, it changes the bill of materials, the choice of platform and sometimes the design itself. On a recent NVIDIA Jetson Orin NX project, the shortage significantly raised the cost of the RAM and of the NVMe storage for one of our customers, and the price of the Jetson modules went up as a consequence.

At AESTECHNO, an electronic design firm in Montpellier, we see the same pattern on every board that carries DRAM or flash. This article explains where the shortage comes from, how far prices have moved with sourced numbers, which memory types are being discontinued, and what a product team should check and decide, from DDR4 versus DDR5 to modules and storage.

It is written for the engineer, buyer or product manager who opened a supplier quote this year and found the memory line several times higher than last year. The short answer is that prices will keep rising into 2027, more slowly, and that the right response depends on where the memory sits in your design.

What the 2026 memory shortage is

The 2026 memory shortage is a supply gap in Dynamic Random-Access Memory (DRAM) and NAND flash caused by memory makers moving wafer capacity to High Bandwidth Memory (HBM) and server DRAM for AI data centres. Conventional DRAM, the kind used in embedded boards, PCs and modules, gets the remaining capacity, and its contract prices rose faster than any forecast predicted.

The three large DRAM makers say so in their own words. Micron announced on 3 December 2025 that it would exit its Crucial consumer business, explaining that "the AI-driven growth in the data center has led to a surge in demand for memory and storage" and that the exit would "improve supply and support for our larger, strategic customers". SK hynix said on its third-quarter 2025 call that its DRAM and NAND capacity was virtually sold out for 2026, and in October 2026 that DRAM and HBM volumes were essentially sold out through 2029.

What the order books say

Samsung told investors at the end of October 2026 that tight memory supply should continue through 2028. None of this is a forecast by a small analyst: it is the suppliers describing their own order books. According to IEEE Spectrum, HBM and other cloud-related memory went from 17% of Micron's DRAM revenue in 2023 to nearly 50% in 2025, which shows how fast the mix moved.

Why HBM takes the wafers

An HBM stack is built from several DRAM dies stacked on a base die, and each gigabyte of HBM consumes considerably more wafer area than a gigabyte of standard DDR5. A memory maker choosing between the two sells HBM to AI accelerator vendors on long-term agreements at higher margins, so advanced nodes go there first. TrendForce describes the result for older memory as "structural tightening in mature-node DRAM supply": fewer wafers for DDR4, LPDDR4 and the other parts embedded products still use.

Why embedded and industrial products are hit harder

Embedded and industrial products are hit harder by the memory shortage because they use older memory types, buy smaller volumes and sit last in suppliers' allocation. DDR4, LPDDR4 and eMMC are exactly the parts memory makers are winding down, and an industrial customer ordering a few thousand pieces has no long-term agreement to protect its share.

TrendForce put it plainly in August 2025: the consumer DRAM segment, which covers industrial control, networking, TV and controllers and "mainly uses DDR4", "ranks behind PC and server applications in suppliers' allocation priorities", and as a result its "supply-demand imbalance is particularly severe". That segment is where most embedded boards buy their memory, so the headline DRAM figures understate what an industrial buyer sees.

The lifecycle mismatch

The lifecycle mismatch makes it worse. An industrial or medical product is designed to sell for ten years or more, with a qualified Bill of Materials (BOM) that cannot change every quarter: a medical device built to IEC 60601-1 and IEC 62304, an automotive unit under ISO 26262 or an industrial controller assessed against IEC 61508 must re-verify a component change under change control, and an ISO 13485 quality system documents every substitution. A consumer phone moves to a new memory generation every year or two, so the memory industry follows the phone, not the industrial controller.

Despite this, many embedded designs from 2020 to 2024 chose DDR4 or LPDDR4 because they were cheap, mature and supported by every SoC, which is precisely what now exposes them.

Small buyers are served last

Large cloud providers sign a Long-Term Agreement (LTA) that reserves capacity years ahead; SK hynix said in October 2026 that such agreements already cover more than half of its 2027 and 2028 capacity. A company building a few thousand industrial units a year buys through distributors or from a module vendor, at spot or short-contract prices, and receives what is left. Unlike a hyperscaler, it cannot trade volume for priority. The practical consequence is that the price an embedded team pays moves faster and further than the headline contract figures.

How far prices moved: the 2025-2026 numbers

Memory prices rose by multiples, not percentages, between mid-2025 and late 2026. According to TrendForce, conventional DRAM contract prices rose 45 to 50% in the fourth quarter of 2025 and 93 to 98% in the first quarter of 2026, with further increases forecast for every quarter of 2026, and NAND flash following the same direction.

The table collects the figures from TrendForce press releases, with the date each was published and whether it is a reported result or a forecast. Read them as compounding: two quarters at +50% and +95% roughly triple a price, before the forecasts for the rest of 2026 are applied.

Two points stand out. First, the surge outran the forecasts: in September 2025 TrendForce expected +8 to 13% for the fourth quarter, and the reported result was +45 to 50%. Second, the slowdown in late 2026 is a slowdown in the rate of increase, not a fall in prices. A buyer waiting for prices to "come back" to 2024 levels has no forecast on their side.

What a board maker's price list shows

Contract prices are abstract; a public price list is not. Raspberry Pi, which uses LPDDR4 on the Raspberry Pi 4 and 5 and on its Compute Modules, has published every memory-driven increase on its news blog. In October 2025 it wrote that memory cost "roughly 120% more than it did a year ago". On 1 April 2026 it cited "a seven-fold increase over the last year in the price of the LPDDR4 DRAM used on Raspberry Pi 4 and 5", and raised 8 GB boards and Compute Modules by $50.

On 2 November 2026 the 16 GB Raspberry Pi 5 went up by another $80, to $385, because of "a lack of alternative supply for 16GB DRAM". The same board was $145 after the December 2025 increase.

Industrial vendors pass the same costs on. KUNBUS, which builds the Revolution Pi industrial controllers on the Raspberry Pi Compute Module, wrote on 1 April 2026 that Compute Module prices had risen "by an average of around 80%" since the start of the year and raised its RAM surcharge for 8 GB versions from 26 to 70 euros; in July 2026 it added a separate eMMC surcharge.

DDR4, LPDDR4 and eMMC: end of life and the long-lifecycle exception

DDR4 and LPDDR4 End of Life (EOL) means the large memory makers have stopped or are stopping production of these generations to free capacity for DDR5, LPDDR5 and HBM. The schedule has moved several times, and one maker has reversed course for long-lifecycle markets, so a design team should verify the status of its exact part rather than rely on headlines.

In mid-2025, press reports relayed by TrendForce gave the exit plan: Samsung taking last DDR4 orders in June 2025, Micron notifying end of life in June 2025, SK hynix taking last orders in October 2025 with final shipments in April 2026. As prices climbed, DDR4 became profitable again and some suppliers extended production for specific customers; reports then conflicted on whether Samsung kept its schedule. For LPDDR4X, TrendForce wrote in August 2025 that US and Korean suppliers planned to "reduce or halt LPDDR4X output between 2025 and 2026".

The long-lifecycle exception

The notable exception is Micron's Manassas, Virginia fab, backed by a CHIPS for America award to secure legacy memory for automotive, industrial and aerospace markets. It began 1-alpha DRAM production in May 2026 for long-lifecycle DDR4 and LPDDR4 aimed at automotive, industrial, medical, networking and aerospace customers, with qualified output expected by the end of 2026. Even though that is good news for industrial designs, it is one supplier, and its output will be allocated.

Where the memory sits in your design decides your exposure

Where the memory sits in a design decides who absorbs the shortage and what can be done about it. A chip-down board exposes you directly to memory prices and end-of-life notices; a system-on-module moves the memory to the module vendor, whose price list then moves with DRAM; and storage, eMMC or NVMe, is a separate exposure on top of both.

Contrary to the common assumption that buying a module shields a product from component shortages, the module price follows the memory soldered on it. According to NVIDIA, the Jetson Orin NX carries 8 or 16 GB of 128-bit LPDDR5 at 102.4 GB/s on the module, and the Raspberry Pi Compute Module carries LPDDR4: when those DRAM prices rise, the module price rises. What the module does change is the engineering work: you do not route the memory, and you do not requalify a new memory part yourself, because the vendor does it.

Storage is a second exposure

Storage deserves its own line. A Jetson Orin NX design takes its storage on an external NVMe SSD on the carrier board, a Compute Module design often on eMMC; both are NAND, and NAND prices moved as sharply as DRAM in 2026. Which storage device can actually boot the module is a carrier board decision, covered in our guide to Jetson carrier board design for Orin NX and Orin Nano.

Field report: a Jetson Orin NX product whose module price rose

On a recent project built on the NVIDIA Jetson Orin NX, the memory shortage significantly raised the cost of the RAM and of the NVMe storage for our customer, and the price of the Jetson modules went up as a consequence. Resin costs rose on the same product. The design did not change; its bill of materials did.

The case is instructive because it is the configuration most teams consider safe: an edge AI product running inference on a module rated up to 157 TOPS, with the memory and its sourcing left to NVIDIA. The Orin NX carries its LPDDR5 on the module, so nobody on the project routed or sourced DRAM, yet the DRAM price reached the product through the module. The NVMe SSD on the carrier board followed NAND prices separately. Contrary to the idea that a module protects you from the memory market, the module passed the market straight through; what it saved was the redesign and requalification work, which is real but different.

Shortages rarely come alone

In our practice, the resin increase on the same product was a reminder that shortages rarely come alone: see our article on the PPE resin shortage and PCB laminates.

What we recommend after this project

Our measurement methodology for this kind of exposure is a bill-of-materials review, not a bench test. We recommend treating memory and storage as two separate cost lines with their own watch dates, asking the module vendor for its memory roadmap and surcharge policy in writing, and checking at design review whether the storage capacity specified is the capacity the product actually needs. Even though the module price is not negotiable for a small customer, the storage size and the number of module variants often are. For the platform itself, our guide to NVIDIA Jetson Orin for embedded AI covers the module choice.

Our method: a memory exposure audit in five steps

A memory exposure audit is a structured review of every DRAM and flash part in a product, its supply status and the cost of changing it. It takes a few days on an existing design and tells a product team whether to buy ahead, second-source, migrate or simply budget for higher prices, instead of reacting to each supplier notice.

Our test procedure follows the same five steps on every product. The audit starts from the bill of materials and ends with a decision per part, so that engineering time is spent only where the exposure is real. Each step produces a written record the buyer and the project manager can use without the engineer in the room.

1. Inventory. List every DRAM, LPDDR, eMMC, NVMe and NOR/NAND part, including those inside modules, with manufacturer part number, density and package.
2. Status. Check each part against the manufacturer's Product Change Notice (PCN) history and lifecycle status, and against at least two distributors' stock and lead time. Record the Last-Time Buy (LTB) date where one exists.
3. Alternatives. For each part, list pin-compatible alternatives from other makers and check them against the SoC vendor's memory support list. A JEDEC-compatible part is not automatically supported: the memory controller and its training firmware decide.
4. Cost of change. Estimate what each alternative costs to adopt: a new DDR training configuration in the firmware and Board Support Package (BSP), a PCB respin for a different package with new impedance and EMC checks, or a module change with carrier adaptation.
5. Decision. Choose per part: buy ahead, second-source, migrate or accept, and set a review date.

Requalifying a memory change

A second-source memory on the same footprint still needs validation. The generations differ electrically too: JEDEC DDR4 runs at 1.2 V up to 3200 MT/s, DDR5 at 1.1 V from 4800 MT/s, and LPDDR4X lowers the I/O supply to 0.6 V, so a newer part is rarely a drop-in replacement. On a chip-down design we rerun the DDR training, check timing margins and measure the bus on our Tektronix oscilloscope with the TekExpress suite, which runs DDR compliance tests, before releasing the change. A different package or a move from DDR4 to DDR5 is a layout change and goes through the full high-speed PCB design and validation flow, including signal integrity simulation of the new memory interface.

Decision matrix: buy ahead, second-source, migrate or wait

The right response to the memory shortage depends on product volume, remaining lifetime and how hard the memory is to change. Buying ahead suits a mature product near the end of its life, second-sourcing suits a chip-down design with compatible alternatives, migrating suits a product with years of sales ahead, and accepting higher prices suits low-volume products where engineering costs more than memory.

The matrix below is the one we use with customers after the audit. It is deliberately coarse: its purpose is to stop a team from choosing a redesign by reflex, or from buying two years of stock that ties up cash for a product that may be redesigned anyway. Each row assumes the audit has already confirmed the part's status; a decision taken on a rumour about an end-of-life date is a decision taken twice.

When does it end? The 2027-2028 outlook

The memory shortage is not expected to end before 2028 for DRAM, while NAND flash may balance earlier. TrendForce estimated in July 2026 that the DRAM supply-demand gap will widen further in 2027 and that new fabs will not make substantial output contributions until 2028; Samsung and SK hynix gave consistent statements to their investors in October 2026.

New capacity is coming, but slowly. Micron expects first wafer output from its Idaho fab in mid-2027 and broke ground in New York in January 2026; SK hynix brought its M15X fab in Cheongju into operation ahead of schedule in late 2025. A DRAM fab takes years to build and qualify, which is why Samsung told investors that higher capital spending in 2027 would add little output before 2028.

NAND may ease first

For NAND, TrendForce expects the supply balance to turn positive in the second half of 2027 as new capacity arrives, with downward price pressure from then on. In practice this means storage costs may ease about a year before DRAM costs do. A product plan for 2027 should therefore assume DRAM prices still rising, more slowly than in early 2026, and treat any NAND relief as upside rather than as the budget.

Bottom line

The 2026 memory shortage is a structural reallocation of wafers to AI memory, not a short spike, and embedded products feel it more than PCs because they use the generations being wound down and buy in small volumes. The response that works is a part-by-part decision based on where the memory sits, not a blanket redesign.

• The cause is HBM. Memory makers move capacity to HBM and server DRAM for AI; conventional DRAM gets what is left.
• Prices multiplied. Conventional DRAM contract prices rose 45 to 50% in 4Q25 and 93 to 98% in 1Q26 according to TrendForce, and keep rising more slowly through 2026.
• A module does not shield the price. On our Jetson Orin NX project the module price rose with DRAM; the module saved the redesign work, not the cost.
• Audit before you redesign. Inventory, status, alternatives, cost of change, decision: most parts need a second source or a budget, not a new board.
• Plan for 2028, not 2027. DRAM stays tight until new fabs ramp; NAND may ease from the second half of 2027.

Related reading: managing electronic component shortages and obsolescence, obsolescence management: last time buy or redesign, LPDDR4 memory design and routing, SoC, SoM, SBC or custom board in 2026, COTS versus custom electronics, what drives electronic product development cost, design for manufacturing, electronic project risk management, the PCI Express bus behind NVMe, deploying embedded industrial AI, the DDR4 DIMM and SO-DIMM pinout, the M.2 pinout for NVMe SSDs, our hardware design services and our embedded firmware expertise.

Written by Hugues Orgitello, an electronic-design engineer with 10+ years of experience who founded AESTECHNO in Montpellier in 2022.

FAQ: the 2026 memory shortage

Why is RAM so expensive during the 2026 RAM shortage?
Because memory makers moved wafer capacity to high-bandwidth memory and server DRAM for AI data centres, leaving less for conventional DRAM. TrendForce reports conventional DRAM contract prices up 45 to 50% in the fourth quarter of 2025 and 93 to 98% in the first quarter of 2026. Micron, SK hynix and Samsung all describe their capacity as sold out or tight into 2028, so the cause is structural rather than a temporary spike.

Is DDR4 end of life?
Largely, yes. Samsung, SK hynix and Micron announced the end of DDR4 production for mainstream markets in 2025, although some extended supply for specific customers as prices rose. Micron also started long-lifecycle DDR4 and LPDDR4 production in Virginia in 2026 for automotive, industrial and medical customers. Check the lifecycle status of your exact part number with its manufacturer and distributors rather than relying on general reports.

Does a system-on-module protect my product from the shortage?
It protects you from the engineering work, not from the price. The memory is soldered on the module, so the vendor handles end-of-life notices and requalification, but the module price follows DRAM prices. On a recent NVIDIA Jetson Orin NX project, the module price our customer paid rose as a consequence of the memory shortage, and the NVMe storage on the carrier board rose separately.

Should I redesign from DDR4 to DDR5 now?
Only if the product has several years of sales ahead and no supported second source exists. A DDR5 migration means a new layout, new DDR training, signal integrity validation and often a different SoC, so it costs far more than a second-source qualification. For a product near the end of its life, a last-time buy is usually cheaper; for a module-based product, the decision belongs mostly to the module vendor.

When will memory prices go down?
Not soon for DRAM. TrendForce expects the DRAM supply gap to widen in 2027 and new fabs to contribute substantially only from 2028, so DRAM prices should keep rising, more slowly than in early 2026. NAND flash, used in eMMC and NVMe storage, may reach balance in the second half of 2027, with prices easing from then on. Budget 2027 on that basis.