Memory no longer feels predictable. What was widely available only a few quarters ago is now increasingly governed by allocation, long lead times, and limited flexibility across several established memory technologies.
In this Q&A with Avnet Silica’s Jozsef Miho, Supplier Business Manager and Memory Technology Specialist, we discuss why memory availability has changed so quickly, how AI is reshaping allocation, and what engineering and procurement teams can do to protect long-life programmes.
The semiconductor market is showing signs of recovery overall, so why does the memory segment feel constrained and unpredictable from both a supply and engineering perspective?
It feels different because memory has moved much faster than the wider semiconductor market. In the first half of 2025, many memory products were still widely available. By the third quarter, the market was starting to recover, but by the fourth quarter, the rapid and largely unpredicted rise in AI-related memory demand had pushed several areas of the market into severe constraint. That is a very sharp change in a short period.
Memory is also often a leading indicator for the rest of the semiconductor market. In normal conditions, it gives us an early view of where the wider market may be heading as memory is a fundamental electronic component. However, what we are seeing now is not a simple return to demand, it is a shift from availability to unprecedented allocation.
This matters because growth and availability are now moving in different directions. Our latest Trendliner shows that the wider served semiconductor market is moving towards an upcycle, while memory is also clearly benefiting from strong demand, especially from data centre and AI investment. But in memory, growth is not translating into easier access. It is the exact opposite, driven by constrained capacity, rising average selling prices (ASPs), high-bandwidth memory (HBM) prioritisation and intense competition for wafer output.
That is why memory feels different from a normal recovery market. In a balanced recovery, higher demand is gradually met by improving supply. In memory, higher demand is absorbing available capacity faster than the supply base can respond.
The change is visible in our Trendliner data. In Q1 2025, legacy DRAM was still widely available, and eMMC was typically at 9 to 16 weeks. By Q2 2026, DRAM, NAND flash memory, eMMC, NOR flash memory, and SSDs are all shown at 26 weeks or more, with several supplier lines under allocation. In practice, however, many of these products are operating beyond what a normal lead-time discussion really captures. Allocation can mean customers receive fewer parts than ordered, reduced delivery visibility, or in some cases difficulty securing order confirmations at all. Some suppliers are only confirming within the current quarter, while others are pushing confirmed deliveries far further into the future. That is the structural change, and AI is not just increasing demand for memory, it is changing which customers and technologies receive priority in production planning and allocation.
To what extent is this being driven by AI demand, and how much reflects deeper structural changes in how memory is produced and allocated?
In my view, this is being driven almost entirely by AI demand. The scale is almost difficult to overstate, and around 70% of global dynamic random-access memory (DRAM) production is expected to go into the AI market in 2026. This was not demand that the market could easily forecast several years ago. The size and speed of AI, and the demands it creates, are far beyond anything the memory market has seen before.
HBM is central to this demand ‘headache’ because it is critical to AI data centre infrastructure. As demand for AI accelerators has grown, memory suppliers have naturally shifted more production towards HBM and other advanced memory technologies. This also happened at a time when much of the wider market was still saturated with memory inventory, meaning demand visibility from other sectors remained relatively weak as many customers were still consuming overstock. Nevertheless, the commercial logic of memory suppliers is understandable, as AI customers are buying at a scale, with a level of commitment, that is difficult for other market segments to match.
Large data centre and hyperscale customers are placing major forward commitments, absorbing the higher pricing, and giving suppliers the visibility they need to allocate capacity. However, that is the core issue for the rest of the market. AI is not only adding demand, it is taking priority in how memory capacity is planned, priced and allocated.
Why are established and widely used memory technologies such as DDR4, LPDDR4, and eMMC becoming increasingly difficult to secure, particularly for long-life industrial and embedded applications?
Because unfortunately established does not mean protected. For many industrial and embedded customers, DDR4, LPDDR4, and eMMC are still highly relevant technologies. They are stable, proven, and well understood. In many applications, the priority is not maximum performance, but more likely reliability, longevity, cost control, and qualification stability.
The problem is that today, many leading-edge technologies and AI-related demand have this new level of priority. Older or more established technologies can still have strong demand, but they may not receive the same level of capacity expansion due to AI-related allocations.
That creates a mismatch. Our customer base may see peak demand for DDR4 and LPDDR4 in 2027 and 2028. At the same time, the supply base is moving investment towards newer technologies. That does not mean DDR4 and LPDDR4 will disappear and Micron, for example, has committed to DDR4 and LPDDR4 for the long term and is transferring LPDDR4 production into its legacy Fab6 facility in Manassas. Once that fab ramps, expected in the second half of 2027, the situation should improve, but there are likely to be some challenging quarters before that additional capacity helps. Importantly, future output will also be shaped by the demand signals customers provide now. If customers hold back forecasts or orders while waiting for lower prices, suppliers may plan too little capacity for the wider market.
The situation with eMMC follows a similar pattern. The underlying NAND flash memory used in eMMC is also heavily needed for SSDs, which are seeing strong demand from large language models (LLMs), inference infrastructure in AI data centres. As more NAND is directed into SSD production, eMMC increasingly becomes a managed and allocated product category.
For long-life industrial, automotive, medical, and embedded programmes, this creates a very real problem. At component level, a change may be manageable if compatible alternatives have already been qualified, which is why multi-sourcing during design is so important. But at technology or generation level, the impact is much greater. Moving from LPDDR4 to LPDDR5, for example, is not a simple substitution. It can mean a full redesign, effectively starting the project again. That is why availability pressure in memory quickly becomes a significant programme risk, rather than a simple procurement issue.
There is renewed discussion around ‘allocation’ in memory. How is allocation behaving today, and how does it differ from previous shortage cycles?
As we have touched on the magnitude is really different. In previous shortage cycles, allocation was often treated as a temporary disruption that would be resolved relatively quickly. Today it feels far more structural. Large memory suppliers are already sold out through the end of calendar year 2027 and beyond, meaning if a customer wants memory during that period and it is not already planned, that capacity has to come from somewhere else.
That is why demand visibility matters so much, and why at Avnet Silica we place significant focus on end-to-end planning alignment, forecast discussions, long-term supply planning, and generally just the way we build relationships with both our suppliers and customers.
Visibility is also key for allocation, suppliers do not respond only to interest. They look for credible demand and commitment. That means purchase orders, order history, realistic forecasts, long-term agreements, and evidence that the customer will take the material. Forecasts still matter, but only if they are defensible.
Over-forecasting is also not a good strategy, and while it may attract attention for a short period, it damages credibility if the orders do not follow. That is why suppliers increasingly use non-cancellable, non-reschedulable (NCNR) terms, and why customers should avoid placing speculative orders across multiple sources, only to cancel once the first confirmation arrives. This behaviour was seen in the last allocation cycle, and it weakens the trust suppliers need when deciding how to allocate constrained supply.
From the supplier side, the investment response is also different to previous cycles. Suppliers are investing incredible sums into new capacity, but fabs do not arrive quickly. A new semiconductor fab typically takes several years to build, equip, qualify, and ramp to meaningful output.
For legacy technologies, Micron is transferring LPDDR4 and DDR4 production into its Fab6 facility in Manassas, which should help once that capacity ramps. At the same time, Micron’s larger, leading-edge investments in Boise, Idaho, with output expected from 2028, and Clay, New York, with output expected from 2030, show how long the capacity cycle really is. Samsung and SK hynix are working to similar timelines for their own leading-edge investments. Together, these three suppliers account for around 90% of global DRAM production, so their investment schedules define much of the market’s medium-term capacity outlook.
From an engineering and programme delivery perspective, where are these market conditions having the most immediate and tangible impact?
The biggest impact is that memory is no longer just a purchasing decision. It is becoming a design decision again. A common misconception is that memory can be swapped easily. Sometimes it can. If a design allows for equivalent parts within the same generation, with the right validation, there may be options. But in many cases memory is tied closely to the processor architecture, memory controller, layout, timing, and qualification work.
Moving from one generation to another is not simple and typically comes with significant design implications that affect the platform itself. This is especially relevant because the market is moving towards leading-edge memory, while many industrial and embedded designs are still best served by DDR4 or LPDDR4. At the same time, microprocessor units (MPUs) that support LPDDR5 are still relatively limited for some target applications. They can also be over-specified, more expensive, and more complex to design with. Higher frequencies make designs more sensitive to signal integrity, layout, and power management challenges, which can increase development cost and extend design cycles.
However, all of this should not be read as an immediate must-move message. The industry’s long-term investment shows these technologies will remain available for years, particularly for industrial and automotive markets. Nevertheless, for future developments, the direction of travel should be part of the discussion. Newer technologies will become more widely adopted, supporting devices will become broader, costs will come down, and the global production base will increasingly favour technologies used at scale. That does not remove the need for legacy support, but it should change the long-term risk profile.
Given these constraints are unlikely to ease in the near term, what can engineering and procurement teams realistically do to protect programmes and reduce exposure to memory risk?
For running projects, it often does not make sense to redesign purely because of the current memory situation. In many cases, the effort and risk would be so high that it could be close to starting the project again.
The most important action is to share demand early and order as early as possible. If demand is visible to suppliers and supported by realistic planning, products can still be secured. But in the current market, every wafer not needed by the industrial and embedded sector can quickly be redirected towards AI demand. Ordering late and hoping for short-term availability is not a good strategy.
It is also important to build flexibility where possible. That means qualifying compatible memory suppliers and validating alternative sources where the design allows it. Flexibility matters because supply can shift quickly, and sometimes availability comes from unexpected pockets of capacity.
That is where Avnet Silica can help, by combining supplier visibility, market insight, and engineering support to help customers plan earlier and reduce exposure before supply options narrow.
