The electronics industry is in the grip of a memory shortage – and this one is different. Unlike the supply shocks of the COVID-19 era, today’s crunch is being driven by a fundamental shift in who consumes memory and how much they need.
Experts who joined a recent Procurement Pro webinar – Claudio Chan, Director of Global Commodities at Smith; Margaret Upshur, Founder and CEO of Mobius Materials; and Nikolaos Florous, Global Product Marketing Director at Memphis Electronic – painted a picture of a market that has structurally changed, and may not return to the conditions procurement teams took for granted just a few years ago.
A structural shift, not just a cycle
The central point of agreement among all three panellists was this: AI is not just adding demand – it is permanently raising the floor.
“We’re talking about a massive long term shift to the demand profile in memory,” said Upshur, pointing to hyperscalers committing hundreds of billions in annual CapEx, with AI capacity expected to account for around 20% of long-term memory demand. “That’s what’s made the shortage more severe and longer lasting than some of the others.”
Chan reinforced this from a supply angle. Manufacturers have not collapsed production – they have redirected it. “Significant portion of leading-edge capacity is being prioritised for AI applications,” he noted, adding that suppliers are now exercising far greater capital expenditure discipline than in previous cycles, choosing profitability over volume growth. Florous pointed to the technical dimension: advanced node migration – the move to 1b, 1c, 1d DRAM and 200-plus-layer 3D NAND – has increased capital intensity per wafer, making it harder and slower to expand effective capacity.
High bandwidth memory (HBM), the format demanded by AI accelerators, is particularly disruptive because it is not additive to supply – it competes directly with conventional DRAM. “As suppliers such as SK Hynix and Samsung relocate wafer starts towards HBM to maximise profits, the commodity DRAM supply tightens,” said Florous. The bottleneck is dual: front-end wafer capacity and back-end advanced packaging, both constrained simultaneously.
Who is feeling it most
The pain is not evenly distributed. AI infrastructure companies, backed by enormous resources and long-term agreements, are largely insulated. It is the markets in the middle – medical devices, automotive OEMs, Industrial IoT, and embedded systems – that are bearing the sharpest cost.
Upshur described two profiles of end markets in crisis: those with low volumes and limited negotiating power, and those operating under heavy regulatory frameworks that make redesign extremely costly. “Medical customers are really in trouble,” she said. “If they do a redesign, it’s incredibly costly – a lot of times there are FDA hurdles to pass – but they don’t have the negotiation power that consumer electronics would have.” Automotive customers face similar constraints, stuck on older nodes like DDR4 with auto-grade specifications that are now acutely constrained.
Meanwhile, the concentration of supply is itself a risk. As Upshur noted, roughly 95% of DRAM capacity sits with just three manufacturers: SK Hynix, Samsung, and Micron. “Their allocation profiles pretty much affect the entirety of the market situation.”
What engineers and procurement teams should do
The practical advice from all three speakers converged on a few key themes.
Act early and stop searching for exact matches. “The biggest misstep I’ve seen is customers coming to us having looked for an exact part number for 10 weeks,” said Upshur. “They’ve lost that 10 weeks – it’s gone.” The moment supply tightens on a single component, teams should be qualifying pin-compatible alternatives and ranking options, not waiting for the original part to return.
Design for procurement flexibility. Florous introduced a phrase that was “design for procurement.” The idea is that engineering teams should minimise temperature and tolerance requirements where possible, and build in architectural flexibility – for example, designing around controllers that can support both DDR4 and DDR5 rather than locking into one.
Stop treating memory as a commodity. “Start treating memory as a first-party strategic concern,” Upshur urged. “We’ve all spent years in a world where memory is easy to get. Now it’s time to shift our thinking.” Florous was equally direct: “Organisations need to treat memory as a strategic resource, not as a commodity.”
Supplier relationships have changed
Perhaps the most striking signal of how much the market has shifted came from Florous, recounting a recent supplier meeting: “They told me that in the old times, customers were choosing suppliers. Now the rules of the game are the opposite – suppliers are choosing their customers.”
Suppliers, particularly those with constrained capacity, now favour customers who provide linear, predictable demand rather than those who double-book orders or chase short-term price advantages. Chan echoed this: “Good buyers should be transparent, consistent, and realistic. They shouldn’t go silent during downturns, and they should engage even more when they’re not buying aggressively.”
The practical implication is dual sourcing – not just across components, but across vendors. A direct relationship with a major manufacturer needs to be supplemented with a vetted secondary source that can provide stock when primary lead times stretch.
On counterfeits, all three panellists sounded a clear warning. During shortages, counterfeit risk rises sharply. “There’s blood in the water,” said Upshur. “This is a prime time to be very careful.” Physical testing of every batch, financial transparency in transactions, and clear chain of custody for components are the minimum bar.
When does it end?
Not soon. All three panellists were broadly aligned: significant relief is unlikely before 2027, and even then, prices will not return to 2022 or 2023 levels. New fab capacity from Hynix and Micron targeting advanced nodes is not expected online until around 2028. At the legacy node end of the market, supply has dropped dramatically, and tier-two suppliers are only slowly moving to fill the gap.
“Memory is going to be a first-party strategic concern for at least another two to three years,” said Upshur. For engineering and procurement teams that have long treated it as a background line item, the adjustment required is significant – and the time to make it is now.
