It's gotten worse – and the memory market isn't done yet
Two months ago, at Evertiq Expo Zürich, Nikolaos Florous left little room for optimism. Tight memory market conditions would persist, he said — and things would likely get worse before they got better.
At Evertiq Expo Berlin, the question put to him was simple: You previously said it would get worse before it gets better. Well, it's gotten worse – what happened?
"The AI cloud investment is still unprecedented," he said. "It is accelerating. It's not slowing down."
Hyperscalers are on track to invest USD 700 billion in 2026, with projections pointing toward USD 1 trillion by 2028. Samsung, SK Hynix, and Micron continue to concentrate their focus on the high-margin products serving that demand – HBM and through-silicon via. Everything else remains under pressure.
The relief that isn't
Announcements of capacity expansions have continued to filter through. Micron, for instance, has signalled an expansion of DDR4 and LPDDR4 capacity out of its Virginia fab. Florous was measured about what that actually means in practice.
"This will only give some relief to dedicated customers in aerospace, industrial, military and automotive applications — within the US timeframe."
Meanwhile, Chinese memory maker CXMT is making strides in DDR4 and LPDDR4 – but not in a direction that helps the broader market. Its expansion is driven by China's semiconductor sovereignty agenda, not by a desire to relieve global supply constraints.
"Basically, nothing changes," Florous said. "Scarcity is expanding. The wave of AI infrastructure cloud investment as a superposition cycle is much stronger than the memory cyclical behaviour."
New entrants, limited impact
The question of whether structural change might open the door to new players is one Florous has heard before. Taiwanese manufacturers Nanya and Winbond are making inroads, he acknowledged, and are positioned to gain market share as the big three shift their attention elsewhere.
The catch is capacity. "Their capacity is not enough to relieve the capacity dislocation from Samsung, SK Hynix, or Micron." Demand will reduce, but not fast enough to offset the supply shortfall in legacy products. Structurally, a shortage in DDR3 and DDR4 remains the baseline.
The market is moving to long-term agreements
One of the more significant shifts Florous described is behavioural rather than structural. The spot market, historically a fallback for buyers caught short, is losing ground to long-term agreements – and not just among the hyperscalers.
"A lot of the supply chain is shifting from the typical spot behaviour to long-term agreements," he said. Smaller players and fabless companies are increasingly seeking MOUs and supply agreements extending from six months to two years. "The capacity is already being sold out, or they are in discussion to be sold out."
For procurement teams still relying on spot availability, this is the most direct warning: the market is being divided between those who have secured supply and those who have not.
Design for procurement
The sharpest point of the Berlin conversation came not from the market data but from a single slide in Florous' presentation – one that asked customers to weigh the strategic importance of memory against risk. His message was a departure from the usual procurement guidance.
"I have asked people to consider designing for procurement," he said. "This means that in the equation now, within their CTQs, they have to add supply chain resilience, flexibility, and price trends as well."
The implication is significant. Supply chain considerations – historically treated as downstream of engineering decisions – now belong in the design specification itself, weighted equally alongside technical requirements.
"This needs to be taken into account as equally important as all the other technical CTQs."
