Memory is no longer a commodity – The industry hasn't fully caught up
Nikolaos Florous of Memphis Electronic delivered a data-heavy, at times bluntly honest overview of the global memory market at Evertiq Expo Zürich – and his central message was uncomfortable: the rules have changed, and most of the industry is still playing by the old ones.
The numbers Nikolaos Florous opened with were striking enough on their own. The global semiconductor industry, previously forecast to reach one trillion dollars by 2030, is now on track to hit USD 1.2 trillion this year, four years ahead of schedule. Memory alone is expected to capture 30 to 35% of that figure.
But Florous, Semiconductor & Electronics Expert and Global Product Marketing Director at Memphis Electronic, was quick to contextualise the growth. "Not all boats are lifted by this tide," was the implicit message running through his keynote at Evertiq Expo Zürich. AI economics, he argued, has fundamentally rewritten the rules of the memory market, and most of the industry has yet to absorb what that actually means.
From perfect competition to oligopoly
To understand the present, Florous took the audience back to 1990, when more than twenty companies – most of them Japanese – were producing DRAM. Today, three dominant players remain: Samsung, SK Hynix, and Micron. A fourth, China's CXMT, has entered with meaningful wafer capacity, though its revenue share tells a different story.
"From a perfect competition, we went to an oligopoly market," Florous said. "And an oligopoly market is characterised by great volatility."
That volatility, he argued, is structural – not cyclical. A single fab line for advanced DRAM nodes now costs USD 14 billion to build, and takes up to five years to become fully operational. To paint a picture just how massive such an investment is, he offered a few comparisons: the Burj Khalifa cost USD 2 billion to construct. Germany's A7 autobahn, stretching the length of the country, cost USD 6 billion. One advanced DRAM fabrication line costs more than both combined.
"How many companies do you think have the capacity and cash flow to invest fourteen billion dollars?" he asked. "Very few. And then ask when the ROI comes back."
The avalanche triggered by a small event
The most immediately relevant section of Florous' presentation concerned what he described as a chain reaction, set off by a decision that, in isolation, appeared routine.
When Samsung, SK Hynix, and Micron announced end-of-life for DDR4, the D1z wafer capacity that had supported it was rapidly repurposed toward HBM – the high-bandwidth memory driving AI infrastructure demand. That shift, in turn, locked up OSAT capacity – the outsourced assembly and testing facilities that smaller, tier two and tier three memory players depend on.
"The OSATs are already sold out," Florous said. "They have demand from the big hyperscalers, and they're leaving the smaller players without expansion capacity."
The cascade continued: PCB material shortages, longer lead times, and tightening foundry access for fabless manufacturers followed. What looked like an orderly product transition became, in practice, a supply constraint across the entire memory ecosystem.
The shortage won't resolve before 2027 – at the earliest
On the supply-demand outlook, Florous was unambiguous. The memory market entered shortage territory in 2025 at -9%, well beyond the -1% to minus 1.5% that he described as the "ideal golden point" for suppliers: tight enough to maintain pricing, loose enough not to starve customers.
"Everything below -1 or 1.5 is already in a critical state," he said. 2026 will offer partial relief, moving to -4%, but meaningful normalisation is not expected before late 2027 or early 2028 – when new fab investments announced in 2022 finally reach full operational capacity.
In the interim, the DDR4 spot market has already shown what shortage pricing looks like in practice. From Q2 2025, OEM contract prices and spot prices diverged sharply, with spot hitting its peak around Q4 2025. Some correction is expected from Q3 2026 onward, but Florous was careful to set expectations.
"We shouldn't expect prices to return to where they were two or three years ago. That will never happen again."
The split no one is talking about
Perhaps the most structurally significant point in the presentation was one that received less attention than the supply crunch data: the memory market is spliting in a way that makes the old commodity model obsolete.
Around 2023, a crossover occurred. DRAM below 8GB became what Florous termed "specialty DRAM" – a sustainability-oriented segment defined by legacy demand, constrained supply, and no new investment. Everything above sixteen gigabit became performance-oriented mainstream DRAM, driven by AI and hyperscaler economics.
"The era of cheap memory is over," he said. "It's not coming back."
For procurement and engineering teams still designing around commodity memory assumptions – cost-driven, spot-market sourced, last-minute – this division represents a material risk. Florous was direct about what the appropriate response looks like: long-term agreements for advanced nodes, last-time buys and redesign strategies for legacy exposure, and a fundamental reclassification of memory from commodity line item to strategic supply decision.
"Memory needs to be a strategic topic," he said. "You need to enable engineering, purchasing, and sales to have minimum and long-term requirements built in before you start designing."
The companies that absorb this shift early will be positioned differently from those that don't. The supply constraints of 2025 and 2026 are, in that sense, less a crisis than a forcing function – accelerating a reckoning that was already overdue.
Nikolaos Florous will return to the Evertiq Expo stage at Evertiq Expo Berlin on June 18, where he will present "The Global Memory Market Reset."
