Pua Khein-Seng, CEO of flash memory controller maker Phison Electronics, is warning that the fallout from the current memory crisis will widen—a warning that goes beyond simple supply-demand forecasting and hints at "the beginning of the end" for the consumer electronics industry. The extraordinary investment fever surrounding AI infrastructure is threatening to fundamentally destroy the supply chains for the smartphones, PCs, and even automobiles we handle every day.
An Unprecedented Memory Grab and the Impossibly High Wall of "Three Years' Prepayment"
Global semiconductor demand is currently being sucked into the enormous black hole that is generative AI. In an interview on the Taiwanese TV program "ChenTalkShow," Pua offered a shocking outlook: the shortage of DRAM and NAND flash memory could persist until 2030, or in the worst case, for the next decade.
What makes this crisis unusual isn't simply that supply is falling short. The foundries and memory makers who control supply have begun imposing an unprecedented and brutal condition on device makers within the electronics industry: "prepayment of three years' worth of costs."
Mega-players with ample cash flow, such as Apple and Samsung, can likely swallow these terms and secure the components they need. But for mid-tier and smaller smartphone makers and PC vendors, locking up three years' worth of cash is tantamount to a death sentence. Pua's prediction that "by the end of 2026, many system vendors will either go bankrupt or be forced to withdraw from their product lines" isn't merely a forecast—it's the logical outcome of a "survival competition" determined by financial strength.
NVIDIA's "Vera Rubin" Is Draining Away the NAND Supply
Until now, the impact of the AI boom on the memory market has largely been viewed as centered on DRAM—particularly HBM (High Bandwidth Memory). But with the arrival of next-generation AI hardware, that battlefront is rapidly expanding into the NAND flash market as well.
The symbol of this shift is NVIDIA's next-generation AI platform, "Vera Rubin." According to Pua's analysis, each Vera Rubin board requires more than 20TB of SSD storage for caching, along with up to 576GB of RAM. If NVIDIA ships 10 million units of Vera Rubin as planned, the SSD demand alone would consume roughly 20% of the entire world's 2026 NAND production capacity.
And this figure doesn't even include storage demand for actual data retention—it's purely demand for "computational support." When a single AI server consumes memory resources equivalent to thousands of ordinary laptops, the memory that's supposed to reach consumers' hands simply ceases to physically exist.
A 2000% Price Surge: The Cost Shockwave Hitting the Automotive and Smartphone Industries
This shortage isn't limited to cutting-edge chips—it's also spreading to eMMC memory, which has traditionally been considered a "cheap commodity." The example Pua cited is staggering: an 8GB eMMC chip that cost around $1.50 (roughly ¥220) each in early 2025 now exceeds $20 (roughly ¥3,000), with automotive-grade chips requiring higher reliability approaching $30 (roughly ¥4,500).
This cost surge translates directly into product pricing. RAM and storage account for over 20% of a smartphone's bill of materials (BOM). By contrast, in the AI server and data center market, the cost ratio for memory and storage is a mere 5–6%.
This gap determines suppliers' priorities. Data center customers—who offer higher margins and can more easily absorb rising costs—get priority, while consumer products are pushed to the back of the line. As a result, global smartphone production in 2026 is projected to fall by 200–250 million units—a massive decline equivalent to roughly 20% of global supply.
Phison's Trump Card, "aiDAPTIV+": A Software Answer to a Hardware Shortage
In response to this dire situation, Phison is presenting a solution built on its own strengths: the middleware technology "aiDAPTIV+."
The core of this technology lies in substituting relatively cost-efficient flash memory for the role normally played by expensive, hard-to-obtain DRAM (HBM or GDDR). By combining dedicated flash memory with optimized middleware, it expands GPU memory up to 320GB in PC environments and up to 8TB in workstation or server environments.
Rather than piling up large numbers of expensive GPU cards, this approach—extracting the potential of existing hardware through "memory tiering"—could offer a realistic workaround for companies struggling with component shortages. However, it also carries a paradox: the approach itself becomes a factor that further pushes up NAND demand.
An Industrial Shift: From Scrap-and-Build to "Repair and Extend"
Memory shortages and soaring prices will likely force a fundamental change in consumer behavior as well. As new products become unavailable—or prohibitively expensive—extending product lifespans will become unavoidable.
The culture of replacing smartphones every two to three years, as was once common, may come to an end, accelerating a shift toward a "repair economy" in which broken parts are fixed and devices are used for far longer. While this may look positive from a sustainability standpoint, in reality it's simply a byproduct of resources that should have gone to consumers instead being diverted to "industrial AI infrastructure"—like Nvidia GPUs that consume 1,400W each.
Strategic Retreat and Restructuring to Survive
Pua Khein-Seng's warning signals that 2026 will be a year of "selection" for the electronics industry. Manufacturers lacking financial strength will vanish from the market, while the survivors will be forced to develop products under constraints reminiscent of computing's early days—grappling with the question of "how to operate with as little memory as possible."
The brighter the light of AI shines, the deeper the supply anxiety grows in the consumer market hidden in its shadow. The next device we hold in our hands may no longer be a "casual piece of consumer electronics"—it may instead be a "rare item" that has survived a fierce, global battle over components.
Sources