| | | Did you see this post from a S. Korean blogger?
(Must read) Gou Jiazhang of Silicon Motion (??): The most severe, panic-level, large-scale shortage in the history of memory is approaching, and the NAND Flash supply shortfall is expected to exceed 80%.
Gou Jiazhang of Silicon Motion (??) stated that this time, all types of memory—DRAM, NAND, SSD, and HDD—are, for the first time ever, experiencing a simultaneous, large-scale shortage. The growth of AI inference has already surpassed AI training, and inference drives demand for memory—especially high-capacity SSDs—much more strongly. He asserted that the current shortage is nearing a “panic” level and that the supply shortfall will not be resolved throughout 2026. He said this, too, is a conservative view, and that if this trend continues, smaller, less competitive firms could face an existential crisis.
AI inference is driving a surge in HBM3E demand
Sharing takeaways from a recent business trip to Korea, Gou said customers kept coming to discuss only capacity and never price; the mood was that simply getting a capacity allocation would be fortunate. He said HBM demand is very strong, and since AI inference demand emerged, HBM3E demand has expanded. It is not only NVIDIA’s GPUs that require HBM3E; all cloud providers’ in-house ASICs also need HBM3E. At present, Samsung, SK hynix, and Micron are prioritizing capacity allocation to HBM and DDR5. From a capacity-consumption perspective, each unit of HBM requires roughly three times the number of wafers compared with DDR5. As a result, HBM is in short supply now—and so is DDR5.
The HDD shortage is shifting demand en masse to the NAND industry
In the China market, starting from the second quarter of this year, SSD demand has rapidly stepped up from the past 16TB/32TB to 64TB/128TB. Looking closely at the cause, because sufficient HDD volumes could not be secured, high-capacity SSDs were purchased as substitutes. Gou also analyzed that, thanks to strong AI inference demand, even if you add up all NAND Flash capacity worldwide, it still cannot meet even one-tenth of data centers’ annual warm-data demand. Therefore, over the next 10 years it will be impossible for SSDs to replace HDDs. He added that timing is the only question.
Tracing the current HDD shortage back, during the pandemic the blow from excessive inventories made HDD makers’ business models more conservative, shifting them to expand capacity only to the extent of the orders they received. This has made it difficult to absorb the sudden volume surge from the AI boom all at once. On top of this, the North American hyperscalers’ demand for ultra-high-capacity SSDs has increased to an even more astonishing level. The industry must wait for NAND technology to transition from TLC to QLC; at the earliest, in 2027 the global bit output of QLC may overtake that of TLC NAND.
The NAND shortage could trigger a reshuffle of the global smartphone market
Apple has removed the 128GB model in the iPhone 17 and pushed capacity straight up to 2TB, and Samsung is expected to follow. For 2026, overall smartphone bit growth is forecast at about 32%. In the context of a NAND Flash shortage, major vendors will defend premium smartphones from impact, but low-end models may see capacity cut. A recent example is smartphone storage in Africa and Southeast Asia falling from 128GB to 64GB.
The NAND Flash shortage will also affect the competitive landscape of the global smartphone market. Apple need not worry about shortages, and Samsung can be supported by its in-house NAND. The focus should be on branded smartphones in the China-based Android camp. Although the local NAND giant YMTC???? is actively expanding, it is estimated it can meet only 50% of local vendors’ demand. Recently, Xiaomi announced price hikes for new models citing rising memory prices. Going forward, whether smartphone vendors can secure sufficient NAND chips—and what choices they make between specifications and cost/selling price—will determine how market shares shift.
Questions about an AI bubble
Gou stated that concerns about an AI bubble will not materialize in the short term. From cloud AI to edge AI, inference consumes memory in very large quantities. The supply gap in the NAND market is 70%–80%, already at a panic-level shortage. Some overbooking naturally exists, but the key is how the major memory makers allocate and plan capacity. Including customer groups such as cloud, smartphones, autos, and PCs, many Chinese automakers still have not recognized the severity of the shortage. Moreover, all capacity cannot be diverted solely to cloud providers; doing so would lead to fragmentation of the technology industry, which is undesirable.
10:29 PM · Nov 3, 2025
x.com |
|
