The Phison CEO has just uttered something that every PC builder should take note of. Pua Khein-Seng is referring to it as a 10-year memory shortage. Not a bad quarter. Not a rough cycle. A decade.
That's a big claim. However, the statistics supporting it are difficult to dispute.
Right now we're sitting in what some analysts are calling "Memflation." The DRAM contract prices are expected to increase by 58 to 63 percent in this quarter alone. NAND flash, which your SSD is composed of, is seeing 70 to 75% growth. If you bought storage or RAM in the last 6 months you were probably feeling lucky. You were.
Gartner sounded an alarm that PC shipments increased by 4 percent in Q1 but termed the increase as artificial. Vendors had been hoarding in advance of the price wall. The expansion was not actual demand. It was suit panic buying.
So why a decade. That is what really counts.
The brief response is HBM. Samsung, SK Hynix, Micron and CXMT are all racing to make High Bandwidth Memory to serve AI GPU stacks, primarily to serve the data center demand of Nvidia. And HBM is cruelly wasteful of wafer. It uses about three times the wafer area of standard DRAM with the same capacity. Each HBM chip built to serve a server is fundamentally three sticks of consumer RAM that are never produced.
That's the core of the problem. It's not a temporary hiccup. It is a structural redistribution of the whole manufacturing base to enterprise and not to you.
And then there's DDR4. Manufacturers are officially shutting down production. Which is okay until you discover that hundreds of millions of low-end laptops and phones continue to use it. Reduced supply, unchanged demand. DDR4 has almost doubled its price since late 2025 and is sometimes trading higher than DDR5 at the moment. That sentence shouldn't make sense. It does.
To builders the fallout is already becoming practical. The new minimum storage is 256GB or 512GB, and entry level laptops are likely to be shipped with this as the new minimum, since sticker prices would scare away buyers. Even GDDR7, the memory in the RTX 50 series cards, is competing with the same limited fabs as server memory. Therefore, the same way is the pressure on the price of GPUs.
According to TrendForce, it is unlikely that meaningful new capacity will be online until late 2027 or 2028. That's not a typo. The present scenario is supply based and the supply issue has a multi year runway.
The wait and see strategy is becoming more risky should you be planning a serious build this year. HBM margins are currently operating far above consumer chip profitability and therefore manufacturers have very little incentive to switch back. The incentive system is not focused on your 32GB kit but on another rack of AI accelerators.
It remains to be seen whether the full decade timeline will be maintained or the AI build out will eventually reach a wall and release capacity earlier than expected. The 10 year figure may be a pessimistic one or it may seem clear in retrospect. What is more difficult to dispute is that the following 18 months will be costly and there is no apparent mechanism to alter that in the short term...
Post a Comment