Something unusual is happening inside the factories that make the world’s computer memory, and the ripples are reaching every device that carries a chip. For the first time in years, memory has become scarce, and the reason says a great deal about where computing is headed.
Three companies, Samsung, SK Hynix, and Micron, produce nearly all of the planet’s DRAM, the fast working memory that every phone, laptop, and server relies on. Over the past year those companies have been quietly steering their production lines toward a specialized product called High Bandwidth Memory, or HBM. HBM is the stacked, ultra-fast memory bolted directly onto the AI accelerators that power data centers, and it has become the most profitable thing a memory factory can make. A single silicon wafer turned into HBM brings in three to five times what the same wafer earns as ordinary DDR5.
That single economic fact has reshaped the entire market. HBM now soaks up an estimated 23 percent of the industry’s total DRAM wafer capacity, a striking jump from just 8 percent in 2024. Every wafer committed to feeding AI is a wafer not making the memory sticks that go into consumer and business machines.
When the shelves go bare
The squeeze became undeniable this autumn. SK Hynix told investors that its HBM, DRAM, and NAND output is essentially sold out for all of 2026. Micron went further, stepping away from the consumer memory business entirely to concentrate on enterprise and AI clients. TrendForce, a research firm that tracks the sector, logged conventional DRAM contract prices climbing 90 to 95 percent from one quarter to the next heading into 2026.
Those wholesale jumps have already reached buyers. A 32GB DDR5 module that Samsung listed around $149 earlier in the year now sells for roughly $239. Contract prices for DDR5 have more than doubled, reaching about $19.50 per unit against roughly $7 in 2025. Even the companies at the front of the queue are paying up: Nvidia’s compact DGX Spark AI desktop rose from $3,999 to $4,699, and Apple pared back the largest memory configurations on some of its machines as supply tightened.
What sets this shortage apart from earlier memory cycles is how long the recovery will take. Chip fabs are among the most expensive and intricate facilities ever built, and they cannot be conjured on demand. New plants from Samsung and SK Hynix are not expected to reach full production before the second half of 2027. Micron has pledged a remarkable $200 billion long-term investment in U.S. manufacturing, yet its first new fab does not begin turning out DRAM until mid-2027. SK Hynix has even cautioned that the imbalance could linger past 2030.
There is a genuinely fascinating story underneath the higher prices. This shortage exists because artificial intelligence arrived faster and hungrier than anyone in the supply chain planned for. Every model that writes, draws, or reasons runs on memory, and the appetite for it has outrun what the factories can currently produce. The billions now flowing into new plants are a wager that this demand is not a passing fad but a permanent change in how computers are built.
For the technology industry, the memory crunch is a reminder that the AI revolution rests on physical foundations. Behind every sleek chatbot and clever assistant sits a wafer of silicon, made in a handful of factories, and for the moment those factories are aimed squarely at the data center. Everyone else is waiting in line, and the queue will not shorten meaningfully until the next generation of fabs comes online toward the close of the decade.
