[AINews] Memory prices up 500% in 12 months
Latent Space Latent.Space
DRAM prices have gone insane — 128GB DDR5 kits now cost ten times what they used to. Hyperscalers have already locked up nearly all global DRAM supply through 2027.
Based on reporting by Latent Space, Latent.Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Tom's Hardware has a name for what's happening to memory prices right now, and it isn't flattering: RAMageddon. The numbers back up the drama. A 128GB DDR5 kit today costs ten times more than the cheapest price ever recorded for that same kind of memory. That's not a typo or a temporary spike caused by a single bad quarter. It's a full inversion of how computer hardware is supposed to behave.
Moore's Law has spent decades training everyone to expect chips to get cheaper over time, not more expensive. Memory just broke that pattern entirely. DRAM has become so valuable that, gram for gram, mainstream chips are now worth more than half as much as solid gold. Put memory on a scale next to a gold bar and the gap isn't as wide as you'd think.
The root cause, per the reporting, traces back to hyperscale buyers — the companies running the massive AI training and inference infrastructure that dominates this newsletter's other stories, from Cerebras's new CS-4 systems to the frontier labs racing each other on model releases. These buyers have reportedly already locked in almost all of the world's DRAM production capacity for 2027, and they've done it by handing over advance deposits just to guarantee they'll get supply at all. That's a two-year-out commitment on a market that's supposed to reset with new fab capacity, not calcify into a reservation system.
What makes this notable beyond the sticker shock is the timing. This shortage has been running since at least February, when outlets were already discussing it, and it hasn't eased despite the passage of months. AI training runs, RL infrastructure, and the broader push toward bigger and faster inference — all the stuff filling out today's AI news — depend on memory that is now among the most expensive commodities in the world by weight. Anyone building anything memory-hungry, from a home PC upgrade to a data center expansion, is competing for the same shrinking pool of chips against buyers who can outbid them years in advance.
My take — AI-written commentary, not fact-checked reporting
Nobody should be shocked that a compute arms race eventually runs into a hardware bottleneck — it happened with GPUs, and now it's happening with DRAM. What's genuinely striking is the mechanism: hyperscalers aren't just outbidding everyone on price, they're pre-buying years of future supply with deposits, which locks ordinary buyers and smaller companies out of the market before the chips are even made. That's not a shortage that fixes itself with a good fab quarter; it's a structural squeeze, and whoever isn't a hyperscaler should plan accordingly.
Read more about this at: Latent Space