DRAM, NAND prices jump sharply

Memory chips are having one of those moments where a boring component suddenly becomes the whole story. DRAM and NAND sit inside servers, SSDs, laptops, phones — basically everything digital. But in early 2026, they stopped behaving like cheap background parts and started acting like scarce strategic assets. The change is sharp enough that cloud budgets, server quotes, and storage plans all have to be redone. ### What actually jumped? TrendForce’s February 2 update was the clearest signal. It raised its 1Q26 forecast for conventional DRAM contract prices to a staggering 90–95% quarter over quarter, up from an already-hot 55–60% view. NAND flash got the same treatment, revised to 55–60% QoQ from 33–38%. Server DRAM was projected to rise around 90% QoQ, and enterprise SSD pricing was seen climbing 53–58% QoQ. Those are not normal cyclical moves — they are shortage moves. (trendforce.com) ### Why are “normal” memory chips suddenly scarce? Because the best capacity is being steered toward AI memory. HBM — the stacked, ultra-fast DRAM used next to AI GPUs — earns far better margins than commodity memory. So Samsung, SK hynix, and Micron all have a reason to prioritize it. The catch is that memory manufacturing is not infinitely flexible. When more wafers, packaging, and engineering attention go to HBM and high-end server parts, less is available for the plain DRAM and NAND that everyone else buys. (trendforce.com) That is the spillover. ### Why does AI inference matter here too? Training got the headlines first, but inference is now a huge part of the squeeze. Running models at scale needs lots of fast memory for active data and lots of flash for checkpoints, vector stores, and KV-cache-heavy storage systems. Samsung explicitly pointed to KV cache storage demand as a driver for next-gen enterprise SSDs, and SK hynix said the shift toward agentic AI and real-time inference is widening demand across both DRAM and NAND. (spglobal.com) So this is not just “NVIDIA needs HBM.” It is the whole AI stack pulling on memory at once. ### Are suppliers really seeing it in the numbers? Yes — and that is what makes this more than an analyst note. Micron’s fiscal Q2 2026 results on March 18 were absurdly strong: $23.86 billion in revenue, 74.9% non-GAAP gross margin, and a forecast for more records in fiscal Q3. SK hynix then posted first-quarter revenue of 52.5763 trillion won with a 72% operating margin, saying strong AI demand lifted sales of HBM, high-capacity server DRAM modules, and eSSDs. (news.skhynix.com) Samsung followed with all-time-high quarterly results and said memory profits were boosted by higher ASPs and limited supply availability. ### Who gets squeezed first? Server buyers and storage buyers. TrendForce said CSPs and server OEMs were already competing for limited supply and negotiating long-term agreements to lock in DRAM. Once that starts, smaller buyers lose flexibility fast. Quote windows get shorter, suppliers protect strategic accounts, and the “optional” server refresh turns into “buy now or pay more later.” Consumer devices may feel it too, but the first pain lands in data centers. (investors.micron.com) ### Does this mean the whole market is broken? Not broken — but repriced. Memory has always been cyclical. What looks different this time is that AI created a premium lane so profitable that it can starve the rest of the market without demand everywhere else needing to boom. That is why even soft PC or phone demand does not automatically bring relief. The industry can stay tight as long as AI memory keeps winning the allocation fight. (trendforce.com) ### So what is the bottom line? The memory story is no longer about a routine rebound. It is about AI turning DRAM and NAND into constrained infrastructure. If you run clouds, buy servers, or model 2026 hardware costs, the new assumption is simple: memory is expensive, lead times are tighter, and relief probably comes later than buyers want. (trendforce.com)