DRAM and HBM prices surge 234%

Memory prices are spiking again — not because consumer gadgets suddenly got hot, but because AI servers are swallowing the best memory supply in the market. That supply shift is now large enough to hit ordinary DRAM and NAND, not just the premium HBM stacks bolted onto AI GPUs. (trendforce.com) The actual news is pretty specific: on March 31, TrendForce raised its Q2 2026 contract-price outlook to 58–63% quarter over quarter for conventional DRAM and 70–75% for NAND Flash. ### What is HBM, exactly? HBM is high-bandwidth memory — a dense, stacked form of DRAM that sits right next to an AI accelerator and feeds it data much faster than ordinary server memory can. (spglobal.com) That matters because modern AI chips are often bottlenecked by memory movement, not raw compute. TrendForce calls this the “memory wall” problem: compute keeps getting faster, but memory bandwidth is not keeping up. ### Why does that spill into normal DRAM? Because the same three companies — Samsung, SK hynix, and Micron — make both the fancy AI memory and a lot of the standard DRAM used in servers, PCs, and phones. When those companies push more advanced capacity toward HBM and server parts, less supply is left for everything else. (trendforce.com) S&P Global’s January note basically framed it as a capacity diversion — AI memory up, legacy DRAM tighter, prices higher. ### What changed this week? The big change is that price pressure is no longer a vague “tight market” story. TrendForce put fresh numbers on it for Q2 2026. (trendforce.com) Conventional DRAM is now expected to rise 58–63% from Q1, and NAND 70–75%. It also said suppliers are using “catch-up pricing” across segments, which is a polite way of saying they’re trying to close old price gaps fast while they still have leverage. ### Why are cloud companies at the center? Because North American cloud providers are locking in supply with long-term agreements as they build more AI infrastructure. TrendForce said high-capacity RDIMMs have become a primary procurement target, and near-term server-memory supply remains tight even as suppliers prioritize it for margin reasons. (spglobal.com) So the biggest buyers are not just buying more — they are reserving future output. ### Does this mean HBM itself is up 234%? Probably not in the clean, literal way that headline suggests. The strongest public figures I found point to huge jumps in mainstream DRAM and NAND contract pricing in 2026, while HBM pricing looks firmer and supply-constrained but not obviously up 234% on the same basis. (trendforce.com) In fact, S&P Global said HBM average selling prices in 2026 may rise much less than conventional DRAM, partly because HBM already started from a very high base. ### So why are people feeling a shock anyway? Because even if HBM unit pricing is not exploding at the same rate as commodity DRAM, total AI-system memory cost is still brutal. (trendforce.com) HBM is already expensive, only a few suppliers can make it, and capacity expansion takes time. Meanwhile, standard DDR5 and other server memory are also rising, so the whole bill of materials for an AI cluster gets hit at once. That is the nasty part. ### Who gets squeezed first? Enterprise buyers without giant prearranged supply deals. TrendForce said OEMs with weaker allocation fulfillment are being pushed into higher-priced procurement, and smartphone brands are already struggling to absorb the increases after two straight quarters of sharp hikes. (spglobal.com) In other words, the market is rewarding scale and punishing everyone else. ### Bottom line? This is less a one-off price spike than an AI-driven memory reshuffle. HBM is the star, but the collateral damage is spreading through the whole memory stack — and that means higher server costs, tighter component supply, and uglier budgeting for anyone building compute in 2026. (trendforce.com 1) (trendforce.com 2)