Conventional DRAM contract prices rose by roughly 93% to 98% quarter over quarter in the first quarter of 2026, and the analyst firm TrendForce projects a further 58% to 63% rise in the second. The cause is not a shortage of factories but a reallocation of them: memory makers are steering capacity toward high-bandwidth memory and high-capacity server modules for AI datacentres, and everyone else is bidding for what is left.
Key facts
- TrendForce reports conventional DRAM contract prices up approximately 93% to 98% quarter over quarter in 1Q26, lifting total memory industry revenue 81% to $97 billion.
- It projects a further 58% to 63% quarter-over-quarter rise for conventional DRAM in 2Q26, with NAND flash contract prices up 70% to 75%.
- TrendForce attributes the move to suppliers "reallocating capacity toward HBM and server applications," leaving PC makers and module vendors short.
- Primary source: TrendForce, June 1, 2026 and TrendForce, March 31, 2026.
Price moves of this size do not happen in commodity components. Memory is famous for gentle multi-year gluts punctuated by mild squeezes; a near-doubling in one quarter, followed by a forecast of another 60%, is a different kind of event. TrendForce's explanation is mundane and therefore credible. Suppliers have extremely low inventory, incremental output is being prioritised for the high-capacity server modules that AI inference deployments want, and cloud providers have shown willingness to accept the higher prices -- which promptly teaches every other buyer to pay up or lose their allocation.
The mechanism is worth being precise about, because "AI is eating the RAM" is only half right. AI accelerators use high-bandwidth memory, a specialised stacked product, not the sticks in a desktop. But HBM and ordinary DRAM come off the same wafers in the same fabs. Every wafer devoted to the higher-margin product is a wafer not making the cheaper one. The analogy is a bakery that discovers wedding cakes pay ten times what bread does: no flour shortage, and the bread shelf still empties.
Why this matters to anyone reading AI news rather than semiconductor news: system memory has quietly become an AI component. The current generation of open-weight designs deliberately pushes bulky model components off the graphics card and into system RAM -- Qwen's newest architecture ships a 97.7 GiB lookup table designed to live there, and community reports show people running it with around 100 GB of combined memory on a mid-range card. That was a clever way around expensive video memory right up until ordinary memory started repricing too.
The graphics-card side of the same squeeze is easier to see. NVIDIA launched the GeForce RTX 5090 at $1,999 in January 2025, according to its own announcement. Retail listings checked during this reporting showed 5090-class cards well above twice that figure. Between the card and the sticks, the cost of a machine that can run a large model at home has moved a long way from where it sat a year ago -- a squeeze consumers have already felt through memory-driven laptop price rises, and one reason a 512 GB Mac Studio reads differently now than it did at launch.
There is a real counter-argument. Contract prices are what large buyers negotiate, not what a retail shopper pays this afternoon, and the two can diverge for months in either direction. TrendForce also notes that HBM is priced annually rather than quarterly, so the headline volatility in the conventional segment partly reflects contract timing rather than pure demand. And there is a plausible bear case: PC demand has been revised downward, so if AI server buildouts slow, capacity swings back and prices unwind quickly.
The practical response in the local-model community has not been to buy more memory. It has been to compress harder -- lean on mixture-of-experts models where only a fraction of parameters are active, quantize aggressively, and budget carefully for the key-value cache. Running models at home was always a fight against memory bandwidth and capacity. It just got more expensive to lose.
The vendor-level numbers show how concentrated the gains are. TrendForce reports Samsung's quarterly revenue up 93.4% to $37.32 billion with a 38.5% share, and SK hynix up 62.5% to $27.98 billion, with the difference partly explained by hynix's heavier mix of high-bandwidth memory, whose contract prices are set annually and therefore did not ride the quarterly spike. That is a slightly counterintuitive result worth holding onto: the supplier most exposed to AI memory captured less of the AI memory boom, because its prices were locked in before it happened.
Originally published on Ground Truth, where every claim is checked against the primary source.
