AINEWS
Module makers warn 2027 supply could fall more than 70% as AI locks up the chips that feed every CPU and GPU, with consumer memory already up roughly 500% in 12 months.
A 128GB DDR5 memory kit now lists at $3,399, up from a multi-year low near $300. That is a roughly 500% climb in 12 months and roughly 10x the cheapest kit Tom's Hardware's memory price index has ever tracked, and the same shape is repeating at every capacity tier from gaming-PC DIMMs to the registered modules that anchor servers (Tom's Hardware). DDR5 is the current-generation DRAM standard for desktops, laptops, and most servers. The $3,399 kit is the kind of memory a workstation or high-end desktop needs to keep a fast CPU fed.
AI training and inference use high-bandwidth memory, the stacked DRAM that sits next to a GPU and feeds it data. The same fab capacity that makes HBM would otherwise make mainstream DDR5, so every wafer that ships as HBM is a wafer that does not ship as a consumer stick. Hyperscalers have reportedly pre-locked almost all of the global DRAM production capacity for 2027 with advance deposits, and consumer and PC DRAM allocations could fall below 30% of 2026 levels (AINews / Latent Space).
Apacer CEO Chang told investors and trade press that DRAM chip supply to module makers could fall more than 70% year over year in 2027, with Q3 2026 DRAM contract prices expected to rise about 30% and NAND flash contracts up more than 20% (Tom's Hardware). The forecast comes from a single module maker, not a measured outcome, and it sits alongside the price-tracker's recorded climb rather than replacing it. The Q3 2026 contract step is the next real price event, and it lands before the 2027 supply contraction does.
Daniel Lemire, a computer-science researcher who tracks memory economics, has pointed out that per-unit RAM cost is now back to roughly 2007 levels, undoing about 20 years of exponential decline that the industry had come to treat as a law of nature (Lemire on X). Tom's Hardware's coverage notes that mainstream DRAM chips are now worth more than half as much per kilogram as solid gold, a comparison the trade press uses to mark the shift from commodity to strategic input. Both observations point the same way: a multi-year cost curve has flattened, and the people who planned around the curve are the ones who have to re-plan.
A price spike ends when the demand behind it cools: a buyer pauses, a fab ramps, the backlog clears, and the long decline resumes. A regime change does not end that way. The capacity stack that supports DDR5 is the same stack that supports HBM, and the buyers at the top of that stack are signing multi-year offtake agreements with cash deposits. Apacer's forecast is the load-bearing piece of evidence for treating this as a regime change. If module makers lose more than 70% of their chip supply in 2027, the only ways consumer prices ease are if HBM demand cools faster than the contracts unwind, or if new fab capacity comes online in 2028 or later. Neither is on the contract calendar.
The second-order effects cut wider than the consumer aisle. A cloud instance priced by the hour bakes memory cost into the rate card, and a memory step that large flows into the per-vCPU economics of every tier from burstable general-purpose VMs to memory-optimized databases. Workstation and small-server builders who treat memory as a build-on-demand line item have to choose between absorbing the step, deferring upgrades, or substituting lower-capacity configurations and pushing more work to swap and SSD. The same arithmetic hits the second-hand market, where used DDR4 kits that were worthless a year ago are now the cheaper option for capacity-bound builds.
Builders and buyers should treat the Q3 2026 DRAM contract window as the real price signal, not the spot listings. Memory that was a build-on-demand line item is now a procurement question: lock the kits you need for late 2026 and 2027 builds, or price in the chance that consumer DRAM becomes the residual claimant on a capacity stack that AI has already cornered.