Memory prices have surged 500% year-over-year, undoing nearly two decades of Moore's Law price reductions. A 128GB DDR5 kit now costs ten times more than the lowest price ever recorded, pushing standard DRAM into "RAMageddon" territory. On a per-kilogram basis, mainstream DRAM chips are now worth more than half the price of solid gold, reflecting unprecedented AI data-center demand crowding out consumer and enterprise segment supply.
Hyperscale buyers have already locked in nearly all of 2027's global DRAM production capacity with advance deposits, signaling that the shortage will persist well into next year. AI data centers are consuming an estimated 70% of the world's memory output in 2026, up from 20–30% a few years ago, according to IDC. Memory now accounts for more than 80% of a GPU's bill of materials, particularly in high-VRAM cards where GDDR7 16GB modules jumped from roughly $65–$80 in mid-2025 to over $200 by year-end 2025 when long-term supply contracts expired.
For architects, this is no longer a cyclical shortage: it is structural. Every dollar spent on VRAM is a dollar not spent on compute, and contention for the world's memory fabs is now the primary binding constraint on scaling. Expect vendors to continue exploring chiplet strategies, memory pooling (via CXL), and alternative architectures that bypass HBM and GDDR altogether. GPU spot rates and rental costs will continue climbing unless supply normalizes—analysts expect no meaningful relief before Q4 2027 at earliest.