DDR5 Prices Hit 2008 Levels as AI Memory Demand Starves PCs

DDR5 memory costs $11.41 to $13.28 per gigabyte, according to the Stanford DAM Project — price levels last seen in 2008. Software performance scientist Daniel Lemire noted the reversal: "RAM on a per unit basis is about as expensive as it was in 2007. I cannot recall a similar historical anomaly." Counterpoint Research puts Q1 2026 DRAM prices up 80 to 90 percent over the prior quarter. Bloomberg reports a 700 percent spot price increase over the past year. The cause is not a manufacturing failure. It is structural demand reallocation.

High-bandwidth memory is the mechanism. Each HBM chip stacks as many as 12 individual DRAM dies connected through silicon vias. Nvidia's B300 GPU uses eight of these chips — 96 DRAM dies per card, 768 dies in a fully configured DGX B300 system. Nvidia's NVL72 rack holds 13.4 terabytes of RAM, equivalent to roughly 1,000 high-end smartphones. When Micron produces one bit of HBM, it gives up three bits of conventional DRAM on the same wafer, according to Sumit Sadana, Micron's business chief. SemiAnalysis estimates HBM accounts for 50 percent or more of packaged GPU cost. Samsung, SK Hynix, and Micron — collectively controlling over 95 percent of global DRAM production — have reallocated manufacturing capacity toward HBM for AI accelerators, leaving consumer-grade DRAM in critically short supply. Data centers now consume an estimated 70 percent of all memory chips produced worldwide.

The industry had no buffer. After the COVID glut reversed, Samsung cut production by roughly 50 percent in 2022–2023 to keep prices above manufacturing cost. The sector made little or no investment in new fab capacity through 2024 and most of 2025, per IEEE Spectrum. New fabs cost $15 billion or more and take 18 months or longer to reach volume. Micron and SK Hynix will have new HBM capacity in 2027 at the earliest. SK Hynix secured demand for its entire 2026 RAM production capacity in October 2025. SK Hynix's CEO warned in July 2026 that demand will outpace supply well beyond 2030.

The demand gap is severe. Memory output grows roughly 20 percent per year. AI demand is growing at approximately 200 percent per year, per Elon Musk on SpaceX's earnings call. IDC projects 2026 DRAM supply growth at 16 percent year-on-year — well below the historical 20–30 percent range. Micron's CEO Sanjay Mehrotra projects HBM market revenue will grow from $35 billion in 2025 to $100 billion by 2028, larger than the entire DRAM market in 2024 and arriving two years ahead of Micron's prior forecast. OEM clients are receiving only 50 to 66 percent of their ordered memory volumes, per CNBC reporting in March 2026.

The squeeze on commodity memory is total. IDC calls it "a zero-sum game: every wafer allocated to an HBM stack for an Nvidia GPU is a wafer denied to the LPDDR5X module of a mid-range smartphone or the SSD of a consumer laptop." Memory now accounts for roughly 20 percent of laptop hardware costs, up from 10–18 percent in the first half of 2025. Smartphone shipments are projected to decline 12.9 percent in 2026; the PC market faces an 11.3 percent contraction. Micron discontinued its consumer PC memory segment to redirect supply toward AI chips.

IDC describes the pivot as "a potentially permanent, strategic reallocation of the world's silicon wafer capacity," not a cyclical correction. Prior supercycles tracked PC and smartphone unit volume. AI model scaling consumes memory as a direct function of parameter count and context length, with no ceiling in sight. The next generation of accelerators will demand more HBM per card, not less. Lemire puts the resolution bluntly: "We either find a way to build AI systems without so much memory, or we find really clever ways to make much more memory much faster."

Architects sizing inference clusters for 2026 should treat HBM and DDR5 allocation — not raw compute — as the primary planning constraint. No meaningful price relief is expected before 2027 at the earliest.