A 32GB DDR5 kit that cost $80 to $100 through most of 2024 was running closer to $800 by January. Prices began climbing in October 2025, and megaton.ai reported no sign they would stop.
The cause is not a fire or a flood at a fab. It is artificial intelligence. Data centers were set to absorb 70% of all memory chips made in 2026, up from roughly 40% two years earlier, according to the same report.
Why a wafer of HBM beats a wafer of RAM
AI clusters want High Bandwidth Memory. A typical server might use 256GB of standard DRAM, while a single AI training node needs 8 to 24 HBM modules with about ten times the bandwidth. HBM takes roughly three times the wafer capacity of standard DRAM to make.
The margins explain the rest. Megaton.ai put HBM gross margins at 85% against about 15% for consumer memory, with a module selling for $1,000 or more against about $50 of equivalent consumer RAM. Samsung, SK Hynix and Micron have announced more than $50 billion in combined HBM expansion through 2027, while spending on consumer DRAM has flatlined.
Micron went furthest. It said it would discontinue its entire Crucial consumer RAM and SSD line, shifting capacity to “larger, strategic customers in faster-growing segments.” Aihola noted the company held nearly 25% of the DRAM market. OpenAI’s Stargate project alone reportedly signed agreements for up to 900,000 DRAM wafers a month.
Who is paying
TrendForce forecast conventional DRAM prices rising 55% to 60% in the first quarter of 2026, with NAND flash up 33% to 38%. Aihola cited DRAM up about 170% year over year. Older memory suffered too: DDR4 jumped 50% in a quarter as suppliers quit making it, and industrial buyers reported 40-week lead times on DDR3.
The bills reach well beyond PC builders. Raspberry Pi raised prices in October, with CEO Eben Upton saying memory costs had climbed 120% in a year. The Register reported that Dell and HPE were expected to pass costs straight to enterprise customers, and a server configuration at $15,000 in December could reach $18,000 by March. Smartphone makers faced component costs up 25% or more. Global DRAM inventories had fallen to 3.3 weeks, matching the 2018 low.
No quick fix
Building a new memory fab takes at least three years, so capacity decided now would not yield chips until late 2028 or 2029. TeamGroup’s general manager estimated supply would not normalize before 2027 or 2028 at the earliest.
Megaton.ai argued this crunch differs from the GPU shortage, because server memory modules are usually destroyed when decommissioned for data security. There is no flood of used parts to pull prices down. Its advice to those still buying: waiting likely means paying more.
Sources
- RAM Prices Triple as AI Data Centers Devour Memory Supply
- Memory Prices Double as AI Eats the World's RAM Supply
