AI's Memory Chips Eat 3× the Factory Space of Yours. Your Laptop Just Noticed.
A 32GB RAM kit that cost about $90 in early 2025 was fetching roughly $530 by that December, and still runs $375-400 today. Factories aren't short on capacity — they're spending it on Nvidia's memory instead of yours.
The wafer trade-off
Drag the slider to shift a fab's wafer starts toward AI's HBM. Top bar: how capacity is split. Bottom two bars: the resulting memory output, on one shared scale — because HBM costs ~3x the wafer capacity per gigabyte, moving fabs toward it produces less total memory, not more. Illustrative model: the 3x cost ratio and the $90→price scaling are grounded in real reporting; the exact allocation split is simplified for illustration, calibrated to land near today's real ~$400 kit price at the default position.
Shift a fab's wafer starts toward HBM:
A 32GB kit priced near $90 at 0% HBM allocation would cost roughly $409 at this split — in the range of today's real street price.
The plain version
Something strange happened to computer memory this year: a 32-gigabyte RAM kit that cost about $90 in early 2025 was selling for roughly $530 by that December, and even after prices cooled slightly, a basic kit still runs $375 to $400 today. This isn't a normal chip shortage — the factories are running flat out. The problem is what they're running flat out to build.
Picture a memory factory as a bakery with one oven. For years that oven turned out standard loaves: the ordinary RAM chips that go into every laptop, phone and PC. Then AI companies started ordering a fancier product — HBM, "high-bandwidth memory," the special stacked chips that sit right next to GPUs like Nvidia's so they can feed them data fast enough. HBM isn't just pricier to make, it's dramatically more wasteful of the same oven space: each stack needs larger, more failure-prone dies threaded with thousands of microscopic vertical wires, so one gigabyte of HBM eats up roughly three gigabytes' worth of an ordinary RAM chip's factory time.
Meta, Google, Microsoft and Amazon signed years-long contracts to lock in that oven space. Samsung and SK Hynix, who together make about 90% of the world's HBM, happily obliged — and normal RAM production got squeezed. Laptop makers responded by quietly shrinking base memory, with 16GB defaults becoming 8GB, or soldering RAM in to control costs. Analysts don't expect real relief before 2028, when new factories are expected to finally catch up. The AI boom isn't just buying up chips anymore. It's buying the memory that used to be yours.
The expert version
This is a wafer-capacity allocation problem, not a raw fab shortage. High-bandwidth memory (HBM) stacks 8 to 16 DRAM dies, connected by through-silicon vias (TSVs), then bonds the stack above a GPU or accelerator on an interposer, trading areal bit density for bandwidth and proximity to compute. TSVs, redundancy circuitry, and thicker inter-die bonding pads consume die area that would otherwise store bits: SK Hynix's D1z-node DDR4 reaches roughly 0.296 Gb/mm², about 85% denser than its HBM3 die at ~0.16 Gb/mm². Add stacking yield loss — one bad die can scrap an entire multi-die stack — and the widely cited industry rule of thumb holds: a gigabyte of HBM output consumes roughly three gigabytes' worth of equivalent DDR5 wafer capacity from the same fab.
Because DRAM fabs are largely fungible between HBM and commodity DDR4/DDR5 at the front-end process level, every wafer hyperscalers redirect toward HBM is a wafer that doesn't become conventional DRAM — and a worse trade than 1:1. SK Hynix (roughly 50% HBM share as of Q2 2026), Samsung (32%), and a fast-scaling Micron (18%, racing toward ~100,000 wafers a month by year-end from a much smaller base) have signed multi-year HBM supply agreements with Meta, Google, Microsoft and Amazon that lock in capacity years ahead, leaving commodity DRAM and NAND to absorb the residual demand.
Pricing reflects that squeeze. Contract DRAM prices rose as much as 89% in a single 2026 quarter; spot 16Gb DDR5 die pricing rose roughly 298% between September and December 2025 alone. Gartner projects a combined 130% DRAM-plus-NAND price increase for full-year 2026 versus 2025, pushing memory's share of PC bill-of-materials from roughly 16% to 23% and finished-device prices up 17% for PCs and 13% for smartphones. Intel and Silicon Motion both peg normalization at 2028, when new fab capacity is expected online; SK Hynix has said HBM demand could outrun its own supply through 2030.