Memory Chips: Why The Next AI Device Wave Will Overwhelm Every Forecast
Every memory forecast published in the last twelve months has been wrong in the same direction: too low. IDC, TrendForce, and Bank of America have each revised DRAM and NAND demand estimates upward multiple times since early 2025, and the pattern is not noise — it is a structural failure of forecasting methodology colliding with a demand curve that refuses to plateau.
The Forecasts Keep Missing In One Direction
The current numbers are already staggering. IDC now expects 2026 DRAM supply growth of only 16% year-on-year, with NAND supply growth at just 17%, both well below the 20-30% historical norms that defined the post-2018 memory market. HBM demand alone is projected to grow 70% year-over-year in 2026, with HBM consuming 23% of total DRAM wafer output, up from 19% the year before. Bank of America forecasts DRAM revenue surging 51% year-over-year and NAND 45%, with ASPs rising 33% and 26% respectively.
None of this accounts for what happens when consumer AI hardware — not data center accelerators — becomes the second leg of demand.
Data Centers Are Only Half The Story
The current shortage narrative centers almost entirely on hyperscaler buildouts. Data centers are projected to consume as much as 70% of all high-end memory in 2026, and Samsung and SK Hynix are reportedly planning to raise server memory prices by up to 70% this quarter alone, following 50% increases in 2025. That is the supply side forecasters have been modeling.
What’s largely absent from these models is the second wave: on-device AI moving from novelty to default configuration across smartphones, PCs, and edge hardware. IDC defines the AI PC as any PC with an NPU, and these devices carry meaningfully more RAM, with Microsoft’s Copilot+ PC requirement set at a 16GB minimum and higher-end systems shifting toward 32GB or more as language models move on-device. That is not a data center story. That is hundreds of millions of consumer units, each requiring double or quadruple the memory footprint of a pre-AI device, entering a market where DRAM capacity growth is already limited to just 10-15% annually.
The Structural Mismatch Compounds
The mechanics make this worse than a simple supply-demand gap. Because HBM production uses larger, more complex dies, a single HBM wafer displaces two or more conventional DRAM wafers, creating a net reduction in global memory bits even as total wafer output stays flat. One gigabyte of HBM consumes four times the manufacturing capacity of standard DRAM, meaning AI’s drain on fabs vastly outpaces its share of memory actually shipped.
Executives across the supply chain are already framing this as unprecedented rather than cyclical. IDC has stated that while the memory industry has historically been defined by boom-and-bust cycles, this inflexion point is different, with demand materially outpacing supply. Lam Research CEO Tim Archer told a conference in South Korea that what lies ahead in memory demand through the end of the decade will overwhelm all other sources of demand the industry has faced.
Why The Next Wave Breaks The Models
Every current forecast is built on today’s mix of demand drivers: hyperscaler training clusters, inference buildouts, and a gradual consumer AI ramp. That mix is about to shift. When AI-native smartphones, always-on wearables, robotics platforms, and edge inference devices hit mainstream volume — not pilot volume — they add a demand layer that current models treat as incremental rather than structural.
The precedent is instructive: NAND and DRAM capacity planning cycles run 3-5 years from groundbreaking to output. Micron is currently building fabs in Boise that won’t produce memory until 2027 and 2028, with a New York facility not expected online until 2030. Consumer AI device adoption curves, by contrast, can inflect in a single product cycle. A memory industry that is already sold out — SK Hynix secured demand for its entire 2026 RAM production capacity back in October, and Micron says it is sold out for 2026 — has no slack to absorb a second demand shock arriving on top of the first.
The base case for memory investors should not be “shortage persists into 2027.” It should be that every forecast revision from here continues to run in one direction, because the models keep pricing in the AI infrastructure wave that’s already visible and keep underpricing the consumer device wave that hasn’t hit volume yet.
The Position
Samsung, SK Hynix, and Micron remain the direct beneficiaries of a structural mismatch that consumer electronics OEMs cannot engineer their way around. The risk to this thesis is not demand — it’s premature capacity expansion collapsing pricing before the second wave arrives. Watch capex guidance and wafer allocation commentary from all three suppliers as the leading indicator of whether that risk is materializing.