Demand and supply have decoupled by segment. AI data-center build-out is pulling logic and high-end memory vertical, while the capacity to serve it — leading-edge wafers, CoWoS packaging, HBM — is the binding constraint. Everything else competes for what's left.
The 2026 up-cycle is unusually narrow and unusually steep. WSTS expects logic to grow ~37% and memory ~28% — the categories tied to AI accelerators and the servers around them — while mature segments like analog, discrete and optoelectronics grow in the mid-single digits. AI chips now drive roughly half of revenue from a tiny slice of unit volume, a structural shift toward high-value silicon.
Memory is where the squeeze is most visible. Samsung, SK Hynix and Micron — together ~95% of DRAM and NAND — are re-tooling fabs toward HBM and server DDR5 to feed AI, starving the conventional market. The result: rising prices across the board and warnings of shortages lasting through 2027. SK Hynix held ~53% of the HBM market in Q3 2025, ahead of Samsung (~35%) and Micron.
The answer to a demand shock is capacity, and capacity costs more every cycle. Global semiconductor capex is forecast to reach roughly $200B in 2026 (up ~20%), with TSMC alone accounting for over a quarter of it. New leading-edge fabs in Arizona, Taiwan, Japan and Korea are coming online, but the gating items — EUV tools, CoWoS packaging, HBM — have long lead times.
HBM, CoWoS advanced packaging, leading-edge (3/2nm) wafers, server DDR5. Customers reserving supply years ahead.
Conventional DRAM & NAND (spillover from HBM re-allocation), substrates, some power devices.
Mature-node analog, microcontrollers and auto chips after the 2022–23 glut; mainstream consumer logic.