AI demand is squeezing chip supply
Market chatter says AI is locking advanced nodes (GPU/HBM) while demand for mature nodes (power semiconductors, MCUs) is tightening too — an industrial rebound is piling pressure across the chain. SEMICON China execs also flagged equipment and MLCC shortages plus talent strain as immediate bottlenecks to scaling production. (x.com) (x.com)
TSMC’s 3nm wafer slots are reported fully booked 18–24 months out, with lead times for 3nm-based AI accelerators pushing past 50 weeks. (siliconanalysts.com) CoWoS packaging and HBM supply are cited as acute bottlenecks, with packaging/substrate constraints creating roughly a 40–50% shortfall versus current AI assembly demand. (siliconanalysts.com) Memory vendors including SK Hynix and Micron have said HBM capacity for calendar 2025–2026 is essentially sold out, prompting fabs to reallocate production away from standard DRAM. (fusionww.com) Analysts estimate AI data centers could consume about 70% of DRAM production this year, a shift that industry reports link to rising DRAM spot prices and tighter consumer-memory availability. (enkiai.com) Mature-node supply is tightening: MCU lead times have stretched back toward crisis-era levels and several manufacturers scheduled price increases effective April 1, 2026, while PMIC and power-IC allocations are rising as industrial orders rebound. (j2sourcing.com) Automotive-grade and AI-server MLCCs now show lead times of 20–26+ weeks with select high-capacitance parts still under allocation, a constraint suppliers and EMS buyers flagged at recent industry gatherings. (whychips.com) At SEMICON China (March 25–27, 2026, Shanghai) executives publicly urged faster equipment delivery and warned of a talent shortfall for advanced-node ramping, with some leaders calling for domestic equipment programs to address the backlog. (semiconchina.org) Market-level effects are appearing: multiple reports show DRAM prices up roughly 20% in recent months and analysts say some consumer GPU production (reported cuts to NVIDIA’s RTX 50-series of ~30–40%) reflects HBM-driven reallocation. (tech-insider.org)
