Memory crunch & geopolitical risk
Analysts flag persistent memory and advanced‑node constraints plus regional instability as risks that could keep GPU/HBM lead times tight for years — SK’s chairman says the memory chip crunch may persist until 2030. The brief also calls out the US‑Iran conflict as a supply‑chain tail‑risk for semiconductors. (bloomberg.com) (tomshardware.com)
SK Hynix reported a record operating profit of ₩11.38 trillion and revenue of ₩24.45 trillion for the quarter when it said customers had already secured its full memory chip lineup for 2026. (turn4search3) Micron’s management confirmed during its fiscal 2025 earnings call that the company’s HBM capacity for calendar 2025 — and strong demand signals into 2026 — are already fully booked. (turn1search15) TechInsights’ teardown of NVIDIA’s HGX B200 identified SK hynix as the HBM3E supplier and showed the card’s use of TSMC’s CoWoS‑L bridge packaging for dual GB100 dies. (turn0search13) Industry trackers and trade press report TSMC’s CoWoS lineup and other advanced‑packaging options are heavily booked — with TrendForce and others saying CoWoS‑L/S slots are effectively sold out into the near term. (turn2search8) TSMC executives and analysts have warned that bleeding‑edge wafer capacity (2nm/3nm class nodes) is tight versus customer roadmaps, with observers saying existing advanced‑node capacity falls materially short of planned AI demand. (turn2search7) The Iran‑related Gulf disruption has forced Qatar’s Ras Laffan LNG/helium hub offline, removing roughly one‑third of global helium supply and prompting official notices that helium and other materials shortages could disrupt chip fabs. (turn5search0) Taken together, vendor confirmations of sold‑out HBM allocations, identified HBM suppliers on flagship accelerators, booked CoWoS‑L packaging, and constrained 2–3nm wafer slots create multiple serial chokepoints for GPU+HBM delivery schedules through 2026–27. (turn0search13)
