\n\n
Key developments: TSMC’s most advanced CoWoS packaging lines are fully booked through 2027, with lead times of 52‑78 weeks, and Nvidia consumes 60‑70% of that capacity. Nvidia’s Vera Rubin AI accelerators rely on a 150‑partner Taiwanese network, making them especially vulnerable to any Taiwan Strait disruption. Meanwhile, Samsung and SK Hynix are testing AMEC equipment in China as a hedge against potential U.S. tool‑servicing bans. China’s rare‑earth export controls, paused only until November 2026, remain in place for gallium, germanium and antimony, further tightening material supplies.
Why it matters: The combination of fully booked advanced‑packaging capacity, geopolitical exposure in Taiwan, and looming material export controls creates multi‑year queues for new AI chips. Memory pricing, already seeing system price hikes above 15% for HBM‑heavy servers, signals that memory scarcity—not just logic fab capacity—will drive costs into 2027.
What to watch next: Monitor whether CoWoS lead times fall below the 52‑week threshold, which would signal easing pressure. Watch the November 2026 reinstatement of China’s rare‑earth controls and quarterly memory earnings for signs of HBM supply relief. Also track U.S.–China AI talks, as any shift could impact tool‑servicing restrictions and the broader supply‑chain risk landscape. [[1]]