Applied Materials CEO says AI race hinges on technology and capacity

The AI compute race has quietly shifted its bottleneck from single-die performance to system-level delivery. AMAT's CEO framing this as a two-front war—technology and capacity—is the equipment industry's acknowledgment that lithography is no longer the sole chokepoint. Hybrid bonding, TSV, and chiplet interconnect capacity will dictate actual AI chip shipments through 2026. Technically, as process nodes approach physical limits, 3D stacking and chiplet interconnect become the only growth vector. AMAT's hybrid bonding, Lam's TSV etch, and TEL's CMP form a new equipment iron triangle rivaling ASML's lithography monopoly. HBM4 ramp is tightly coupled to CoWoS-L capacity, heavily concentrated in Taiwan, China's packaging fabs—a stark single-point supply risk. On compliance, advanced packaging equipment sits outside current export controls but will likely enter the restricted list by 2025. That would stretch non-US foundry expansion cycles by 18 to 24 months, creating a structural AI compute supply gap. Competitively, Intel's Foveros, Samsung's X-Cube, and TSMC's SoIC will collide head-on in 2026. AMAT's cross-customer coverage is a natural moat, but Lam is accelerating its packaging push—equipment share battles will intensify sharply. Within 18 months, packaging capacity replaces wafer capacity as the core pricing variable for AI silicon. UCIe versus CXL camp fragmentation will spawn two independent ecosystem trees, forcing equipment vendors to pick sides and accelerating industry consolidation.