AI’s Trillion-Dollar Chip Boom Is Getting Bigger, and NVIDIA Isn’t the Only Winner - benzinga.com

The "trillion-dollar" framing is itself the tell: AI chip competition has shifted from "who builds the best GPU" to "who locks down the entire compute supply chain." NVIDIA's moat is being eroded from three directions simultaneously. On technology, the HBM4 production window (H2 2025) is migrating value from logic dies toward memory and advanced packaging (CoWoS-L). TSMC in Taiwan, China and SK Hynix in Korea capture the highest margin elasticity this cycle. Hyperscaler ASICs—TPU v6, Trainium2, Maia—are demoting NVIDIA from sole supplier to one-of-many, while open-source inference stacks dilute CUDA's lock-in. On compliance, US export restrictions on advanced chips to China have evolved from targeted throttling into structural decoupling, inflating global supply-chain redundancy costs so that any single foundry's capacity hiccup triggers cross-continental price transmission. On competition, AMD's MI325X value play and Intel Gaudi 3's "free software" strategy are fundamentally about price-anchoring to slice into NVIDIA's premium. The real 18-month variable isn't silicon—it's power and thermal. Data-center PUE ceilings will determine which clusters actually run at full utilization. That, not chip performance, is the invisible cap on the next capex wave.