[DC] What Changed in AI Infra — Week 2026-W17

- **Nvidia invested $2B in Marvell** for NVLink Fusion interconnect, signaling a strategic shift to silicon photonics and disaggregated GPU topologies to overcome bandwidth bottlenecks in 100k+ GPU clusters. - **Google committed up to $40B to Anthropic** and deployed 134,000 TPU v8 chips via Virgo network (47 Pbps fabric), materially expanding its captive AI capacity and vertical integration ahead of rivals. - **Meta deployed millions of Amazon Graviton CPUs** for AI agents, marking a major shift from GPU-centric to heterogeneous compute for inference workloads, with implications for CPU-GPU ratio planning. - **Applied Digital secured 300MW hyperscaler lease** in Louisiana, while **Microsoft's Fairwater DC launched early**—both indicate hyperscalers are accelerating buildouts despite power constraints and regulatory headwinds (Maine moratorium). - **Arista doubled 2026 AI revenue target to $3B+** on open Ethernet, while **UALink 2.0 spec finalized**—together signaling a credible alternative to NVLink and InfiniBand, pressuring Nvidia's networking margins. - **AI chip capacity crisis deepens**: SemiAnalysis notes 10GW shortfall through 2030, prices up double digits, while **Nvidia B200 costs $6,400 to produce at 82% gross margin**—implying hyperscalers may front-run supply via multi-year prepayments and custom ASIC diversification.