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

- **Google commits up to $40B to Anthropic; Anthropic separately locks 5GW AWS compute in a $100B+ deal.** This dual-cloud strategy signals Anthropic is preparing for hyperscale training/inference demand well beyond current capacity, and that Google/AWS are willing to finance frontier-model infrastructure at unprecedented scale. - **Meta deploys millions of Amazon Graviton CPUs for AI agents**, marking a shift from GPU-only inference to CPU-based agentic workloads. This validates ARM server CPUs for real-time, low-latency AI tasks and pressures Intel/AMD in the AI inference market. - **Nvidia B200 gross margin hits 82%** at $6,400 production cost, while AI chip capacity crisis drives 10GW power shortfall and double-digit price increases through 2030. Nvidia’s pricing power remains extreme, but the supply-side bottleneck is now shifting from silicon to power/grid capacity. - **Maine passes first U.S. statewide AI data center moratorium**; Microsoft and Google shift to range-based capacity planning at DC World 2026. Regulatory and planning uncertainty is hardening, forcing hyperscalers to adopt flexible, modular buildouts rather than fixed megaprojects. - **Arista doubles 2026 AI revenue target to $3B+ on open Ethernet; UALink 2.0 finalized to challenge NVLink.** Competition to Nvidia’s networking monopoly is accelerating, with open standards gaining real revenue traction and hyperscaler backing. - **Foxconn to mass-produce 10,000+ CPO optical switches in Q3 2026; Cisco says AI GPU networking needs 14x DCI bandwidth.** Silicon photonics and co-packaged optics are moving from lab to production, solving intra-cluster bandwidth bottlenecks but requiring massive retooling of data center interconnect.