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

- **Google commits $40B to Anthropic**; Anthropic separately secures 5GW AWS compute in a $100B+ deal. Implication: two hyperscalers are now directly underwriting Anthropic’s capacity, signaling a shift from spot GPU leasing to long-term, multi-GW reserved clusters for frontier model training. - **Maine passes first U.S. statewide AI data center moratorium**; Microsoft’s Fairwater data center launches early. Implication: regulatory pushback is accelerating in some regions, while hyperscalers race to bring capacity online ahead of potential permitting freezes elsewhere. - **Meta deploys millions of Amazon Graviton CPUs for AI agents**; Arista doubles 2026 AI revenue target to $3B+ on open Ethernet. Implication: disaggregated inference and CPU-driven AI agent workloads are becoming material, challenging the GPU-centric narrative and boosting ARM/Open Ethernet ecosystems. - **Nvidia B200 cost at $6,400, gross margin 82%**; SemiAnalysis claims LPU surpasses GPU for inference and disaggregated architectures gain traction. Implication: Nvidia’s margin strength faces structural pressure as custom inference silicon and disaggregated designs erode its unit economics over the next 18 months. - **Foxconn to mass-produce 10,000+ CPO optical switches in Q3 2026**; Cisco says AI GPU networking demands 14x DCI bandwidth. Implication: co-packaged optics and silicon photonics are moving from lab to factory, signaling a near-term interconnect bottleneck shift from compute to fabric. - **Gas-fueled AI data centers could emit more than entire nations**; X-energy raises $1B+ IPO for Amazon-backed SMRs. Implication: nuclear small modular reactors are gaining real capital and hyperscaler backing as the only scalable low-carbon option, while gas buildout faces growing environmental and regulatory headwinds.