[DC] What Changed in AI Infra — Week 2026-W18

- **AWS CEO confirms no A100 server ever retired;** signals extreme fleet utilization amid GPU scarcity, implying secondary market supply constraints persist and hyperscalers prioritize depreciation extension over refresh cycles. - **Nvidia invests $2B in Marvell for NVLink Fusion Interconnect**, signaling shift toward custom silicon partnerships to lock in high-bandwidth fabric advantage against UALink 2.0 (spec finalized this week). - **Google commits up to $40B to Anthropic** while deploying 134K TPU v8 Virgo fabric (47 Pbps), indicating vertical integration of compute supply with frontier model funding to secure exclusive access. - **Maine passes first US statewide AI data center moratorium**; gas-fired AI DC emissions could exceed entire nations — regulatory precedent may slow greenfield builds in other states. - **Applied Digital lands 300MW hyperscaler lease in Louisiana**; Phononic (AI DC startup) enters sale talks at multi-billion valuation — consolidation wave accelerating as 10GW capacity gap through 2030 drives double-digit price hikes. - **Foxconn to mass-produce 10K+ CPO optical switches by Q3 2026**; Arista doubles AI revenue target to $3B on open Ethernet — disaggregated networking gains momentum as NVLink alternative matures.