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

- **Google splits TPU line** into 8t (training) and 8i (inference), signaling explicit disaggregation of compute for AI workloads; Virgo network links 134k TPU v8 chips at 47 Pbps — second-order: hyperscalers architecting purpose-built fabrics for training vs. inference, raising bar for network silicon. - **Nvidia invests $2B in Marvell** for NVLink Fusion interconnect; B200 cost at $6,400 with 82% gross margin — operator move: Nvidia locking in custom interconnect supply chain; implication: NVLink ecosystem deepens, challenging InfiniBand and Ethernet alternatives. - **Meta deploys millions of Amazon Graviton CPUs** for AI agents — operator move: Meta shifting inference to Arm-based CPUs for cost efficiency; second-order: hyperscalers increasingly decouple CPU-based agent serving from GPU training clusters. - **AWS CEO confirms never retired an A100 server** amid chip shortage — implication: capacity scarcity persists, pushing operators to extend legacy GPU lifecycles; signals delayed refresh cycles for inferencing fleets. - **AI chip capacity crisis: 10GW left through 2030, prices up double digits** — operator move: Applied Digital lands 300MW hyperscaler lease in Louisiana; second-order: site-level power constraints driving pre-emptive leasing and SMR investments (X-energy IPO, Amazon-backed). - **Maine passes first US statewide AI data center moratorium** — regulatory shift: state-level permitting risk now live; second-order: hyperscaler site selection pivoting to deregulated or pre-permitted jurisdictions, inflating land/energy costs in remaining markets.