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

- **Google splits TPU line into v8t (training) and v8i (inference),** with Virgo network linking 134K chips at 47 Pbps. Materially shifts inference efficiency play against Nvidia; second-order: disaggregated AI infrastructure becomes standard, pressuring unified GPU architectures. - **Nvidia invests $2B in Marvell for NVLink Fusion interconnect,** signaling intent to lock down intra-cluster fabric. Second-order: hyperscalers may accelerate open Ethernet adoption (Arista doubling 2026 AI revenue target to $3B+) to avoid vendor lock-in. - **Meta deploys millions of Amazon Graviton CPUs for AI agents,** marking largest non-GPU AI compute shift. Implication: CPU-based inference scales for agentic workloads, reducing GPU dependency for low-latency tasks. - **Maine passes first US statewide AI data center moratorium,** while Utah hyperscale site to exceed state power use. Regulatory and grid-capacity friction now material; second-order: gas-fueled centers face emissions scrutiny (could emit more than entire nations). - **Applied Digital lands 300MW lease with hyperscaler in Louisiana,** and OpenAI's 'Freebird' Texas data center costs $470M for 549K sq ft. Hyperscaler buildout accelerates in secondary markets; implication: power availability, not chip supply, becomes binding constraint. - **PayPal cuts LLM inference cost 50% with EAGLE3 speculative decoding on H100,** and Alibaba opens Qwen app via China Eastern. Second-order: inference optimization (speculative decoding, disaggregation) becomes competitive moat; cost halving pressures hyperscaler margins.