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

- **Google breaks ground on $15B India DC; splits TPU line into 8t (training) & 8i (inference).** Virgo network links 134K TPU v8 chips at 47 Pbps — topo and disaggregation pattern now public. Second-order: Google signals vertical silicon + fabric lock-in for hyperscale AI, challenging NVLink dominance. - **Nvidia invests $2B in Marvell for NVLink Fusion interconnect; B200 cost ~$6,400 with 82% gross margin.** SemiAnalysis notes Nvidia customer data drives disaggregated inference, with LPU surpassing GPU in specific workloads. Implication: Nvidia is building a proprietary inter-chip fabric moat, while margin data suggests pricing power persists despite rising production costs. - **AWS CEO confirms never retired an A100 server amid chip shortage; Meta deploys millions of Amazon Graviton CPUs for AI agents.** Second-order: Hyperscalers are extending legacy GPU lifetimes and offloading inference to custom ARM CPUs — reduces Nvidia dependency for non-training workloads. - **Maine passes first US statewide AI data center moratorium; Utah hyperscale DC to exceed state power use.** Gas-fueled DCs could emit more than entire nations. Implication: Regulatory and energy bottlenecks are now material — operators face siting delays and potential carbon compliance costs. - **Applied Digital lands 300MW lease with hyperscaler in Louisiana; X-energy raises $1B+ IPO for Amazon-backed SMRs.** Second-order: Small modular nuclear and underutilized grid sites become strategic assets — hyperscalers are locking long-term power before capacity clears. - **PayPal cuts LLM inference cost 50% with EAGLE3 speculative decoding on H100; OpenCLAW-P2P v6.0 cuts paper lookup latency <50ms.** Implication: Inference optimization at the software layer is compressing cost faster than hardware generations — may delay demand for next-gen silicon in some workloads.