[DC] What Changed in AI Infra — Week 2026-W24

- **xAI pivots Colossus to rental compute**, chasing 30%+ margins; Google to pay $920M/month for capacity — signals hyperscaler demand outstripping internal supply, enabling secondary GPU leasing markets. - **NVIDIA NVFP4 on Blackwell cuts JAX training 1.8x** in MaxText — precision compression now a first-class optimization lever; expect rapid adoption across LLM pre-training stacks. - **MLCC shortage threatens AI server ramp**: prices hiking, lead times stretching — passive component bottleneck may delay Q3/Q4 cluster deliveries; hyperscalers are pre-buying capacity. - **UK doubles sovereign AI cloud providers**, deploys 65MW Nebius cluster — government-backed compute diversification; second-order effect: regulatory pressure on US hyperscalers to localize infrastructure. - **Liquid cooling crosses 50% by 2027**; CoolIT coldplate hits 15kW, Cerebras re-engineers wafer-scale cooling — rack densities force shift; air-cooled retrofits become stranded assets. - **Google’s 1 GW Texas AI campus tests ‘power-first’ model** — hyperscalers now securing grid capacity before chip procurement; shifts data center site selection from latency to energy arbitrage.