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

- **xAI pivots Colossus to rental compute**: xAI shifts Colossus from internal training to third-party rental, targeting 30%+ margins. Google commits $920M/month for capacity, signaling hyperscaler appetite for spot GPU access and validating secondary compute markets. - **Google’s Virgo interconnects 134K TPUv8t chips at 47 Pbps**: Google deploys massive TPU fabric in production, pushing interconnect bandwidth limits. Implication: network topology becomes the bottleneck for next-gen training; rivals must match or risk model scale disadvantage. - **Google’s 1 GW Texas campus tests ‘power-first’ model**: Hyperscaler prioritizes site power availability over proximity to grid. PJM warns load assumptions are broken. Second-order: regulatory pushback likely; operators may pre-commit to behind-the-meter generation. - **MLCC shortage threatens AI server ramp**: Capacitor prices rising, lead times stretching. Implication: server assembly delays compound GPU shortages; hyperscalers may pre-book passive components or vertically integrate supply chains. - **UK doubles sovereign AI cloud providers, deploys 65MW Nebius cluster**: Government-backed capacity expansion diversifies European cloud supply. Nebius enters as a sovereign operator; second-order: regulatory pressure on US hyperscalers to offer local data sovereignty solutions. - **Liquid cooling crosses 50% by 2027; CoolIT quadruples coldplate capacity**: Rack densities force thermal transition. Foxconn/Intel partner on rack systems, Cerebras reengineers wafer-scale cooling. Implication: air-cooled designs become stranded assets; cooling infrastructure vendors gain pricing power.