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

- **xAI pivots Colossus to rental compute** chasing 30%+ margins, signaling hyperscaler overbuild is being monetized via secondary compute markets; Google paying SpaceX $920M/month for xAI capacity confirms this shift. - **Google Books Intel for 3M+ TPUs in 2028** as TSMC CoWoS hits capacity wall, forcing hyperscalers to diversify foundry partners; second-order: Intel's foundry ramp becomes critical for AI silicon supply chain. - **MLCC shortage threatens AI server ramp** with prices hiking and lead times stretching; passive component bottleneck adds 4-6 month delays to GPU cluster deployments, compounding existing memory and CoWoS constraints. - **Crusoe claims 5 GW pipeline, pauses Wyoming site**; gas-flare-to-compute model faces permitting and grid interconnection hurdles, suggesting modular data center buildouts are hitting real-world deployment ceilings. - **TensorWave raises $350M Series B for AMD-powered clusters**; AMD GPU ecosystem gains dedicated operator, challenging Nvidia's rental compute dominance as xAI and Crusoe both pivot to GPU-as-a-service models. - **CATL invests in DeepSeek**; battery giant pivots to AI energy, signaling energy storage + compute co-location as next hyperscaler frontier; second-order: AI power demand is reshaping industrial conglomerate strategy.